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

Abstract

The FDI step identifies the presence of a fault, its level, type, and possible location. Gas turbine gas-path fault detection and isolation can improve the availability and economy of gas turbine components. Data-driven FDI methods are studied in this paper. Some notable gas turbine FDI challenges include: insensitivity to operating conditions, robust separation of faults, noisy sensor readings and missing data, reliable fault detection in time-varying conditions, and the influence of performance gradual deterioration. For conventional ML methods, the problem with handling time series data is its volume and the associated computational complexity; therefore, the available information must be appropriately compressed via the transformation of high-dimensional data into a low-dimensional feature space with minimal loss of class separability. In order to improve the detection and isolation sensitivity, this paper develops a method for FDI based on CNNs. Work in this paper includes: (1) Defining the problem and assembling a dataset. (2) Preparing data for training, validation and test: data generation, feature engineering, data pre-processing, data formatting. (3) Building up the model. (4) Training and validating the model (evaluation protocol). (5) Optimizing: a. deciding the model size. b. regularizing the model by getting more training data, reducing the capacity of the network, adding weight regularization or adding dropout. c. tuning hyperparameters. (6) Evaluation.

Convolutional Neural Network Denoising Auto-Encoders for Intelligent Aircraft Engine Gas Path Health Signal Noise Filtering

“…The name "Convolutional Neural Network (CNN)" indicates that the network employs a mathematical operation called convolution. In this research, to accommodate 2D data of shape (timesteps, measurements), which can be viewed as a 1D grid taking samples of shape (measurements, ) at a regular time interval, One-dimensional Convolutional Neural Networks (1DCNNs) [28,29] are used.…”

“…MLP1-DAE cannot consider temporal order; hence, the 3D tensors of shape (samples, timesteps, features) must be "flattened" as 2D tensors of shape (samples, timesteps*features). Details of the flatten process can be found in reference [28]. There should be 350 neurons for the flattened MLP-DAE input layer in the case study.…”

Aircraft Engine Performance Monitoring and Diagnostics Based on Deep Convolutional Neural Networks

“…The role of this layer in the CNN operation is mainly to produce a meaningful output from the extracted features. Since the fully-connected layer does not allow multi-dimensional data input, the feature output from the feature-learning section needs to be flattened [30].…”

“…Here, the network training should stop when the validation error begins to increase while the training error keeps on decreasing. Zhao and Li [30] applied the cross-validation technique in their suggested CNN-based gas turbine diagnostic method. Using deep ensemble methods could also be considered as an alternative approach to overcome overfitting, as well as quantify the uncertainty in the predictions made [49].…”

“…Grouped convolutional denoising autoencoders were used to reduce measurement noise and extract useful features from aircraft communications, addressing, and reporting system (ACARS) data [8]. A 1D CNN was employed for abrupt fault diagnostics based on time series data [30]. Zhong et al [31] and Yang et al [32] evaluated the effectiveness of the transfer learning principle with a CNN for engine fault diagnostics with limited fault samples.…”