Final Report Summary - SENCES (An open simulation environment for self-diagnosis and prognosis in composite aero-structures)
Executive Summary:
The design of self – diagnostic structures usually requires several physical tests in order to evaluate the performance of an SHM system and the structure itself. The SENCES project focused on the development of a software tool called OSAS platform that will use validated analysis tools instead of expensive physical tests. With this platform, the designer will be able to model a self – diagnostic structure and perform diagnostics and prognostics using simulated data coming from Multilevel Structural Analysis. Moreover, the platform will have the ability to perform design optimisation and to select the best SHM system properties.
In the course of realising the OASAS platform the following activities have been carried out:
1. Approximation techniques have been developed to build computationally inexpensive surrogate models (or meta-models) to replace expensive-to-run multilevel analysis code. The developed meta-models will be used to produce valuable information to perform Diagnostics and Prognostics. Two meta-models are going to be generated for the multilevel analysis model. A global and a local meta-model. These surrogate models use sampled information from high–fidelity analysis codes in order to generate faster and lower–fidelity approximations. Sampled information is determined by the use of Design of Experiments (DoE), which create a set of points that the analysis response will be computed. A Permutation Genetic Algorithm (PermGA) method has been developed in order to generate a high – quality DoE. In order to build reliable surrogates, a Kriging predictor has been combined with a typical Response Surface Method (RSM). The developed Matlab code is developed in a Graphical User Interface environment and has three main procedures: Design of Experiments, Kriging Approximation and Adaptive LHD (for inherited DoE).
2. A Diagnostic theory and tool has been developed to detect and identify damage using simulated signal information and could be used to assess and optimise the reliability of SHM system. The model takes into account the signal characteristics and uncertainties (e.g. signal noise, signal threshold, etc.) and determines the detected damage parameters (type, size and location). It also takes into account the quantification of the uncertainty coming from the signal. Further to the original expectations, the theory has been complemented with the capability of determining the actual damage size characteristics using the results of the detected state as well as with updating schemes of the damage state when additional monitoring data becomes available, based on Bayesian statistics. Thus, it is able to represent the “true” damage state and its characteristics as a function of time/cycles.
3. A comprehensive and detailed Prognostic theoretical tool has been developed to compute the structural behaviour over time. Initially, the Probability of detecting a life – limiting damage is computed using updated information (multiple monitoring events) coming from different SHM systems with different attributes (mixed mode procedures). The next step is to compute the Probability of structural failure due to a non – detected damage for the above cases (multiple – mixed mode). The third step is to calculate the Probability of not detecting a life-limiting damage with a specified rate of occurrence. In the end the synthesis of the above is generalized in order to compute the Probability of Failure for various damage types in different subcomponents (or zones) of the structure. This quantity is very critical in the determination of the maintenance schedule.
Project Context and Objectives:
The scope of the work is to develop an integrated environment called Open simulation platform for the evaluation and design of Sensorised Aero – Structures (called OSAS platform from now on). This platform will be able to perform diagnostics and prognostics using exclusively simulated instead of real data. Moreover it will be able to optimise the design of composite aero – structures and choose the best SHM system properties. In Figure 1, a general flowchart of the platform function is shown. Impact simulation is performed using an appropriate structural model, which has geometry and material as input variables. The response will initially be used to reproduce realistic sensors signal using a sensors model that uses sensors topology and threshold values as inputs. The resulted signal will be correlated with the response in order to build a diagnostic model that will be able to provide information on possible damages, Probability of Detection (PoD) and Probability of False Alarm (PoFA). Damage information will be used to evaluate the future behaviour of the structure and to follow a sensor – advised maintenance strategy. On the first hand, PoD and PoFA will be used as objectives in order to optimise the SHM system properties, and on the other hand, cost and weight objectives will form a multi - objective structural optimization
The processing of data from an SHM system necessarily will be done within a Probabilistic framework, since uncertain information is usually available from the sensors. This probabilistic approach is expected to be more beneficial compared to the traditional “safety factor” approach in terms of cost reduction and safety. Probabilistic calculations require intensive and often unbearable simulation effort, which are tackled by the introduction of advanced meta-modeling techniques and by making use of the resources of High Performance Computing (HPC) systems. The ability to perform probabilistic multi – scale calculations within a reasonable time frame, enables realistic optimisation studies that provide useful design alternatives to decision makers. The developed platform is not meant to be a single – purpose commercial software but a flexible and evolving open framework, able to receive different types of information and methods. The platform will be able to act as a “test bed” for the evaluation of new design concepts, SHM technologies and methods.
The work did not only focus on developing methods, but also on the integration of the existing knowledge. Following a holistic design approach, the vision for the future is the OSAS platform to constitute a part of a life – cycle virtual platform which is planned to integrate information coming from design, manufacturing, operational and disposal phase of a component or an aircraft.
Project Results:
WP1.1 Multi-level Analysis
Objective: The purpose of this task was to develop and use a Multilevel / Multi-scale Analysis model of a sensorised composite panel to compute i) structural response, ii) sensors integrity and iii) simulated signal coming from virtual sensors placed on the model of the structure, using a High Performance Computing (HPC) system. The model will be used to produce data in order to build accurate Meta-models (in WP1.2).
Outcome: D1.1 – Report and model for Multilevel/Multi-scale Analysis
WP1.2 Meta-modeling
Objective: The application of Approximation Techniques to build computationally inexpensive surrogate models (or meta-models) to replace expensive-to-run multilevel analysis code(s) (WP1.1). The developed meta-models were used (in WP1.3 and WP1.4) to produce valuable information to perform Diagnostics and Prognostics.
Outcome: D1.2 – Report and S/W module to construct meta-models (surrogate models)
WP1.3 Diagnostics
Objective: A Diagnostic tool has been developed to detect and identify damage using simulated signal information from WP1.2 and ii) assess the reliability of SHM system. The developed tool can be used (in WP1.4) to perform Prognostics.
Outcome: D1.3 – Report and S/W Diagnostics module
WP1.4 Prognostics
Objective: A Prognostic tool has been developed to compute the structural behaviour over time. The results could be used to follow a maintenance scheduling (in WP1.5).
Outcome: D1.4 – Report and S/W Prognostics module
WP1.5 Maintenance
Objective: A probabilistic maintenance strategy based on the sensor readings has been devised. The reduction of the operational costs will have a major impact on the computation of life – cycle costs (in WP2.2).
Outcome: D1.5 – Report and S/W Maintenance module
WP2.1 Platform specifications
Objective: The purpose of this work was to i) identify the nature of the elements that will be embedded inside the OSAS platform and ii) to plan their appropriate integration. The outcome of this task has been used in integrating the OSAS platform (in WP2.3).
Outcome: D2.1 – Report on OSAS platform’s components /platform specifications
WP2.2 Platform specifications
Objective: The purpose of this work was to provide simple Weight and Cost models to be integrated with Structural Reliability in the Optimisation framework (in WP2.3).
Outcome: D2.1 – Report on OSAS platform’s components /platform specifications
WP2.3 Platform specifications
Objective: The purpose of this work was the development of optimisation methods. Optimisation process has two main objectives; to optimise the SHM system properties (sensors topology, threshold values) and the structure itself (geometry, materials). For this purpose, advanced multi-objective optimisation methods have been developed.
Outcome: D2.1 – Report on OSAS platform’s components /platform specifications
WP2.4 Platform Integration
Objective: The purpose of this work was to integrate the developed modules into the OSAS platform.
Outcome: D2.2 –OSAS Platform (Completed software -offline function demonstration)
WP2.5 Implementation
Objective: The purpose of this work was to produce and present the results of a test case using the OSAS platform.
Outcome: D2.3 – Report on OSAS platform results
A sequential overview of the technical progress on a subject basis is given in the sections that follow.
3.1.1 WP2.1 Multi-level Analysis
WP 2.1 – Multi-level Analysis
The purpose of this task is to develop and use a Multilevel / Multi-scale Analysis model of a sensorised composite panel to compute i) structural response, ii) sensors integrity and iii) simulated signal coming from virtual sensors placed on the model of the structure, using a High Performance Computing (HPC) system. The model could be used to produce data in order to build accurate Meta-models (in WP1.2).
Work performed
Various multi-level fusion techniques, namely, the voting scheme, Bayesian inference, Dempster–Shafer rule and fuzzy inference have been presented individually for decision fusion to assess damaged structures based on multiple information sources. To avoid the risk of using inappropriate fusion methodology during damage assessment, further higher level of decision fusion could be applied to assessments based on different approaches. The key technique is to combine information (i.e. information fusion) from different sources, whatever is quantitative or qualitative, to make an overall judgment. Information fusion should be implemented not only at the level of raw data from sensors and features extracted from the raw data, but also at different levels of decision making. Both may reduce the risk of a malfunction of individual sensors and misjudgment by inappropriate assessment procedures. The voting algorithm, Bayesian inference, Dempster–Shafer rule, and fuzzy inference are the major techniques for decision fusion, whereby probability or possibility patterns of damage on structures can be established and the area most likely to be damaged can be identified.
Key achievements:
Several techniques of multi-level data fusion and decision have been examined and compared. The validation is based on a specimen of T650/F584 CF/EP [0/45/-45/90]s quasi-isotropic laminate which was manufactured with an artificial through-hole defect introduced. The responses of all passive sensors to the interrogating stress wave emitted from an active sensor were measured.
Although a better sensitivity to damages is obtained by the use of Bayesian fusion technique, it is still hard to conclude which technique would perform better for others. Therefore, it is also important to implement a high level of decision fusion, which is here in return the combination of results from all four fusion techniques. The technique of multilevel decision fusion shows much promise. Nevertheless, the establishment of a more comprehensive and broad knowledge database for various scenarios should be done, before further improvement of the accuracy and robustness of damage identification using multilevel decision fusion techniques along with distributed sensor networks is established. It should be also pointed out that the multilevel decision-making fusion process is not limited at local acquired sensing data, but also may include more information on design and even maintenance provided that they have direct or indirect links with structural damages and may provide further confidence on damage detection. Quantification of all kinds of information into a same description domain is necessary before any fusion is processed.
WP 2.2 – Meta-modeling
The work concerns the development of a methodology and of the appropriate S/W module that will be used in order to replace the “computationally expensive” multilevel finite element analysis (D1.1) with Kriging surrogate models. Surrogate models use sampled information from high–fidelity analysis codes in order to generate faster and lower–fidelity approximations. Sampled information is determined by the use of Design of Experiments (DoE), which create a set of points that the analysis response will be computed. In order to build reliable surrogates, a Kriging predictor has been combined with a typical Response Surface Method (RSM).
Work performed
A design of experiment (DoE) technique has been developed in order to compute the original model response at a representative amount of design sites (points). Kriging model acts as a surrogate model (or meta-model) and replaces the “expensive” analysis model by using these responses. Kriging approximation accuracy is evaluated through a cross validation procedure and if the accuracy has not reached the desired levels, refinement is performed using additional evaluations points. Additional points are arranged with the old ones (inherited) and the next Kriging approximations are performed. The appropriate S/W module has been developed, checked and validated.
Key achievements:
• A Permutation Genetic Algorithm (PermGA) method has been developed in order to generate a high – quality Design of Experiments (DoEs) in conjunction with a Latin Hypercube Sampling, or Latin Hypercube Design (LH or LHD or LHS). This method was found to be more accurate than random sampling and stratified sampling to estimate the means, variances and distribution functions of an output. Moreover, it ensures that each of the input variables has all portions of its range represented. It can cope with many input variables and is computationally cheap to generate. After optimization, a high quality Latin Hypercube Design is formed due to the fact that its points follow a distribution which is very close to uniform.
• A method for generating meta-models using the Kriging model has been developed. The method is a powerful tool that replaces a certain analysis model or a data set by a Surrogate Model (or Meta-model) of maximum accuracy for a large number of variables using relatively few data points. Kriging approximation model is a combination of a regression model and a stochastic process. It interpolates given ns data points and its parameters θk are evaluated through the Maximum Likelihood Estimation Criterion. The Kriging approximation has been augmented with the Low Discrepancy Adaptive Latin Hypercube Design (LDA-LDH) for improved distribution of design points.
• All of the work has been implemented into a Matlab code and representative screens are shown below:
WP 2.3 – Diagnostics
A correlation model will be developed to compute damage parameters (type, size, location) from signal. The uncertainty coming from the signal will be quantified. As a result, the detected damage parameters (size and location) will be characterized by statistical distributions instead of deterministic values. In order to speed up the process, the meta–models developed (in WP1.2) will replace the “sensors” and “structure” models. The developed correlation model will be able to receive raw signal as input and compute the following damage parameters: i) damage type as a discrete variable, ii) size and location as statistical distributions.
Work performed
The complete theoretical foundations for determining the probability distribution of damage size have been developed. The approach takes into account the signal threshold, saturation and noise. The signal noise is expressed via a probabilistic definition. The decision whether a signal indicates damage is made on the basis of a decision threshold. The combination of signal and decision thresholds and the signal noise determine whether damage has been identified or it is probably a false alarm. In the latter case due to the presence of the signal uncertainty, there is a (small) possibility of exceeding these bounds without having any damage on the structure. This probability is called Probability of False Alarm (PoFA) and it considered as a very important index when designing SHM systems. In the former case the estimated quantity is the Probability of Indication (PoI) or Probability of Detection (PoD) of damage sizes given specified signal and decision thresholds. Once these quantities are estimated by employing Maximum Likelihood (ML) techniques we can determine the probability distribution of damage size. PoD and PoFA are going to be computed using simulated data. In addition, there is the possibility for determining the probability distribution of actual damage sizes given a detected damage size density function and the detection density. This differentiates between the detected and actual damage sizes effect in estimating the damage tolerance of the structure since the analysis could be based on the actual damage sizes and not the detected ones. Last but not least an updating scheme for the distribution of damage size has also been implemented. Hence, if initial detected damage size probability distributions exist for the various SHM techniques the partners/users could update these definitions based on their own data – of course some underlying assumptions regarding the acquisition of data (noise, method etc.) should be considered.
Key achievements:
• A detailed approach has been developed for estimating damage size probability distribution functions taking into account signal characteristics and uncertainties as well as decision and signal thresholds. The key parameters of noise signal and signal saturation have been incorporated into the formulation as well as the application of various filters such as Kalman.
• A mathematical model based on Bayesian statistics has been developed for updating the damage size distributions given new data. This feature provides the possibility of updating damage size distributions when new detected damage size data is available from the SHM system.
• A mathematical formulation has been implemented for obtaining the actual damage size distributions given the detected damage size density and the detection density. This could help in evaluating the remaining life of a structure that takes into account the SHM characteristics as they are manifested in the detected damage size distributions and provide us with a basis for optimising the SHM system and its monitoring and detection capabilities.
• All of the above developments have been tested and provide us with the flexibility to account for renewal of data, SHM optimisation and the basis for sound remaining life estimations. The location and type of damage are treated as distinct variables.
• The necessary theoretical developments regarding the SHM Data Analysis, PoDs and damage size distributions determination have been presented. The implementation of these developments will constitute the so-called “Expert” Module of the probabilistic analysis software. An analysis screen of the software (Figure 3b) along with a schematic of the “Expert” Module (Figure 3b) is shown below.
Status
All theoretical development has been completed. Additional features such as updating schemes for damage size distributions and projected distribution for actual damage sizes have been incorporated. Preliminary testing of concepts has been performed and seems to provide us with a lot of insight on how to treat the developments, modeling and optimisation of SHM systems.
WP 2.4 – Prognostics
Since damage is described by statistical distributions, a probabilistic approach will be followed for the computation of the residual life of the structure. The approach will take into account the damage size distributions (either detected and/or actual), the SHM system’s characteristics (signal, noise, thresholds, detection capabilities), damage tolerance characteristics (required operational cycles, critical damage sizes and thresholds), as well as rate of occurrence of damage types. The presence of various types of damages in the structural component and their effects in estimating the residual life will also be taken into account. In order to represent the variation of damage parameters and compute failure probabilities, several runs of the structural analysis model will be performed using the meta-model of the structure, derived using the tools developed (in WP1.2) instead of the original analysis model (FEM). A representative set of values of damage size and location will be generated to compute the structural response. Residual life will be computed for every “point” using certain failure criteria and eventually, the probability of exceeding allowable response values (or Probability of Failure, PoF) will be computed.
Work performed
The first step in developing a complete and most generally applicable prognostics platform to estimate the residual life of a structural component was to develop the theoretical tools to appropriately represent the SHM events. Thus, in the current developments we have formulated the probability of not detecting a life limiting damage size escaping the SHM monitoring and leading to failure for various cases. The monitoring event could take place at any time and we have developed the theoretical tools for the simplest case, i.e. one monitoring event with a specified damage tolerance and signal threshold, to the most complex one where we have multiple mixed modes monitoring with various damage tolerance and signal thresholds.
A mixed mode procedure is defined as the one where several steps of monitoring are occurring and it may happen that in some of the intermediate steps an initial damage is detected and hence it triggers a subsequent monitoring to ultimately determine whether an damage is present or not. Since we are seeking the probability of an initial damage not being detected the final step it is a “no detection” step.
The probability of failure is then estimated as the joint event of having a life limiting damage and/or combination of damages on various components and/or zones - on the notion that given a SHM state and detection state are employed - not being detected and subsequently leading to failure. The theory accommodates various rates of occurrence for various damage types, location variability in terms of component zonal information as well as different monitoring schemes per zone and per damage type. Several approaches for estimating the probabilities of failure have also been implemented such as the meta-models (GKO) and limit state approximation (Most Probable Point – MPP) methods.
Key achievements:
• A detailed treatise of the SHM events – from the very simple to the most complex of sequences – has been formulated. It takes into account not only the signal characteristics and uncertainties but as well as damage tolerance thresholds, damage types, occurrence rates, and all other influences in SHM system’s performance. For the first time a mixed-mode procedure has been defined to account for back-up SHM contributions, and/or adjacent and alternative monitoring solutions and their effect in residual life estimation. The most simple detection event and the most complex are shown below along with the resulting expressions for the probability of not detecting a life limiting damage escaping the SHM system’s capabilities.
WP 2.5 – Maintenance
The Maintenance strategy will no longer follow a predefined schedule but will be continuously updated with information coming from the sensors. In this way, unnecessary inspection events will be avoided since the whole structure will be monitored by the SHM system. This unscheduled approach is expected to cause a significant elongation of the time between successive ground inspections, which will dramatically reduce maintenance costs. Conditional Probability of Failure (CPF) will be continuously updated by the information coming from the SHM system. Inspection interval will be computed having as constraint that the sensor - updated CPF must not exceed an upper bound.
Work performed
The entire theoretical framework for implementing the maintenance approach has already been set-up in the previous tasks. The tools for estimating the conditional probabilities of failure have been developed and linked to the monitoring and structural response platforms. In addition, the probabilities of failure based on updating of monitoring data and referral to actual damage sizes constitute the complete set of tools for devising a maintenance schedule based on changing probabilities of failure as they are estimated from data and structural response simulation nomographs or look-up tables. In addition, the thus far developed theory is additional complemented with the capability of taking into account the effect of replacing/repairing a number or all of the detected damages. The methodology for combining the diagnostic and prognostic methods developed, in devising a maintenance schedule
WP 3.1 – Platform specifications
The purpose of this work is to i) identify the nature of the elements that will be embedded inside the OSAS platform and ii) to plan their appropriate integration. A detailed flowchart of the interdependencies between the different modules will be created. It is very important to identify their input and output format in order to establish their connectivity. The outcome of this task will be used in integrating the OSAS platform (in WP2.3).
Work performed
The different modules composing the OSAS platform are depicted in the Figure 6 below. Initially, there is an EXTERNAL feeder module that pertains to a multi-level structural analysis and incorporates FEMs results at different levels of abstraction, including damage identification (type, location, etc.). This module has to provide the input for the OSAS platform analysis. The OSAS platform incorporates the following modules:
Meta-modeling (used to obtain less complex but still accurate SHM signal patterns and structural models representations)
This module uses DoE and GKO to generate optimized probabilistic meta-models for the signal and/or structural response. It could generate any type of a Kriging model (Response Surface-like model). The code has been developed in Matlab and generates graphical results as well as appropriate functional representations for the meta-models.
These meta-models are fed in both the Diagnostics and Prognostic modules - with the diagnostic and prognostic parameters defined - and used to calculate the probability of non-detecting an actual damage (future structural behaviour) that could lead to failure by using various modes of inspections as well as predicting the probability of failure under various conditions. In short,
Diagnostics (used to identify the characteristics of damage such as size distributions, probability of false alarms, probability of detections, etc.). The diagnostic module has been developed in Matlab and feeds information to the prognostics module.
Prognostics (used to identify the probability of failure taking into account the presence of damage, signal characteristics as well as various inspection techniques and design parameters). The Prognostics module has two independent sub-modules providing the choice of selecting the analysis route: (1) a part based on meta-models and developed in Matlab which has been integrated with the Diagnostics module and (2) a part based on “limit state approximations” approach and developed in FORTRAN.
The Fault-Tree-Analysis (FTA) module aims at synthesizing the multiple events that occur with certain probabilities and has been integrated into the Prognostics module. The FTA is configured and employed to couple the conditional failure probability (CPF) for various impact scenarios.
Subsequently, there is the Maintenance module aims at reducing the number of unnecessary inspection events. Maintenance module employs the damage information, obtained by the SHM, and the results from the Prognostics module in order to evaluate the future behavior of the structure. Therefore, significant maintenance cost reductions will be induced since the inspection interval will be continuously updated by the output of the Prognostics module. In addition, there is the Weight & Cost module, which targets to estimate the Weight of the structure and the Cost corresponding to the life-cycle (manufacturing, operation) of the component. Finally, the Optimization module incorporates the JPDM software tool, developed in Matlab, that provides optimum solutions on multi-criteria (weight, cost) problems by using several techniques, such as Latin-Hypercube Sampling, Monte-Carlo simulation, permutation Genetic Algorithm, and Kriging surrogate models.
WP 3.2 – Cost and weight models
The purpose of this work is to provide simple Weight and Cost models to be integrated with Structural Reliability in the Optimisation framework (of WP2.3). Cost models will be considered in a life – cycle framework. Costs coming from manufacturing and operation of the component will be analysed and computed. For the computation of the operational cost, maintenance strategy (from WP1.5) will be taken into account. Weight values will come from the FEM multilevel analysis.
Work performed
Based on the above notions an optimization framework for the cost/weight optimization of aircraft structures has been devised. The aim is to combine the cost portions of the DOC that are driving the design.
Key achievements:
The approach is easily integrated into the developed framework. Various models for the DOC could be used and “hooked” onto the analysis procedure.
WP 3.3 – Optimisation
Optimisation is an important task for the OSAS platform in order to highlight the minimisation of weight and cost with the use of an SHM system. Two main optimisation processes will be carried out. The first one deals with safety issues and will optimise the properties of the SHM system, while the second one involves structural optimisation with respect to weight and cost. For the optimisation of the SHM system, sensors topology (sensors location, sensor types and threshold values will change in order to achieve the optimal combination of PoD, PoFA and manufacturing cost. For the structural optimisation, Weight and Cost objectives (WP2.2) will form a multi – objective design optimisation problem.
Work performed
The Optimization module incorporates the Joint Probabilistic Decision Making (JPDM) software tool that provides optimum solutions on multi-criteria (weight, cost) problems by using several techniques, such as Latin-Hypercube Sampling, Monte-Carlo simulation, permutation Genetic Algorithm, and Kriging surrogate models. JPDM utilizes the information generated by modern probabilistic design procedures and combines this information into one evaluation criterion the (joint) Probability of Success (POS). POS is the objective function used by traditional optimization methods for multi-objective optimization or the selection criterion based on which the best design is identified among a closed set of alternatives. The objective function used by JPDM is not based on a summation of criteria, but rather the probability of satisfying all criteria at the same time, a notion similar to a Pareto-optimality. The main difference with respect to Pareto-optimality lies in the optimizable objective function, called probability of success (in satisfying all criteria). This multi-criteria approach to decision making lends itself more suitably to aircraft design than a probabilistic single-criterion approach, since customers typically like to see all decision criteria satisfied (Figure 7).
The POS function is approximated by low fidelity approximations (Metamodels, Surrogate Models). These approximations are generated sequentially until the desired accuracy of optimal solution is reached. A series of steps that incorporate adaptive methods of Sampling and Metamodeling are necessary in order to build sequential Surrogate models. Gradual Kriging Optimization (GKO) is used for this reason.
Key achievements:
• A new innovative optimization tool has been developed based on a combined evaluation criterion embedding all constraints, called POS. If the decision making problem at hand is an optimization, i.e. finding the best solution within the design space spanned by a set of design variables, the joint probability serves as the objective function. The optimal solution is found when the design with the highest POS is found within the design range.
WP 3.4 – Integration
The scope of the work was the integration of all developed concepts resulting to the first OSAS platform. A step-by-step presentation of the OSAS platform diagnostics-prognostics software is given. After the Probability of detecting a life-limiting damage is computed using updated information (multiple monitoring events) coming from different SHM systems with different attributes (mixed mode procedures). The next step is to compute the Probability of structural failure due to a non-detected damage for the above cases (multiple-mixed mode). The subsequent step is to calculate the Probability of non-detecting a life-limiting damage with a specified rate of occurrence. In the end the synthesis of the above is generalized in order to compute the Probability of Failure for various damage types in different subcomponents (or zones) of the structure. This quantity is very critical in the determination of the maintenance schedule.
Potential Impact:
Structural Health Monitoring (SHM) is a relatively new method in the aircraft industry of monitoring the condition of a structure in real time while the structure is in service. Diagnostic methods attempt to autonomously detect and localize damages in large structures. The current state of the art in Diagnostics includes several methods for damage detection based on data coming from Non – Destructive Inspection (NDI) or Structural Health Monitoring (SHM). Prognostics are also addressed in many projects, but most of them are related to NDI and not to a continuous self – diagnostic process. In addition, Diagnostics and Prognostics are mostly studied through deterministic rather than probabilistic approaches. Optimisation is also a very interesting topic, but is only used for choosing the best material or geometry by taking into account selected objectives without following a holistic approach. The innovative characteristics of the SENCES project and the attempt to develop an Open simulation platform for the evaluation and design of Sensorised Aero – Structures (called OSAS platform) still remains ahead of current developments as they are driven from:
• The vision for the future with the OSAS platform to constitute a part of a life – cycle virtual platform which is planned to integrate information coming from design, manufacturing, operational and disposal phase of a component or an aircraft leading to analysis and optimisation of aircraft components with embedded structural health monitoring sensors
• The Multilevel methodology -the structure is broken down into hierarchically ordered levels which may represent different analysis aspects of different subsets of the global structure, of varying fidelity and resolution
• A Probabilistic framework for Diagnostics, Prognostics and Optimisation
• Advanced meta-modeling techniques by making use of the resources of HPC systems and thus promoting such computationally demanding systems
• The complexity and accuracy requirements of the analysis and its integration ability
In some cases, additional concepts have been introduced than the ones described in the Technical Annex (e.g. Kalman filters in the diagnostics related activities, Limit State Approximations for estimating the probability of failure in the prognostics and maintenance related activities, etc.). The generality introduced into the theoretical developments may allow for the application into other areas of structural design and monitoring/inspection techniques.
The expected impact on the market of this type of system is to provide a set of procedures, architectures and components that will be implemented in an integrated SHM system. This will allow a significant reduction of the maintenance costs and a simplification-optimization of technical interventions, because the system will permit to pass from a fixed inspection plan on the basis of statistical hours of flights to a “condition maintenance based on diagnostic/prognostic methods”.
The Structural Health Monitoring with Embedded sensors (ESHM) approach on aeronautical structures can have a strong impact as a means of possibly revolutionizing the current structural monitoring, maintenance and design processes. While the OSAS platform elements do primarily target aircraft panels, the results could be used to study, produce and disseminate “best practices” and “highlights” for the use of CFRP materials in a variety of applications spanning multiple domains. Moreover though OSAS is mainly addressed to CFRP composite materials its embedded sensors architecture could be evaluated for an exploitation also addressed to other classes of materials like polymers and complex composites, even belonging to emerging technologies. Thus, the results could have the following cross-domain impacts:
• Reduce time to market despite the increasing contribution of embedded systems and software and their increasing size and complexity
• Increase the quality and reliability of products and services while providing novel functionalities to the user
• Improve cross-domain fertilisation
Aerospace, aviation, automotive and industrial SECTORS are application domains directly targeted by the OSAS architecture. The study of potential applications in these domains will boost cross-domain fertilisation through sharing of knowledge and best practices. OSAS will produce and disseminate “blueprints” detailing the use and benefits of CFRP components in multiple domains.
Dr. Dimitri Bofilios
INASCO
18 N. ZERVA ST.,
16675 - GLYFADA GREECE
+30 210 99 43 427
The design of self – diagnostic structures usually requires several physical tests in order to evaluate the performance of an SHM system and the structure itself. The SENCES project focused on the development of a software tool called OSAS platform that will use validated analysis tools instead of expensive physical tests. With this platform, the designer will be able to model a self – diagnostic structure and perform diagnostics and prognostics using simulated data coming from Multilevel Structural Analysis. Moreover, the platform will have the ability to perform design optimisation and to select the best SHM system properties.
In the course of realising the OASAS platform the following activities have been carried out:
1. Approximation techniques have been developed to build computationally inexpensive surrogate models (or meta-models) to replace expensive-to-run multilevel analysis code. The developed meta-models will be used to produce valuable information to perform Diagnostics and Prognostics. Two meta-models are going to be generated for the multilevel analysis model. A global and a local meta-model. These surrogate models use sampled information from high–fidelity analysis codes in order to generate faster and lower–fidelity approximations. Sampled information is determined by the use of Design of Experiments (DoE), which create a set of points that the analysis response will be computed. A Permutation Genetic Algorithm (PermGA) method has been developed in order to generate a high – quality DoE. In order to build reliable surrogates, a Kriging predictor has been combined with a typical Response Surface Method (RSM). The developed Matlab code is developed in a Graphical User Interface environment and has three main procedures: Design of Experiments, Kriging Approximation and Adaptive LHD (for inherited DoE).
2. A Diagnostic theory and tool has been developed to detect and identify damage using simulated signal information and could be used to assess and optimise the reliability of SHM system. The model takes into account the signal characteristics and uncertainties (e.g. signal noise, signal threshold, etc.) and determines the detected damage parameters (type, size and location). It also takes into account the quantification of the uncertainty coming from the signal. Further to the original expectations, the theory has been complemented with the capability of determining the actual damage size characteristics using the results of the detected state as well as with updating schemes of the damage state when additional monitoring data becomes available, based on Bayesian statistics. Thus, it is able to represent the “true” damage state and its characteristics as a function of time/cycles.
3. A comprehensive and detailed Prognostic theoretical tool has been developed to compute the structural behaviour over time. Initially, the Probability of detecting a life – limiting damage is computed using updated information (multiple monitoring events) coming from different SHM systems with different attributes (mixed mode procedures). The next step is to compute the Probability of structural failure due to a non – detected damage for the above cases (multiple – mixed mode). The third step is to calculate the Probability of not detecting a life-limiting damage with a specified rate of occurrence. In the end the synthesis of the above is generalized in order to compute the Probability of Failure for various damage types in different subcomponents (or zones) of the structure. This quantity is very critical in the determination of the maintenance schedule.
Project Context and Objectives:
The scope of the work is to develop an integrated environment called Open simulation platform for the evaluation and design of Sensorised Aero – Structures (called OSAS platform from now on). This platform will be able to perform diagnostics and prognostics using exclusively simulated instead of real data. Moreover it will be able to optimise the design of composite aero – structures and choose the best SHM system properties. In Figure 1, a general flowchart of the platform function is shown. Impact simulation is performed using an appropriate structural model, which has geometry and material as input variables. The response will initially be used to reproduce realistic sensors signal using a sensors model that uses sensors topology and threshold values as inputs. The resulted signal will be correlated with the response in order to build a diagnostic model that will be able to provide information on possible damages, Probability of Detection (PoD) and Probability of False Alarm (PoFA). Damage information will be used to evaluate the future behaviour of the structure and to follow a sensor – advised maintenance strategy. On the first hand, PoD and PoFA will be used as objectives in order to optimise the SHM system properties, and on the other hand, cost and weight objectives will form a multi - objective structural optimization
The processing of data from an SHM system necessarily will be done within a Probabilistic framework, since uncertain information is usually available from the sensors. This probabilistic approach is expected to be more beneficial compared to the traditional “safety factor” approach in terms of cost reduction and safety. Probabilistic calculations require intensive and often unbearable simulation effort, which are tackled by the introduction of advanced meta-modeling techniques and by making use of the resources of High Performance Computing (HPC) systems. The ability to perform probabilistic multi – scale calculations within a reasonable time frame, enables realistic optimisation studies that provide useful design alternatives to decision makers. The developed platform is not meant to be a single – purpose commercial software but a flexible and evolving open framework, able to receive different types of information and methods. The platform will be able to act as a “test bed” for the evaluation of new design concepts, SHM technologies and methods.
The work did not only focus on developing methods, but also on the integration of the existing knowledge. Following a holistic design approach, the vision for the future is the OSAS platform to constitute a part of a life – cycle virtual platform which is planned to integrate information coming from design, manufacturing, operational and disposal phase of a component or an aircraft.
Project Results:
WP1.1 Multi-level Analysis
Objective: The purpose of this task was to develop and use a Multilevel / Multi-scale Analysis model of a sensorised composite panel to compute i) structural response, ii) sensors integrity and iii) simulated signal coming from virtual sensors placed on the model of the structure, using a High Performance Computing (HPC) system. The model will be used to produce data in order to build accurate Meta-models (in WP1.2).
Outcome: D1.1 – Report and model for Multilevel/Multi-scale Analysis
WP1.2 Meta-modeling
Objective: The application of Approximation Techniques to build computationally inexpensive surrogate models (or meta-models) to replace expensive-to-run multilevel analysis code(s) (WP1.1). The developed meta-models were used (in WP1.3 and WP1.4) to produce valuable information to perform Diagnostics and Prognostics.
Outcome: D1.2 – Report and S/W module to construct meta-models (surrogate models)
WP1.3 Diagnostics
Objective: A Diagnostic tool has been developed to detect and identify damage using simulated signal information from WP1.2 and ii) assess the reliability of SHM system. The developed tool can be used (in WP1.4) to perform Prognostics.
Outcome: D1.3 – Report and S/W Diagnostics module
WP1.4 Prognostics
Objective: A Prognostic tool has been developed to compute the structural behaviour over time. The results could be used to follow a maintenance scheduling (in WP1.5).
Outcome: D1.4 – Report and S/W Prognostics module
WP1.5 Maintenance
Objective: A probabilistic maintenance strategy based on the sensor readings has been devised. The reduction of the operational costs will have a major impact on the computation of life – cycle costs (in WP2.2).
Outcome: D1.5 – Report and S/W Maintenance module
WP2.1 Platform specifications
Objective: The purpose of this work was to i) identify the nature of the elements that will be embedded inside the OSAS platform and ii) to plan their appropriate integration. The outcome of this task has been used in integrating the OSAS platform (in WP2.3).
Outcome: D2.1 – Report on OSAS platform’s components /platform specifications
WP2.2 Platform specifications
Objective: The purpose of this work was to provide simple Weight and Cost models to be integrated with Structural Reliability in the Optimisation framework (in WP2.3).
Outcome: D2.1 – Report on OSAS platform’s components /platform specifications
WP2.3 Platform specifications
Objective: The purpose of this work was the development of optimisation methods. Optimisation process has two main objectives; to optimise the SHM system properties (sensors topology, threshold values) and the structure itself (geometry, materials). For this purpose, advanced multi-objective optimisation methods have been developed.
Outcome: D2.1 – Report on OSAS platform’s components /platform specifications
WP2.4 Platform Integration
Objective: The purpose of this work was to integrate the developed modules into the OSAS platform.
Outcome: D2.2 –OSAS Platform (Completed software -offline function demonstration)
WP2.5 Implementation
Objective: The purpose of this work was to produce and present the results of a test case using the OSAS platform.
Outcome: D2.3 – Report on OSAS platform results
A sequential overview of the technical progress on a subject basis is given in the sections that follow.
3.1.1 WP2.1 Multi-level Analysis
WP 2.1 – Multi-level Analysis
The purpose of this task is to develop and use a Multilevel / Multi-scale Analysis model of a sensorised composite panel to compute i) structural response, ii) sensors integrity and iii) simulated signal coming from virtual sensors placed on the model of the structure, using a High Performance Computing (HPC) system. The model could be used to produce data in order to build accurate Meta-models (in WP1.2).
Work performed
Various multi-level fusion techniques, namely, the voting scheme, Bayesian inference, Dempster–Shafer rule and fuzzy inference have been presented individually for decision fusion to assess damaged structures based on multiple information sources. To avoid the risk of using inappropriate fusion methodology during damage assessment, further higher level of decision fusion could be applied to assessments based on different approaches. The key technique is to combine information (i.e. information fusion) from different sources, whatever is quantitative or qualitative, to make an overall judgment. Information fusion should be implemented not only at the level of raw data from sensors and features extracted from the raw data, but also at different levels of decision making. Both may reduce the risk of a malfunction of individual sensors and misjudgment by inappropriate assessment procedures. The voting algorithm, Bayesian inference, Dempster–Shafer rule, and fuzzy inference are the major techniques for decision fusion, whereby probability or possibility patterns of damage on structures can be established and the area most likely to be damaged can be identified.
Key achievements:
Several techniques of multi-level data fusion and decision have been examined and compared. The validation is based on a specimen of T650/F584 CF/EP [0/45/-45/90]s quasi-isotropic laminate which was manufactured with an artificial through-hole defect introduced. The responses of all passive sensors to the interrogating stress wave emitted from an active sensor were measured.
Although a better sensitivity to damages is obtained by the use of Bayesian fusion technique, it is still hard to conclude which technique would perform better for others. Therefore, it is also important to implement a high level of decision fusion, which is here in return the combination of results from all four fusion techniques. The technique of multilevel decision fusion shows much promise. Nevertheless, the establishment of a more comprehensive and broad knowledge database for various scenarios should be done, before further improvement of the accuracy and robustness of damage identification using multilevel decision fusion techniques along with distributed sensor networks is established. It should be also pointed out that the multilevel decision-making fusion process is not limited at local acquired sensing data, but also may include more information on design and even maintenance provided that they have direct or indirect links with structural damages and may provide further confidence on damage detection. Quantification of all kinds of information into a same description domain is necessary before any fusion is processed.
WP 2.2 – Meta-modeling
The work concerns the development of a methodology and of the appropriate S/W module that will be used in order to replace the “computationally expensive” multilevel finite element analysis (D1.1) with Kriging surrogate models. Surrogate models use sampled information from high–fidelity analysis codes in order to generate faster and lower–fidelity approximations. Sampled information is determined by the use of Design of Experiments (DoE), which create a set of points that the analysis response will be computed. In order to build reliable surrogates, a Kriging predictor has been combined with a typical Response Surface Method (RSM).
Work performed
A design of experiment (DoE) technique has been developed in order to compute the original model response at a representative amount of design sites (points). Kriging model acts as a surrogate model (or meta-model) and replaces the “expensive” analysis model by using these responses. Kriging approximation accuracy is evaluated through a cross validation procedure and if the accuracy has not reached the desired levels, refinement is performed using additional evaluations points. Additional points are arranged with the old ones (inherited) and the next Kriging approximations are performed. The appropriate S/W module has been developed, checked and validated.
Key achievements:
• A Permutation Genetic Algorithm (PermGA) method has been developed in order to generate a high – quality Design of Experiments (DoEs) in conjunction with a Latin Hypercube Sampling, or Latin Hypercube Design (LH or LHD or LHS). This method was found to be more accurate than random sampling and stratified sampling to estimate the means, variances and distribution functions of an output. Moreover, it ensures that each of the input variables has all portions of its range represented. It can cope with many input variables and is computationally cheap to generate. After optimization, a high quality Latin Hypercube Design is formed due to the fact that its points follow a distribution which is very close to uniform.
• A method for generating meta-models using the Kriging model has been developed. The method is a powerful tool that replaces a certain analysis model or a data set by a Surrogate Model (or Meta-model) of maximum accuracy for a large number of variables using relatively few data points. Kriging approximation model is a combination of a regression model and a stochastic process. It interpolates given ns data points and its parameters θk are evaluated through the Maximum Likelihood Estimation Criterion. The Kriging approximation has been augmented with the Low Discrepancy Adaptive Latin Hypercube Design (LDA-LDH) for improved distribution of design points.
• All of the work has been implemented into a Matlab code and representative screens are shown below:
WP 2.3 – Diagnostics
A correlation model will be developed to compute damage parameters (type, size, location) from signal. The uncertainty coming from the signal will be quantified. As a result, the detected damage parameters (size and location) will be characterized by statistical distributions instead of deterministic values. In order to speed up the process, the meta–models developed (in WP1.2) will replace the “sensors” and “structure” models. The developed correlation model will be able to receive raw signal as input and compute the following damage parameters: i) damage type as a discrete variable, ii) size and location as statistical distributions.
Work performed
The complete theoretical foundations for determining the probability distribution of damage size have been developed. The approach takes into account the signal threshold, saturation and noise. The signal noise is expressed via a probabilistic definition. The decision whether a signal indicates damage is made on the basis of a decision threshold. The combination of signal and decision thresholds and the signal noise determine whether damage has been identified or it is probably a false alarm. In the latter case due to the presence of the signal uncertainty, there is a (small) possibility of exceeding these bounds without having any damage on the structure. This probability is called Probability of False Alarm (PoFA) and it considered as a very important index when designing SHM systems. In the former case the estimated quantity is the Probability of Indication (PoI) or Probability of Detection (PoD) of damage sizes given specified signal and decision thresholds. Once these quantities are estimated by employing Maximum Likelihood (ML) techniques we can determine the probability distribution of damage size. PoD and PoFA are going to be computed using simulated data. In addition, there is the possibility for determining the probability distribution of actual damage sizes given a detected damage size density function and the detection density. This differentiates between the detected and actual damage sizes effect in estimating the damage tolerance of the structure since the analysis could be based on the actual damage sizes and not the detected ones. Last but not least an updating scheme for the distribution of damage size has also been implemented. Hence, if initial detected damage size probability distributions exist for the various SHM techniques the partners/users could update these definitions based on their own data – of course some underlying assumptions regarding the acquisition of data (noise, method etc.) should be considered.
Key achievements:
• A detailed approach has been developed for estimating damage size probability distribution functions taking into account signal characteristics and uncertainties as well as decision and signal thresholds. The key parameters of noise signal and signal saturation have been incorporated into the formulation as well as the application of various filters such as Kalman.
• A mathematical model based on Bayesian statistics has been developed for updating the damage size distributions given new data. This feature provides the possibility of updating damage size distributions when new detected damage size data is available from the SHM system.
• A mathematical formulation has been implemented for obtaining the actual damage size distributions given the detected damage size density and the detection density. This could help in evaluating the remaining life of a structure that takes into account the SHM characteristics as they are manifested in the detected damage size distributions and provide us with a basis for optimising the SHM system and its monitoring and detection capabilities.
• All of the above developments have been tested and provide us with the flexibility to account for renewal of data, SHM optimisation and the basis for sound remaining life estimations. The location and type of damage are treated as distinct variables.
• The necessary theoretical developments regarding the SHM Data Analysis, PoDs and damage size distributions determination have been presented. The implementation of these developments will constitute the so-called “Expert” Module of the probabilistic analysis software. An analysis screen of the software (Figure 3b) along with a schematic of the “Expert” Module (Figure 3b) is shown below.
Status
All theoretical development has been completed. Additional features such as updating schemes for damage size distributions and projected distribution for actual damage sizes have been incorporated. Preliminary testing of concepts has been performed and seems to provide us with a lot of insight on how to treat the developments, modeling and optimisation of SHM systems.
WP 2.4 – Prognostics
Since damage is described by statistical distributions, a probabilistic approach will be followed for the computation of the residual life of the structure. The approach will take into account the damage size distributions (either detected and/or actual), the SHM system’s characteristics (signal, noise, thresholds, detection capabilities), damage tolerance characteristics (required operational cycles, critical damage sizes and thresholds), as well as rate of occurrence of damage types. The presence of various types of damages in the structural component and their effects in estimating the residual life will also be taken into account. In order to represent the variation of damage parameters and compute failure probabilities, several runs of the structural analysis model will be performed using the meta-model of the structure, derived using the tools developed (in WP1.2) instead of the original analysis model (FEM). A representative set of values of damage size and location will be generated to compute the structural response. Residual life will be computed for every “point” using certain failure criteria and eventually, the probability of exceeding allowable response values (or Probability of Failure, PoF) will be computed.
Work performed
The first step in developing a complete and most generally applicable prognostics platform to estimate the residual life of a structural component was to develop the theoretical tools to appropriately represent the SHM events. Thus, in the current developments we have formulated the probability of not detecting a life limiting damage size escaping the SHM monitoring and leading to failure for various cases. The monitoring event could take place at any time and we have developed the theoretical tools for the simplest case, i.e. one monitoring event with a specified damage tolerance and signal threshold, to the most complex one where we have multiple mixed modes monitoring with various damage tolerance and signal thresholds.
A mixed mode procedure is defined as the one where several steps of monitoring are occurring and it may happen that in some of the intermediate steps an initial damage is detected and hence it triggers a subsequent monitoring to ultimately determine whether an damage is present or not. Since we are seeking the probability of an initial damage not being detected the final step it is a “no detection” step.
The probability of failure is then estimated as the joint event of having a life limiting damage and/or combination of damages on various components and/or zones - on the notion that given a SHM state and detection state are employed - not being detected and subsequently leading to failure. The theory accommodates various rates of occurrence for various damage types, location variability in terms of component zonal information as well as different monitoring schemes per zone and per damage type. Several approaches for estimating the probabilities of failure have also been implemented such as the meta-models (GKO) and limit state approximation (Most Probable Point – MPP) methods.
Key achievements:
• A detailed treatise of the SHM events – from the very simple to the most complex of sequences – has been formulated. It takes into account not only the signal characteristics and uncertainties but as well as damage tolerance thresholds, damage types, occurrence rates, and all other influences in SHM system’s performance. For the first time a mixed-mode procedure has been defined to account for back-up SHM contributions, and/or adjacent and alternative monitoring solutions and their effect in residual life estimation. The most simple detection event and the most complex are shown below along with the resulting expressions for the probability of not detecting a life limiting damage escaping the SHM system’s capabilities.
WP 2.5 – Maintenance
The Maintenance strategy will no longer follow a predefined schedule but will be continuously updated with information coming from the sensors. In this way, unnecessary inspection events will be avoided since the whole structure will be monitored by the SHM system. This unscheduled approach is expected to cause a significant elongation of the time between successive ground inspections, which will dramatically reduce maintenance costs. Conditional Probability of Failure (CPF) will be continuously updated by the information coming from the SHM system. Inspection interval will be computed having as constraint that the sensor - updated CPF must not exceed an upper bound.
Work performed
The entire theoretical framework for implementing the maintenance approach has already been set-up in the previous tasks. The tools for estimating the conditional probabilities of failure have been developed and linked to the monitoring and structural response platforms. In addition, the probabilities of failure based on updating of monitoring data and referral to actual damage sizes constitute the complete set of tools for devising a maintenance schedule based on changing probabilities of failure as they are estimated from data and structural response simulation nomographs or look-up tables. In addition, the thus far developed theory is additional complemented with the capability of taking into account the effect of replacing/repairing a number or all of the detected damages. The methodology for combining the diagnostic and prognostic methods developed, in devising a maintenance schedule
WP 3.1 – Platform specifications
The purpose of this work is to i) identify the nature of the elements that will be embedded inside the OSAS platform and ii) to plan their appropriate integration. A detailed flowchart of the interdependencies between the different modules will be created. It is very important to identify their input and output format in order to establish their connectivity. The outcome of this task will be used in integrating the OSAS platform (in WP2.3).
Work performed
The different modules composing the OSAS platform are depicted in the Figure 6 below. Initially, there is an EXTERNAL feeder module that pertains to a multi-level structural analysis and incorporates FEMs results at different levels of abstraction, including damage identification (type, location, etc.). This module has to provide the input for the OSAS platform analysis. The OSAS platform incorporates the following modules:
Meta-modeling (used to obtain less complex but still accurate SHM signal patterns and structural models representations)
This module uses DoE and GKO to generate optimized probabilistic meta-models for the signal and/or structural response. It could generate any type of a Kriging model (Response Surface-like model). The code has been developed in Matlab and generates graphical results as well as appropriate functional representations for the meta-models.
These meta-models are fed in both the Diagnostics and Prognostic modules - with the diagnostic and prognostic parameters defined - and used to calculate the probability of non-detecting an actual damage (future structural behaviour) that could lead to failure by using various modes of inspections as well as predicting the probability of failure under various conditions. In short,
Diagnostics (used to identify the characteristics of damage such as size distributions, probability of false alarms, probability of detections, etc.). The diagnostic module has been developed in Matlab and feeds information to the prognostics module.
Prognostics (used to identify the probability of failure taking into account the presence of damage, signal characteristics as well as various inspection techniques and design parameters). The Prognostics module has two independent sub-modules providing the choice of selecting the analysis route: (1) a part based on meta-models and developed in Matlab which has been integrated with the Diagnostics module and (2) a part based on “limit state approximations” approach and developed in FORTRAN.
The Fault-Tree-Analysis (FTA) module aims at synthesizing the multiple events that occur with certain probabilities and has been integrated into the Prognostics module. The FTA is configured and employed to couple the conditional failure probability (CPF) for various impact scenarios.
Subsequently, there is the Maintenance module aims at reducing the number of unnecessary inspection events. Maintenance module employs the damage information, obtained by the SHM, and the results from the Prognostics module in order to evaluate the future behavior of the structure. Therefore, significant maintenance cost reductions will be induced since the inspection interval will be continuously updated by the output of the Prognostics module. In addition, there is the Weight & Cost module, which targets to estimate the Weight of the structure and the Cost corresponding to the life-cycle (manufacturing, operation) of the component. Finally, the Optimization module incorporates the JPDM software tool, developed in Matlab, that provides optimum solutions on multi-criteria (weight, cost) problems by using several techniques, such as Latin-Hypercube Sampling, Monte-Carlo simulation, permutation Genetic Algorithm, and Kriging surrogate models.
WP 3.2 – Cost and weight models
The purpose of this work is to provide simple Weight and Cost models to be integrated with Structural Reliability in the Optimisation framework (of WP2.3). Cost models will be considered in a life – cycle framework. Costs coming from manufacturing and operation of the component will be analysed and computed. For the computation of the operational cost, maintenance strategy (from WP1.5) will be taken into account. Weight values will come from the FEM multilevel analysis.
Work performed
Based on the above notions an optimization framework for the cost/weight optimization of aircraft structures has been devised. The aim is to combine the cost portions of the DOC that are driving the design.
Key achievements:
The approach is easily integrated into the developed framework. Various models for the DOC could be used and “hooked” onto the analysis procedure.
WP 3.3 – Optimisation
Optimisation is an important task for the OSAS platform in order to highlight the minimisation of weight and cost with the use of an SHM system. Two main optimisation processes will be carried out. The first one deals with safety issues and will optimise the properties of the SHM system, while the second one involves structural optimisation with respect to weight and cost. For the optimisation of the SHM system, sensors topology (sensors location, sensor types and threshold values will change in order to achieve the optimal combination of PoD, PoFA and manufacturing cost. For the structural optimisation, Weight and Cost objectives (WP2.2) will form a multi – objective design optimisation problem.
Work performed
The Optimization module incorporates the Joint Probabilistic Decision Making (JPDM) software tool that provides optimum solutions on multi-criteria (weight, cost) problems by using several techniques, such as Latin-Hypercube Sampling, Monte-Carlo simulation, permutation Genetic Algorithm, and Kriging surrogate models. JPDM utilizes the information generated by modern probabilistic design procedures and combines this information into one evaluation criterion the (joint) Probability of Success (POS). POS is the objective function used by traditional optimization methods for multi-objective optimization or the selection criterion based on which the best design is identified among a closed set of alternatives. The objective function used by JPDM is not based on a summation of criteria, but rather the probability of satisfying all criteria at the same time, a notion similar to a Pareto-optimality. The main difference with respect to Pareto-optimality lies in the optimizable objective function, called probability of success (in satisfying all criteria). This multi-criteria approach to decision making lends itself more suitably to aircraft design than a probabilistic single-criterion approach, since customers typically like to see all decision criteria satisfied (Figure 7).
The POS function is approximated by low fidelity approximations (Metamodels, Surrogate Models). These approximations are generated sequentially until the desired accuracy of optimal solution is reached. A series of steps that incorporate adaptive methods of Sampling and Metamodeling are necessary in order to build sequential Surrogate models. Gradual Kriging Optimization (GKO) is used for this reason.
Key achievements:
• A new innovative optimization tool has been developed based on a combined evaluation criterion embedding all constraints, called POS. If the decision making problem at hand is an optimization, i.e. finding the best solution within the design space spanned by a set of design variables, the joint probability serves as the objective function. The optimal solution is found when the design with the highest POS is found within the design range.
WP 3.4 – Integration
The scope of the work was the integration of all developed concepts resulting to the first OSAS platform. A step-by-step presentation of the OSAS platform diagnostics-prognostics software is given. After the Probability of detecting a life-limiting damage is computed using updated information (multiple monitoring events) coming from different SHM systems with different attributes (mixed mode procedures). The next step is to compute the Probability of structural failure due to a non-detected damage for the above cases (multiple-mixed mode). The subsequent step is to calculate the Probability of non-detecting a life-limiting damage with a specified rate of occurrence. In the end the synthesis of the above is generalized in order to compute the Probability of Failure for various damage types in different subcomponents (or zones) of the structure. This quantity is very critical in the determination of the maintenance schedule.
Potential Impact:
Structural Health Monitoring (SHM) is a relatively new method in the aircraft industry of monitoring the condition of a structure in real time while the structure is in service. Diagnostic methods attempt to autonomously detect and localize damages in large structures. The current state of the art in Diagnostics includes several methods for damage detection based on data coming from Non – Destructive Inspection (NDI) or Structural Health Monitoring (SHM). Prognostics are also addressed in many projects, but most of them are related to NDI and not to a continuous self – diagnostic process. In addition, Diagnostics and Prognostics are mostly studied through deterministic rather than probabilistic approaches. Optimisation is also a very interesting topic, but is only used for choosing the best material or geometry by taking into account selected objectives without following a holistic approach. The innovative characteristics of the SENCES project and the attempt to develop an Open simulation platform for the evaluation and design of Sensorised Aero – Structures (called OSAS platform) still remains ahead of current developments as they are driven from:
• The vision for the future with the OSAS platform to constitute a part of a life – cycle virtual platform which is planned to integrate information coming from design, manufacturing, operational and disposal phase of a component or an aircraft leading to analysis and optimisation of aircraft components with embedded structural health monitoring sensors
• The Multilevel methodology -the structure is broken down into hierarchically ordered levels which may represent different analysis aspects of different subsets of the global structure, of varying fidelity and resolution
• A Probabilistic framework for Diagnostics, Prognostics and Optimisation
• Advanced meta-modeling techniques by making use of the resources of HPC systems and thus promoting such computationally demanding systems
• The complexity and accuracy requirements of the analysis and its integration ability
In some cases, additional concepts have been introduced than the ones described in the Technical Annex (e.g. Kalman filters in the diagnostics related activities, Limit State Approximations for estimating the probability of failure in the prognostics and maintenance related activities, etc.). The generality introduced into the theoretical developments may allow for the application into other areas of structural design and monitoring/inspection techniques.
The expected impact on the market of this type of system is to provide a set of procedures, architectures and components that will be implemented in an integrated SHM system. This will allow a significant reduction of the maintenance costs and a simplification-optimization of technical interventions, because the system will permit to pass from a fixed inspection plan on the basis of statistical hours of flights to a “condition maintenance based on diagnostic/prognostic methods”.
The Structural Health Monitoring with Embedded sensors (ESHM) approach on aeronautical structures can have a strong impact as a means of possibly revolutionizing the current structural monitoring, maintenance and design processes. While the OSAS platform elements do primarily target aircraft panels, the results could be used to study, produce and disseminate “best practices” and “highlights” for the use of CFRP materials in a variety of applications spanning multiple domains. Moreover though OSAS is mainly addressed to CFRP composite materials its embedded sensors architecture could be evaluated for an exploitation also addressed to other classes of materials like polymers and complex composites, even belonging to emerging technologies. Thus, the results could have the following cross-domain impacts:
• Reduce time to market despite the increasing contribution of embedded systems and software and their increasing size and complexity
• Increase the quality and reliability of products and services while providing novel functionalities to the user
• Improve cross-domain fertilisation
Aerospace, aviation, automotive and industrial SECTORS are application domains directly targeted by the OSAS architecture. The study of potential applications in these domains will boost cross-domain fertilisation through sharing of knowledge and best practices. OSAS will produce and disseminate “blueprints” detailing the use and benefits of CFRP components in multiple domains.
Dr. Dimitri Bofilios
INASCO
18 N. ZERVA ST.,
16675 - GLYFADA GREECE
+30 210 99 43 427
