An Integrated Principal Component Analysis, Artificial Neural Network and Gas Path Analysis Approach for Multi-Component Fault Diagnostics of Gas Turbine Engines

Gas path diagnostics is a key aspect of the engine health monitoring (EHM) process that aims to detect, identify and predict engine component faults, using information from installed sensors, in order to guide maintenance action, maintain engine efficiency and prevent catastrophic failures. To achieve high prediction accuracies, current data-derived diagnostic models tend to be engine specific while the model-based methods are known to be time-consuming, especially for complex engine configurations. This paper proposes an integrated approach for accurate and accelerated isolation and prediction of multiple-degraded gas turbine component faults that comprises 3 steps — feature extraction using the Principal Component Analysis (PCA), machine learning classification with a multi-layer perceptron, artificial neural network (MLP-ANN) and model-based fault prediction via the non-linear Gas Path Analysis (GPA) technique. In this hybrid approach, the PCA first transforms the measurement fault signature into a fault-feature domain, which becomes an input to the multi-label ANN classifier used to isolate the potential faulty components. The non-linear GPA finally quantifies the magnitude of degradation that produced the recorded fault signature. Once trained and validated, the PCA-ANN model is deployed as part of the data processing mechanism prior to the actual GPA calculation. This method was assessed and validated using the thermodynamic performance model of a 2-shaft, high-bypass ratio, turbofan engine. For training and testing the PCA-ANN classifier, a total of 28,000 final samples for 14 measurement parameters, each averaged from 10 data points with Gaussian noise of zero mean and unit standard deviation, and implanted with single-, double- and triple-component fault cases of various magnitude, were generated by steady-state performance simulation of the engine model at its reference operating condition. Correlation analysis of this data set revealed the optimum sensor subset to be used for multi-component diagnostics. A quantitative analysis of the PCA-ANN fault isolation on the test set produced a classification accuracy of 96.6% and performed better on all metrics, compared to other multi-label classification algorithms. Finally, the proposed integrated approach achieved an average of 94.35% reduction in processing time, when compared to the conventional non-linear GPA by component-fault-cases (CFCs), while predicting implanted faults to the same accuracy.