Engine Diagnostics in the Eyes of Machine Learning

Abstract

Machine learning (ML) is a data-driven approach to discovering patterns and knowledge, and it is different from the physics-based approach, which uses the principles of physics to describe a phenomenon. Physics-based approach has dominated the field of engine diagnostics because of the maturity of scientific and engineering knowledge embodied in the design and manufacturing of the engine and its components. Nevertheless, development of ML techniques has accelerated in the last three decades, and the techniques can potentially lower development time and are applicable to a wide variety of industries. This paper examines some of the most commonly cited ML techniques for handling numerical data and applies them to a gas turbine engine diagnostic problem. The diagnostic problem is to isolate the symptom of engine performance degradations to a root-cause fault or failure. This fault isolation problem is a type of classification problem in the ML world. A hypothetical engine model for commercial airplanes is used in simulation to create a standard dataset. This dataset is then used by all of the selected techniques. The results from the ML algorithms are evaluated in terms of classification accuracy and misclassification rates.

A Review on Gas Turbine Gas-Path Diagnostics: State-of-the-Art Methods, Challenges and Opportunities

“…This process may include separating different sensor faults [74], distinguishing sensor and actual component faults, and classifying different component faults [62]. Like the detection, measurement residuals can be used in the isolation process based on proper threshold selection [75] or the fault isolation problem can be treated as a classification problem, as reported in [61,76,77]. However, the fault detection and isolation activities do not provide quantitative information about the health status of the engine.…”

A comparative study of data-driven and physics-based gas turbine fault recognition approaches

“…diagnostic features. 21,27 They are extracted from raw data to reveal engine problems such as deterioration, faults, sensor faults, and system malfunctions. A deviation is defined as a relative difference…”

“…Common elements present in the networks are a crossentropy function controlling the network performance by the training error minimization; a hyperbolic tangent sigmoid transfer function f 1 changing within an interval of [1,21] for hidden layer neuron activations; an output layer softmax transfer function f 2 (a generalization of logistic sigmoid function), which varies within an interval of [0, 1]. More detailed information about the chosen network options can be found in Reference.…”

“…Comparisons between some of the above techniques and particular recommendations on the selection of the best algorithm can be found, for example in References. [21][22][23] However, a comparative study covering two main differences directly between the approaches, namely, type of diagnostic space and diagnostic decision rule, has not been addressed so far. A diagnostic space is a multidimensional space where an engine fault classification is represented by engine measurement deviation vectors (data-driven) or estimated fault parameter vectors (physics-based).…”

Abrupt Fault Detection and Isolation for Gas Turbine Components Based on a 1D Convolutional Neural Network Using Time Series Data

“…To extract useful knowledge and make appropriate decisions from Big Data, ML techniques have been regarded as a powerful solution. ML is a "data-driven" approach, that is, ML algorithms try to construct a "set of rules" or "blackbox" model of the "system" under analysis from data, and then use the model to predict the behavior of the system [8]. ML algorithms, for example, ANNs [9], SVM, KNN, BNN, Ensemble methods (Random Forest, etc.…”