C. Romessis

Bayesian Network Approach for Gas Path Fault Diagnosis

A method for solving the gas path analysis problem of jet engine diagnostics based on a probabilistic approach is presented. The method is materialized through the use of a Bayesian Belief Network (BBN). Building a BBN for gas turbine performance fault diagnosis requires information of a stochastic nature expressing the probability of whether a series of events occurred or not. This information can be extracted by a deterministic model and does not depend on hard to find flight data of different faulty operations of the engine. The diagnostic problem and the overall diagnostic procedure are first described. A detailed description of the way the diagnostic procedure is set-up, with focus on building the BBN from an engine performance model, follows. The case of a turbofan engine is used to evaluate the effectiveness of the method. Several simulated and benchmark fault case scenarios have been considered for this reason. The examined cases demonstrate that the proposed BBN-based diagnostic method composes a powerful tool. This work also shows that building a diagnostic tool, based on information provided by an engine performance model, is feasible and can be efficient as well.

Probabilistic neural networks for validation of on-board jet engine data

Abstract A method is presented for identification of faults in the readings of sensors used to monitor the performance and the condition of jet engines. Probabilistic neural networks are used to detect the presence and identify the location and magnitude of faults (biases) in sensor readings. The faults can be detected on sets comprising a limited number of instruments, typical of those available for on-board monitoring of jet engines. An engine performance model is used to support the constitution of a network. Training information is built using the model to produce data for a comprehensive set of healthy and faulty situations. The network performance in detecting and quantifying sensor faults is validated on a large number of fault cases, also generated by a model, which are used for testing the network and cover a wide range of conditions that can be encountered in practice. An engine, representative of current large civil engine designs (large bypass, partially mixed turbofan), serves as the test vehicle for demonstration of the way the method is materialized.

A Parametric Investigation of the Diagnostic Ability of Probabilistic Neural Networks on Turbofan Engines

Fault identification through the use of Artificial Neural Networks has become very popular recently. Probabilistic Neural Networks (PNN) is one of the architectures, which have mostly been investigated for gas turbine diagnostics. In this paper, the influence of parameters related to the structure and training on the diagnostic performance of a probabilistic Neural Network (PNN), is investigated. In particular, the parametric investigation examines the effect of the training set on the diagnostic performance of a PNN. The effect of noise level was also examined and found to be important. Another parameter examined is the severity of a fault, which was found to affect seriously the performance of the diagnostic PNN. Other parameters also examined are the effect of the operating conditions as well as the considered output parameters of the network. Guidelines useful for setting up this type of network, are derived.Copyright

Combining Classification Techniques With Kalman Filters for Aircraft Engine Diagnostics

A diagnostic method consisting of a combination of Kalman filters and Bayesian Belief Networks (BBN) is presented. A soft-constrained Kalman filter uses a priori information derived by a BBN at each time step, to derive estimations of the unknown health parameters. The resulting algorithm has improved identification capability in comparison to the stand alone Kalman filter. The paper focuses on the way of combining the information produced by the BBN with the Kalman filter. An extensive set of fault cases is used to test the method on a typical civil turbofan layout. The effectiveness of the method is thus demonstrated and its advantages over individual constituent methods are shown.© 2004 ASME

Fusion of Gas Turbines Diagnostic Inference: The Dempster-Schafer Approach

This paper proposes a fusion technique allowing the merge of conclusions provided by diagnostic methods that act independently for the detection of gas turbine faults. The proposed technique adopts the principles of Dempster-Schafer theory for the fusion of two diagnostic methods output; these are the method of Bayesian Belief Networks (BBN) and the method of Probabilistic Neural Networks (PNN). The proposed technique has been applied for the detection of thermodynamic as well as mechanical faults on gas turbines. First, the case of a turbofan engine of civil aviation is examined. The proposed technique allows the fusion of diagnostic inference on the presence of several faults of thermodynamic nature. Then the case of a radial and an axial compressor are examined, where several mechanical faults are deliberately implemented. In all cases, the effectiveness of the proposed fusion technique demonstrates that the merge of diagnostic information from different sources leads to better and safer diagnosis.Copyright

Setting Up of a Probabilistic Neural Network for Sensor Fault Detection Including Operation With Component Faults

The diagnostic ability of Probabilistic Neural Networks (PNN) for detecting sensor faults on gas turbines is examined. The structure and the features of a PNN, for sensor fault detection, are presented. It is shown that with the proposed formulation, a powerful tool for sensor fault identification is produced. A particular feature of the PNN produced is the ability to detect sensor faults even in the presence of engine component malfunction, as well as on deteriorated engines. In such situations, the size of bias that can be identified increases. The way to establish the limits of sensor bias that can be detected is presented along with results from application to test cases with realistic noise magnitudes. The diagnostic procedure proposed here is also supported by an engine performance model. The data used for setting up and testing the PNN are generated by such a model.Copyright

Detection of Gas Turbines Malfunctions From Emission Concentration Distributions

A method for detecting gas turbines malfunctions through engine emissions concentration plots is presented. The method is materialized through the use of a bank of Probabilistic Neural Networks (PNNs). The main idea comes from the fact that specific operating and health conditions of an engine lead to specific concentrations of emissions on the exhaust area. By comparison of an emission concentration plot with emission plots of known engine health conditions, diagnostic conclusions can be extracted. The stochastic nature of emission concentrations can be handled by PNNs, a specific type of Artificial Neural Networks which are known to be efficient probabilistic classifiers. The diagnostic problem and the overall diagnostic procedure are first described. A detailed description of the way the diagnostic procedure is set-up, with focus on building the PNNs, follows. The case of an operating family of turbofan engines is used to evaluate the effectiveness of the method. The examined case demonstrates that the proposed method can act as an additional tool on the existing methods for better and safer fault diagnosis.Copyright

Implementation of Stochastic Methods for Industrial Gas Turbine Fault Diagnosis

Implementation of stochastic diagnostic methods for diagnosis of sensor or component faults is presented. Two industrial gas turbines are considered as test cases, one twin and one single shaft arrangement. Methods based on Probabilistic Neural Networks (PNN) and Bayesian Belief Networks (BBN), are implemented. The ability for successful diagnosis is demonstrated on specific cases of sensor malfunctions, as well as on two types of compressor deterioration, fouling and variable vane mistuning. The examined diagnostic problem and the methods of PNN for sensor fault diagnosis and BBN for the diagnosis of component faults are first described. For each gas turbine case, the implementation of the diagnostic methods is shown and application to fault cases that occurred is presented. The effectiveness of the stochastic diagnostic methods demonstrates that they offer a powerful alternative diagnostic tool.© 2005 ASME