Applied Energy 79 (2004) 109–126APPLIED ENERGY www.elsevier.com/locate/apenergy Prospects for aero gas-turbine diagnostics: a review Luca Marinai *, Douglas Probert, Riti Singh Department of Power, Propulsion and Aerospace Engineering, Cranfield University, Bedford MK43 0AL, UK Accepted 21 October 2003 Available online 27 February 2004 Abstract Despite inflating unit-fuel costs, the long-term prospects for the aircraft industry remain buoyant. Nevertheless reducing direct operating-costs is crucial to ensure competitive advantages for airlines and manufacturers, and so effective advanced engine-condition monitoring methodologies are desirable. Hence gas-path diagnostic methods are reviewed and the specifications for such effective tools deduced, together with pertinent future prospects. Ó 2004 Elsevier Ltd. All rights reserved. Keywords: Performance; Diagnostics; Gas-path analysis; Aftermarket 1. Introduction At present, there are deep financial uncertainties in the civil aircraft market and therefore intense competition among airlines. Hence the development of advanced maintenance-techniques in order to reduce operating-costs [1,2]. Engine-related costs contribute a large fraction of the direct operating-costs (DOCs) of an aircraft, because the propulsion system requires a significant part of the overall maintenance effort that has to be expended for each aircraft – see Fig. 1. The world market for transportation by air is expanding, despite the difficulties and changes following the horrific terrorist attack on September 11th 2001 in New * Corresponding author. Tel.: +44-1234-750-111-526; fax: +44-1234-752-407. E-mail address: l.marinai.2001@cranfield.ac.uk (L. Marinai). 0306-2619/$ - see front matter Ó 2004 Elsevier Ltd. All rights reserved. doi:10.1016/j.apenergy.2003.10.005 / Applied Energy 79 (2004) 109–126 Abbreviations A/C aircraft AGARD Advisory Group for Aerospace Research and Development AI artificial intelligence AIAA American Institute of Aeronautics and Astronautics ANN artificial neural-network ASEE American Society of Engineering Education ASME American Society of Mechanical Engineers BBN Bayesian-belief network COMPASS condition-monitoring and performance analysis software system DOC direct operating-cost EKF extended Kalman-filter ES expert system FFBPNN feed-forward back-propagation neural-network FL fuzzy-logic FMPT fleet-management programme GA genetic-algorithm GPA gas-path analysis GT gas-turbine HPC high-pressure compressor HPT high-pressure turbine ICM influence coefficient matrix IEKF integrated extended Kalman-filter IFAC International Federation of Automatic Control IPC intermediate-pressure compressor IPT intermediate-pressure turbine ISABE International Society for Air-Breathing Engines KES knowledge-based engineering systems KF Kalman-filter LPT low-pressure turbine M number of measurements MCPHT maintenance cost per hour MFI multiple-fault isolation MLP multi-layer perceptron N number of performance parameters NASA National Aeronautics and Space Administration OEM original equipment manufacturer Pin inlet pressure Pout exit pressure PNN probabilistic neural-network RPK revenue arising from passenger-kilometres .110 L. Marinai et al. Economic growth is the primary drive for increased travel by air. considerable interest has been devoted to lowering operating-costs by introducing improvements in engine reliability. More recently. / Applied Energy 79 (2004) 109–126 111 RPM SAE SFC SFI Tin Tout WLS revolutions per minute Society of Automotive Engineers specific fuel consumption single-fault isolation inlet temperature exit temperature weighted least-squares Fig. York. integration of engineering and manufacturing functions reduced costs. . 2. and will be. Hence the present knock-on effect of severe competition among gas-turbine manufacturers. will increase at an average annual rate of 4. Marinai et al. In the current financial climate.3 trillion compared with the 3. established airlines are. world annual RPKs will reach 8. achieving extended ‘‘life on wing’’ and upgrading maintenance schedules [5]. By 2020. and. in response to increasingly severe cost-pressures and competition. Previously the business was driven by technological progress (such as the introduction of cooled blades or the move from pure jets to turbofans). measured in revenue arising in US dollars from passenger-kilometres (RPKs). In the 1970s and 1980s.3 trillion in 2000 – see Fig. it would appear that the new generation of lowcost airlines will prosper. driven even further to improve the efficiencies of their route networks and to use low-unit-cost aircraft [4]. Airbus [4] predicts that. during the next 20 years. such world traffic.L. 1.7%. Typical costs as fraction of a civil aircraftÕs DOC [3]. Nowadays the airlines demand is for high quality fleet-management and comprehensive engine-aftercare service. companies would have had to make compensating higher profits on the original equipment sale [5]. Consequently improvements concerning the in-service operations of engines have already had significant impacts on the business [2].112 L.g. These programmes provide engine maintenance on a flat rate per engine flight-hour basis. 25 years). this would have resulted in the partial or complete loss of the engine-manufacturersÕ aftermarket business. are increasingly being demanded. Among them. at present. New role for the gas-turbine after-sales market The gas-turbine business model has been based on revenue strictly related to the aftercare. Similarly the highly successful General ElectricÕs ÔMaintenance Cost per HoureÕ (MCPHe) contracts and Pratt and Whitney ÔFleet-management ProgrammeeÕ (FMPe) contracts offer longterm service agreements. / Applied Energy 79 (2004) 109–126 Fig. a key competitive advantage for manufacturers will be their understanding of this market and one consideration within this will be concerned with engine-condition monitoring methodologies. Marinai et al. based on an agreed prior to the engine delivery rate per engine flying hour. ÔPower by the HoureÕ (trade mark held by Rolls-Royce) type of contracts. which includes the capital cost plus a blend of financing and maintenance after the engineÕs sale. This transfers much of the technical risk from the airline to the gas-turbine manufacturer. The engine is paid for during the period when the aircraft is in the air and so producing revenue. so enabling airlines to accurately forecast operating-costs. The new business scenario could therefore be a win–win opportunity for airlines and manufacturers. Predicted air-world traffic growth [4]. In the previous after-sales maintenance scenario. Prolonging a gas-turbineÕs life could lead to a scenario in which the engine will not require a major service during the aircraftÕs life (e. partic- . In these circumstances. 2. reduce cost of ownership and improve asset utilization [6. 2.7]. So. L. efficiency and flow capacity) given a set of measurements (e. A recent update of gas-path diagnostics methodologies is contained in the Von Karman Institute lecture series 2003-01 on gas-turbine condition monitoring and fault diagnosis edited by Mathioudakis and Sieverding [10]. w the vector of environment and power-setting parameters. ð1Þ where z is the measurement vector. However accurate assessment is complicated by (i) only having relatively few measurements available. maintenance. Many pertinent tools have been devised during the last three decades and a critical review of the most used techniques and their applications now follows – see also Table 1 – highlighting similarities.g. As a consequence of progressive performance-losses. b the sensor bias and hð Þ a vector valued non-linear function: hð Þ is provided by the simulation program [9]. 3. m the measurement noise vector. Marinai et al. The performance of an aero-engine deteriorates over time as a consequence of its componentsÕ degradation. Therefore monitoring and maintenance techniques are employed to ensure that the gas-turbine operates both cost effectively and safely. temperatures. overhaul or replacement of the engine or one of its components. noise and biases). Engine gas-path diagnoses have been recognised. Linear gas-path analysis with ICM inversion This is based on the assumption that the changes in the (independent) healthparameters are relatively small and the set of governing equations can be linearized around a given steady-state operating point. An engine gas-path diagnostic process calculates changes in the magnitudes of the component performance parameters (e. Gas-path diagnostic methodologies A survey of pertinent techniques. differences and limitations. developed since the pioneer Urban [8] first explored these problems is now presented. Deterioration can affect factors such as thrust (or power) and specific fuel-consumption (SFC). shaft speed and fuel flow) through the engine. as important means for making more informed decisions on the usage. for some time. and (ii) errors in the measurements (e. / Applied Energy 79 (2004) 109–126 113 ular consideration should be given to gas-path diagnostics that plays a primary role in an aero-engine performance oriented business. 3. wÞ þ m þ b.g. The identification of the exact component(s) responsible for the performance loss facilitates the choice of the recovery action to be undertaken.1.g. operation of the engine can become cost ineffective or even unsafe. x the performance parameters vector. The relationship between measurements and performance parameters can be expressed analytically as follows: z ¼ hðx. z ¼ Hx ð2Þ . due to uncertainties. These linearized equations can be expressed in matrix form. pressures. M measurements Single/multiple fault(s) Smearing versus concentration Difficulty and dependence on training/tuning Artificial-intelligence based Computational burden (slow in recall mode) Model-free High dimensionality problems Data-fusion capability Black-box not observable Good accuracy in predefined ranges only Expert knowledge capability On-wing Linear X Non-linear GPA with ICM inversion Non-linear Linear Kalmanfilter Linear X X X M <N MFI Smearing Prior knowledge Linear WLS Non-linear Kalmanfilter Non-linear Non-linear model-based with GA Non-linear Artificial neural-networks Bayesian-belief networks Expert systems Fuzzy-logic L.114 Table 1 Summary of GPA methodologies (X means that the methodology involves that feature) Strategy involved Methodology Linear GPA with ICM inversion Linear/non-linear model Small changes of health parameters Measurements of random noise Bias N parameters. Marinai et al. / Applied Energy 79 (2004) 109–126 Linear X X X M <N MFI Smearing Prior knowledge Non-linear Non-linear Non-linear X X M <N MFI Smearing Prior knowledge X X M <N X X M <N X X X M <N X X M <N M PN MFI M PN MFI M <N SFI/limited MFI SFI/limited MFI SFI/limited MFI SFI/limited MFI SFI/limited MFI Concentration Concentration Concentration Concentration Concentration Number of string assignment X X X X X X X X Long training and data selection X Effort in gathering info for setting-up X X X X X X X X X X X Rules explosion X X X X X X X X X X X X X X . L. These have been adopted by major OEMs such as Rolls-Royce. • It relies on the assumption of linearity and is only acceptable for very small ranges of values of the influential parameters about the operating condition. This formulation of the problem leads to a simple solution: x ¼ H 1z ð3Þ The matrix H 1 is referred to as fault-coefficient matrix (FCM). N 6 M). Pratt and Whitney and General Electric.g. • It does not deal with sensor noise or bias. Non-linear gas-path analysis with ICM inversion One way of improving the accuracies of the predictions is to try to solve the nonlinear relationship between the considered health-parameters and measurements using an iterative method as described by Escher [12]. Several other authors. Essentially. Otherwise estimation techniques. Marinai et al. From the results calculated. have been used. via this approach.3. EHM and ADEM). estimation techniques like the Kalman-filters (KF) method. the gas-path diagnostic tool currently used in Rolls-Royce is based on a modified version of the KF . Similarly. a new ICM is generated and this process is recursively repeated until the solution converges to a fixed limit. have described applications of this approach and highlighted the following limitations: • The method requires many pertinent measurements for the analysis.2. / Applied Energy 79 (2004) 109–126 115 The matrix H has been referred to by several names. 3. This method is based on the assumption that the ICM is invertible and the measurements are free of noise. discussed later. an ICM is generated taking into account a small deterioration in the engine-componentÕs performance. 3. To overcome the limitations that face the two techniques. when only a small number of measurements is available (i. through Hamilton Standards. Kalman-filter and weighted least-squares based GPA: linear approach Pratt and Whitney. exchange-rate table and influence-coefficient matrix (ICM) being two of the most popular. This consideration led to the development of non-linear GPA. TEAMIII. should be used. The linear approximation is employed recursively and an exact solution obtained by the Newton–Raphson technique. The assumption of linearity becomes increasingly false when deteriorations cause the engine to operate further away from the condition for which the matrix was calculated. lack of observability) and in the presence of measurement uncertainties (e. noise and biases). The ICM is then inverted to calculate the vector of change in the engine-componentÕs performanceparameters. have been pioneering the implementation of a KF based method. Several adaptations have been designed to cope with some of the filterÕs limitations [13] and they are integrated into the available software (MAPIII.e. and their variants. The inversion of the ICM requires that the number of performance parameters is less than or equal to the number of measurements (i. weighted least-squares (WLSs) approach. such as Passalacque [11] and Escher [12].e. which is theoretically similar to those of non-linear model-based methods described earlier. The most non-linear squares estimation algorithms require a choice between an optimal solution and a recursive formulation.20]: the KF algorithm tends to ‘‘smear’’ the fault over many components. the WLSs estimation has been applied [17] and implemented within a diagnostic tool (TEMPER). a non-linear version of the KF can be used to try to model more accurately the engine behaviour. The solution is obtained when an objective-function. As far as General ElectricÕs approach is concerned. The diagnostic system (COMPASS). • Non-linearity: the error due to the assumed approximation to a linear model may not be negligible. If recursivity is a paramount requirement. Kalman-filter based GPA: non-linear approach If the effect of non-linearity on the estimation accuracy is ascertained. then optimality is compromised. achieves its minimum value: a schematic diagram of the diagnostics strategy is shown in Fig. Non-linear model-based optimal estimation by using genetic-algorithms A genetic-algorithm (GA) based diagnostic method was devised at Cranfield [18]. The GA is applied as an effective optimization tool to obtain a set of engine-component parameters that produce a set of predicted dependent parameters through a non-linear gas-turbine model that leads to predictions which best match the measurements. this leads to a low accuracy estimation. The methodology is based on non-linear modelling and optimal-estimation theory. From a practical point-of-view. 3. has then been used for ‘‘on-wing’’ applications. according to Zedda [18]. 3. Concentration on the faulty components may be difficult. developed for test-cell diagnostics of aero-engines.5.4. which is a measure of difference between predicted and measured parameters. the so called ÔConcentratorÕ applied to a linear GPA [14–16]. Limitations in the application of KF and WLS techniques to GPA are. / Applied Energy 79 (2004) 109–126 technique. It is a model-based approach. Marinai et al. A third possible solution has been discussed by Zedda [18]: the proposed estimation technique splits the problem of cost-function minimization into a linear first step and a non-linear second step by defining new first step states that are non-linear combinations of the unknown states. However. • Prior knowledge and tuning needed: the choice of the covariance matrix (tuning) is often arbitrary. The previously-mentioned applications are able to estimate both component performance changes and sensor biases. The problem is undetermined and the KF solution is a maximum likelihood one: an estimated state vector with a larger number of fault-affected elements is more likely. The most commonly-used filtering techniques are the extended Kalman-filter (EKF) and the iterated extended Kalman-filter (IEKF).116 L. The procedure estimates the performance parameters expressing the fault . 3. • The ‘‘smearing’’ effect [19. it can be shown that both produce biased and sub-optimal estimates due to the linearization of the cost functions. This makes the method difficult to use. Some of these limitations have been overtaken by further developments of the technique. The method suffers from the following limitations [18]: • The methodology is more computationally burdensome than classic estimationtechniques. Marinai et al. The GA uses an accurate non-linear steady-state model of the engineÕs behaviour. The diagnostics strategy uses an engine-performance model that is based on characteristics of the gas-turbine components [18]. / Applied Energy 79 (2004) 109–126 117 Fig. The only statistical assumption required by the technique concerns the measurement noise and the maximum allowed number of faulty sensors and engine-components. active awareness of these issues is necessary for the correct utilization of the technique. • Care must be taken when using the GA in assigning the number of strings. Even though the rule for the assignment of the number of strings for different fault classes can easily be established by trial-and-error and the achieved accuracy is not a strong function of the rule of assignment itself. as discussed later. and requires a trained person for its worthwhile operation.L. • Although multiple-faults can be detected. In fact. 3. the theory here described has been implemented . condition of the engine-components in the presence of measurement noise and biases. The measurement uncertainty is supposed to affect even the parameter setting and the operating condition. Estimation is performed through an Ôad hocÕ GA. the technique is limited to four parameters experiencing simultaneous deteriorations. ANNs are unable to perform creditably outside the range of data to which they have been exposed: this implies that a massive amount of data from encountered and foreseeable fault conditions of operation would be required in each ANN development. The hybrid model improves the accuracy. An ANN consists of parallel distributed processors able to store knowledge as experience and make it available for use.6. 3. This could mean. retraining after a machine overhaul.118 L. ANNs require retraining when machine operating conditions change. The first stage uses response surfaces for computing objective-functions to increase the exploration potential of the search space. The purpose of the learning phase is to determine the NN parameters. It was applied to a RB199 engine and showed good results. / Applied Energy 79 (2004) 109–126 for development engines where many measurements are available. The multi-layer perceptron (MLP) with back-propagation training is the most common architecture used for GPA purposes.20]: it provided a high level of accuracy. Further developments of the method have led to the study discussed by Sampath [24] in which an hybrid approach has been adopted for the intercooled recuperated WR21 engine. Generally. The use of ANNs in gas-path diagnostics experiences the following limitations: • Like other AI tools. ANNs are trained to map inputs to outputs via a non-linear relationship in a framework that loosely mimics the learning process performed by the brain. Artificial neural-network based GPA Artificial neural-networks (ANNs) have been investigated extensively for use in fault diagnoses. [21. It is also called the feed-forward back-propagation neural-network (FFBPNN). reliability and consistency of the results obtained. thereby gaining by reducing the computational burden. The application of evolution strategy also has been investigated by Sampath et al. which will enable the network to function properly in the operating phase. though this depends on the network type. This fault-diagnostics model has been integrated with a nested neuralnetwork to form a hybrid diagnostics model. size and the amount of training data. Marinai et al. This approach is suitable for problems where only limited instrumentation is available. It was applied to a three-spool military turbofan engine RB199 and a two-spool low by-pass military turbofan engine EJ200 [19. Gulati et al. The nested neural-network is employed as a pre-processor or filter to reduce the number of fault classes to be explored by the GA-based diagnostics model. The third stage uses the elitist model concept applied to a GA to preserve the accuracy of the solution in the face of randomness. The second stage uses the heuristics modification of genetics algorithm parameters through a master-slave type configuration. .22] combined a multiple-point diagnostic approach [23] and a GA approach to produce a GA-based multiple operating-point analysis (MOPA) method for gasturbine fault diagnostics. [25]. while reducing the computational load. • Training times are long. Sampath [24] described a diagnostics model that operates in three distinct stages. for instance. the NN operates in two phases – a learning phase and an operating phase. A wide range of different neural-networks has been implemented in gas-turbine engine-fault detection. The authors suggested the use of different networks to isolate the sensor-and-component faults as this would provide a better result than using a single network for the combined task. • It is sometimes difficult to provide the confidence level associated with the output result. as reported by Zedda and Singh [26] is the requirement for a large amount of training data and the long time required to train the networks. the nodes can represent anything. This approach is interesting considering the fact that. [29]. A belief network is a graphical representation of a probability distribution that represents the cause and effect relationships among predisposing factors. diagnostics and prognostics capabilities for the engines. a fault or some intermediate value. Marinai et al. Each node contains a conditional probability distribution that describes the relationship between the node and the parents of that node. an observation. An application of FFBBNNs to gas-turbine diagnosis is discussed by Torella and Lombardo [28].L. it is a mathematically correct way of combining probability estimates even if they come from multiple sources. even for healthy engines. As such. Trained networks with frozen network-weights would require retraining the networks if the engine undergoes an overhaul. the diagnostics error is bound to increase unless an alternative means of data correction is devised [27]. Eustace and Merrington [32] applied a probabilistic neural-network (PNN) to diagnose faults in any engine within a fleet of 130 GE low-bypass F404 military engines. [30] introduced a hybrid neural-network where part of the model was replaced by influence coefficients: they reported that the accuracy of such a network was favourable compared with a back-propagation net and KF approach. This has been further extended by Ogaji [27] to generate a cascaded network to isolate component-and-sensor faults. The authors used a statistical correlation technique to select five out of eight available engine-monitoring parameters as inputs to the network. faults and symptoms. Zedda and Singh [26] introduced a modular neural-network system to tackle large-scale diagnostic problems and applied it to predict the behaviours of a Garrett TFE 1042 engine. Bayesian-belief network based GPA Bayesian reasoning is based upon formal probability theory. In fact. Volponi et al. that need to be diagnosed. A network architecture to perform sensor and component-fault diagnosis step-by-step using multiple neural-networks is described by Kanelopoulos et al. • Another major drawback of neural-networks. the values of the parameters vary from engine to engine. The list of possible states . diagnosis and accommodation. Green [31] discussed the need to incorporate an ANN with other AI techniques to perform the estimations of active life. 3. A network consists of nodes representing variables and arcs for the probabilistic dependencies between these variables.7. / Applied Energy 79 (2004) 109–126 119 • Its deficiency in providing descriptive results: there is no way of accessing the neural-networkÕs ‘‘reasoning’’: it is only possible to inspect the predictions it makes. increases. • As the number of engine operating points. Expert systems The development of expert systems (ESs) has been one of the key breakthroughs of artificial intelligence (AI). Marinai et al. Romessis et al. so as to be able to provide an Fig. 4. The independent parameters are designated as the parent nodes and the dependent parameters as the child nodes. A typical layout of a BBN is shown in Fig. 3. The relationships between these have been defined through the links and each link has a probability associated with it. / Applied Energy 79 (2004) 109–126 for each node must be mutually exclusive and collectively exhaustive. The use of BBNs in gas-path diagnostics experiences the following limitations: • Sensor bias is not dealt with in the BBN-based studies reviewed here. . 4.8. Kadamb [33] aimed at diagnosing relatively large deteriorations of up to four performance parameters relying on a fully non-linear model. [35] provided a thorough description of a Bayesian-belief network (BBN) for turbofan engine diagnostics. • Substantial time and effort are required to gather the information needed for setting it up.120 L. The typical form of an ES encompasses expert knowledge in a pattern-matching process. The links between a parent node and a child node are established only if that particular child node (measurement) is affected by the fault. Typical BBN layout [33]. [34] and Mathioudakis et al. 48] presented a linearized fuzzy-logic intelligent process for gas-turbine fault-isolation. The ES process was coded by means of a belief network. This is often achieved by means of rules. The input and output are discretized and this enables complex mathematical problems to be simplified [43]. taking into account faults in gas-path components. A model-based fuzzy-logic for sensor-fault accommodation was presented by [44] and a fuzzy ruleand-case based expert-system for turbo-machinery diagnosis was developed by Siu et al. The goal of the process is to identify the faulty component without quantifying the deterioration: the output fuzzy sets are not further decomposed (defuzzified) using linguistic variable. Ogaji et al. The system was first designed to isolate and quantify single-component faults. Ganguli [47. The use of fuzzy-logic technology in gas-path diagnostics experiences the following limitations: • The model-free feature. The method uses rules developed from a model of performance-influence coefficients via a linearized approach to isolate five specified engine-faults while accounting for measurement uncertainties along the gas-path. regards diagnosis and troubleshooting of GTs from a probabilistic relationship between failures (causes) and symptoms (behaviour) because an observed symptom could be the result of several different causes having several degrees of probabilities and vice versa. This led Marinai et al.9. [6. [45]. Results from tests carried out on the IPC showed good fault-estimation capabilities. Marinai et al. A similar accuracy level was achieved by both systems. and rotor or oil subsystems. [6.7]. comes with the restrictions that a fuzzy system does not admit modelbased proofs of stability and robustness. Recent studies have been dedicated to the implementation of ES for fault diagnostics [36–41] in cases in which only a qualitative answer is sufficient. the input) via an inference engine. pattern recognition) in the presence of uncertainty. thereby highlighting the potentialities of a fuzzy-logic approach. 3.e. a non-linear model-based process was devised by Marinai et al. Fuzzy-logic based diagnostics Recently these methodologies have been devised by taking advantage of a convenient way of mapping an input space to an output space (i. discussed by Torella and Torella [42]. Karvounis and Frith [50] described a model-based process for assessing the overall condition for the T700 engine and determining when to service it. An ulterior fuzzy-logic process was developed [49] for helicopter rotor-system fault isolation. it uses FL technology for GPA of a three-spool turbofan engine.7] to the implementation of a process capable of multiplefault isolation. An example of an ES application. They also have indicated how the quality of the solution depends on the quality and the number of the rules.e. / Applied Energy 79 (2004) 109–126 121 interpretation of a new situation (i. Most recently. instrumentation. A fuzzy-logic formulation is implemented to automate the detection and provide an end-of-flight estimate for future prognostics. [51] carried out a comparison of the ANN-based method and the FLbased method. Tang [46] described a process based on fuzzy-logic and a neural-network for engine diagnosis.L. . that allows data-fusion and gains computational time reductions. ESs. as well as AI-based methods. A distinction can be made here [52] between techniques more suitable for estimating (i) gradual deteriorations and others for (ii) rapid deteriorations. Moreover AI-based algorithms can be excessively time-consuming. or in the training phase as for an ANN. We refer to such methods as MFI (multiple-fault isolation) and SFI (single-fault isolation) ones respectively. Other techniques. the tendency to smear the faults over a large number of the engineÕs components and sensors) that estimation techniques suffer from. fuzzy systems are unable to approximate creditably outside the range of data to which they have been exposed: this implies that a massive amount of data from encountered and foreseeable fault conditions of operation would be required in their development. Besides they are not able to deal with measurement uncertainty. An appropriate learning algorithm can restrict the number of rules required [53]. Comparison of the methodologies Ten of the key techniques. are described in Table 1.122 L. The former implies that all the enginecomponents (whose shifts in performance we are estimating) deteriorate slowly. / Applied Energy 79 (2004) 109–126 • Like other AI tools.e. such as a diagnostic system. but on the contrary have good ÔconcentrationÕ capabilities to identify the faulty component. BBNs and FL ones are referred to as model-free systems. Nevertheless rules-reduction strategies can be adopted. in fault diagnostics. used in a wide range of applications. AI-based techniques do not show the ÔsmearingÕ problem (i. because they approximate all the possible solutions with a limited number of cases used to train the system. Estimation techniques. whereas the latter implies a rapid trend shift. bearing in mind that. • Fuzzy systems face the problem that the number of rules increases according to the complexity of the process that is being approximated. Algorithms based on ICM inversion are suitable only if the number of measurements is more than the number of health parameters. AI-based methods are more suitable for SFI problems. This model-free . can deal with diagnostics with only a few measurements. The extension to all the possible combinations (even in a limited search-space) is theoretically possible. precision in isolating the faulty component is more significant than the actual accuracy of the estimate. Similarly AI-based methods require particular care during the set-up phase. they need a large number of rules for systems with many non-linearly-related inputs and outputs. The inadequacy of this model has led to non-linear methods being devised. but extremely burdensome computationally and highly inconvenient. Marinai et al. ANNs. • The achieved accuracy is the result of a trade-off between the computational speed and burden. both in the actual calculation as in the case of a GA. As far as fuzzy systems are concerned. 4. probably due to a single entity (or perhaps two) going awry. Some of the approaches are based on the assumption that the changes in the health-parameters are relatively small and the set of governing equations can be linearized. Estimation techniques require prior information and the solution can be dramatically affected by this choice. such as WLE and Fuzzy-logic approaches are particularly suitable for dealing with measurement uncertainties. • Capable of detecting with reasonable accuracy significant changes in performance. Some of the techniques are complementary and each has its own advantages and limitations. they cannot be accurate out of the range of variability for which they have been trained or set-up. in an industry in which. Although fuzzy systems have only recently appeared as an advanced gas-path diagnostics methodology with quantitative capability. that are typically suitable for these circumstances. Marinai et al. the need for an effective approach cannot be ignored. • Fast in undertaking diagnosis for on-wing applications. This . • Easily-satisfied computational requirements. it has been recognised that the capability of making diagnoses on-line for on-wing applications requires a blend of fast calculation and the ability to achieve a worthwhile solution with only a limited number of measurements which are affected by a considerable level of noise. The techniques. difficulties and dependences for the setting-up parameters. 5. BBN and FL approaches have the additional quality that they can be used to encompass expert knowledge in the system. • Exempt from training and tuning uncertainties. significant benefits have already been achieved from their pattern-recognition model and their intrinsic capability of dealing with uncertainty.L. but comes with the limitation that no model-based proofs of stability and robustness are possible. there is no single technique which addresses all the issues. Besides. • Capable of data-fusion. Hence. Numerous tests must be conducted to validate these techniques. are indicated in Table 1. / Applied Energy 79 (2004) 109–126 123 feature is responsible for data-fusion capability and gains in computational burden. Besides. Conclusions Engine-fault diagnostics is a mature technology. the core of the businesses is based on aero-thermal performance. A summary of the available literature (see Table 1) suggests the requirements for an advanced diagnostics process should be: • Based on a non-linear model. • Able to deal with random noise in the measurements and sensor bias. • Able to incorporate expert knowledge. ES. In consideration of the primary role that gas-path diagnostics plays. • Competent to make a worthwhile diagnosis using only few measurements (N > M). Nevertheless no technique provides a satisfactory and complete answer to all the problems: the limitations of the most popular approaches have been described earlier. Eventually. • Free from a lack of comprehensibility due to black-box behaviour. it would be interesting to try and combine more than one technique to offset the limitations of one with the advantages of another within a combined engine-health monitoring (COEHM) scheme [5]. • Designed specifically for SFI or MFI. • Possess a ÔconcentrationÕ capability on the actual fault. [51]. Langenhagen. USA. Advances and opportunities in gas-path diagnostics. PhD Thesis. Mathematical methods for relative engine-performance diagnostics. In: ISABE-2003-1149. Marinai et al. so further increasing the market value of these technologies. Belgium. Future systems are expected to offer a high level of diagnostic competence. ISABE 2003-1008. XVI International Symposium on air-breathing engines. Institution of Mechanical Engineers. 1974. Civil aero gas-turbines: strategy and technology. Gas-turbine engine and sensor fault diagnosis. J Eng Gas-turbines Power 1994. [5] Singh R. 2003. [3] Rupp O. Sieverding CH. [13] Urban LA. 1969. [11] Passalacque J. 2003. School of Mechanical Engineering. On the other hand. School of Mechanical Engineering. COMPAS: a generalized ground-based monitoring system. Curnock B. RollsRoyce. Printed in France. 13–17 January 2003. [14] Provost MJ. Germany.124 L. [9] Li IY. integrated to a prognostic capability including life-cycle management and perhaps economic interactions. Hannover GmbH. 101. 1999. April 2001.A. Probert D.116:82–9. Cleveland. Brussels. 1995. Cranfield University. The use of optimal estimation techniques in the analysis of gas-turbines. PhD thesis. SAE 1992. Global-market forecast 2002–2000. Description of automatic gas-turbine engine-trends diagnostic system.S. 1988. Symposium on Gas-turbine Operation and Maintenance. References [1] Cribbes TD. International Gas-turbine Congress 2003 Tokyo. One is to use critically the results from different methods to work in concert with one another. Fuzzy-logic-based diagnostic process for turbofan engines. exploiting a potential synergistic effect. AGARD-CP-449. ChairmanÕs Address.216(A5):363–77. [8] Urban LA. November 2–7. [12] Escher PC. PhD Thesis. Engine-condition Monitoring. [2] Singh R. Performance technical report. [17] Doel DL. 1987. In: TS-005. IMechE J – Power Energy 2002. Detection and prediction of theperformance deterioration of a turbofan engine. Ó AIRBUS S. 2002. [7] Marinai L. [4] Airbus SAS. [16] Provost MJ. [10] Mathioudakis K. Kalman-filtering for performance engineering. MTU Maintenance. showed that a simple average among the results usually improves upon the result of a single technique. The second way is to design hybrid systems [24]. London.. School of Mechanical Engineering. [18] Zedda M. Singh R. industry is showing a major interest in engine-health monitoring and prognostics (HMP) schemes. Singh R. 2002. 31 August–5 September 2003. Aerosp Eng 1997. Cranfield University. Transaction Journal of Aerospace. Gas-turbine Condition Monitoring & Fault Diagnosis. [15] Provost MJ. National Research Council of Canada. Changes in engine-maintenance management. 1992. Derby. Performance analysis based gas-turbine diagnostics: a review. OH. / Applied Energy 79 (2004) 109–126 task can be thought of in two ways. . USA: Hamilton Standard Division of United Aircraft Windsor Locks. vol. Gas-turbine engine-parameter interrelationships. section 1. UK. Aerospace Division.17(12):7–9. Von Karman Institute Lecture Series 2003-01. 1995. [6] Marinai L. by comparing results from 40 test cases from a method based on fuzzy-logic and one based on an ANN. Cranfield University. technical paper 922048. Volponi AJ. Pythia: an object-oriented gas-path analysis computer program for general applications. Curnock B. TEMPER: a gas-path analysis tool for commercial jet-engines. Ogaji et al. The future challenge is to design effective combined engine-health monitoring and prognostic (COEHMP) procedures. Singh R. Gas-turbine engine and sensor fault diagnosis. Zedda M. ASME 2000-GT-547. Singh R. [37] DePold R. [40] Torella G. Fault diagnostics for an advanced cycle marine gas-turbine using Genetic-Algorithms. Multiple operating-point analysis using Genetic-Algorithm optimisation for gas-turbine diagnostics. June 3–6. Singh R. International Gas-turbine & Aero-engine Congress & Exhibition. [29] Kanelopoulos K. Stamatis A. 19 July 2000. USA. Bayesian belief network for aero gas-turbine module and system-fault isolation. [36] Ali M. Setting-up a belief network for turbofan diagnosis with the aid of an engine-performance model. Florence. Mathioudakis K. An Expert System for Fault Diagnosis in a Space-Shuttle Main Engine 26th AIAA/ SAE/ASME/ASEE Joint Propulsion Conference. Stamatis A. [28] Torella G. ISABE-2001-1139. June 8–13. Merrington G. The Gas-turbine and Aero-engine Congress and Exposition. 13–15 July. School of Engineering. [32] Eustace R. Incorporating neural-networks into gas-turbine performance diagnostics. Papailiou K.957085. ASME 97-GT-35. July 16–18. J Eng Gas-Turbines Power 1999:121–612. In: Seventh International Conference on Knowledge-Based Intelligent Information & Engineering Systems – KES. OH. Brussels. PhD Thesis. Probabilistic methods for gas-turbine fault diagnostics. 13–17 January 2003. [31] Green A. Mathioudakis K. 3–5 September 2003. Oxford. Utilization of neural-networks for gas-turbine engines. Romessis C. AIAA 2000-3716. 1993. September 20–24. XII ISABE 95-703. AIAA 99-253020. The role for expert systems in commercial gas-turbine engine monitoring. [24] Sampath S. Fault diagnosis of a turbofan engine using neural-networks: a quantitative approach. USA. [33] Kadamb A. School of Engineering. ISABE-2001-1032. Belgium. Florida. USA. Cranfield University. Alabama. AIAA-901890. Cranfield University. XII ISABE 1993. [35] Mathioudakis K. Singh R. 1990. 24 June 1999. Brussels. MSc Thesis. Mathioudakis K. 1999. In: Proceedings of the ASME Turbo Expo. USA. [23] Stamatis A. 2001. [34] Romessis C. The use of expert systems for gas-turbine diagnostics and maintenance. [20] Zedda M. School of Engineering. 1998. or ISABE 89-7133. 2000. 6–9 July 1997. [21] Gulati A. Cleveland. Tokyo. Amsterdam. In: 34th AIAA/ASME/SAE/ASEE Joint Propulsion Conference & Exhibit. Gas-turbine engine and sensor fault diagnosis using optimization techniques. FL. Orlando. PhD Thesis. 1997. Japan. Torella G. Jet engine-fault detection with discrete operating points gas-path analysis. 35th AIAA/ASME/SAE/ASEE Joint Propulsion Conference and Exhibit. Singh R. [30] Volponi A. Gupta U. ISABE 99-7238. Taylor D. Marinai et al.17-Huntsville.L. 2003. Gass D. The use of Kalman-filter and neural-network methodologies in gas-turbine performance diagnostics: a comparative study. 2003. Venturini M. Marinai L.7(6):1043–8. Stamatis A. Los Angeles. Daguang C. [22] Gulati A. J Propulsion 1991. 2001. Fault diagnosis of fleet engines using neural-networks. Italy. [25] Sampath S. Gas-turbine Condition Monitoring & Fault Diagnosis. [26] Zedda M. USA. 1995. 2003. 36th AIAA/ASME/SAE/ASEE Joint propulsion conference and exhibition. Artificial-intelligence for real-time diagnostics and prognostics of gas-turbine engines. 13th ISABE. AIAA 98-3602. Belgium. [27] Ogaji S. Advanced gas-path fault diagnostics for stationary gas-turbines. Application of expert systems and neural-networks to gas-turbine prognostics and diagnostics. . Orlando. ISABE 93-7133. 2003. The Netherlands. Singh R. WA. [39] Spina PR. AIAA 97-2899 Conference Seattle. / Applied Energy 79 (2004) 109–126 125 [19] Zedda M. [38] Doel DL. Berios G. CA. Ogaji SOT. Lombardo G. Evolution strategy applied to aero-engine-fault diagnosis. USA. Von Karman Institute Lecture Series 2003-01. 2002. Expert Systems for the Simulation of Turbofan Engines. Ganguli R. Cranfield University. United Kingdom. DePold H. Gas-turbine engine and sensor multiple operating-point analysis using optimisation techniques. June 1990. Comparison of artificial neural-network and fuzzy-logic approaches to engine diagnostics. LA. [52] Volponi A. 1997. 556–63. Application of fuzzy-logic for fault isolation of jet engines. Brussels. AIAA 99-2533. 2003. Larkin L. 1997. 1997. Frith P. ASME 97-GT-222. / Applied Energy 79 (2004) 109–126 [41] Torella G. USA: NASA CR-1432. Sampath S. [48] Ganguli R. [47] Ganguli R. A fuzzy-logic intelligent system for gas-turbine module and system fault isolation. Singh R. [51] Ogaji SOT. DC. New Jersey: Prentice Hall. 1969. Gas-turbine Condition Monitoring & Fault Diagnosis. ISABE 97-7148. 1999. Extending gas-path analysis coverage for other fault conditions. Toward a theory of fuzzy systems.40(12):2373–81. A practical intelligent system for condition monitoring and fault diagnosis of jet engines. June 20–24. [53] Kosko B. Marinai L. Health monitoring of a helicopter rotor in forward flight using fuzzy-logic. 2001. 2001. Shen Q. [50] Karvounis G. Probabilistic expert systems for the diagnostics and trouble-shooting of gasturbine apparatuses. In: ISABE-2003-1232. Int J Eng Intelligence Syst. [43] Zadeh LA. vol. 1997. 1999. In: SAFEPROCESSÕ97. Kerr L. ISABE2001-1112. USA. 31 August–5 September 2003. Washington. p. Marinai et al. Torella R. [46] Tang G. California. submitted for publication. TMDOCTOR: a fuzzy rule. [42] Torella G. XVI International Symposium on air-breathing engines.and case-based expert system for turbomachinery diagnosis.126 L. 13– 17 January. Cleveland. Glossary Bias: Measurement systematic error Computational burden: Negative property of excessively time consuming software Health parameters: Engine-componentsÕ efficiency and flow capacity as indicative of their health condition Rules explosion: Number of rules in a fuzzy system increases according to the complexity of the process that is being approximated Smearing: Tendency to smear the faults over a large number of the engineÕs components and sensors String: A coded potential solution of the GA String assignment: The process of assigning a string in a GA setting-up procedure . Von Karman Institute Lecture Series 2003-01. [49] Ganguli R. Expert systems and neural-networks for fault isolation in gas-turbines. [44] Healy A. 35th AIAA/ASME/SAE/ASEE Joint Propulsion Conference. [45] Siu C. Automated detection of engine heath using hybrid model-based and fuzzy-logic approach. In: AIAA 99-2942. OH. ASME 2001-GT-0013. Fuzzy engineering. Belgium. 1. Milne R. Model-based fuzzy-logic sensor fault accommodation. AIAA J 2002.
Report "Prospects for Aero Gas-turbine Diagnostics a Review"