Green AI for industry 4.0: Energy-efficient generalised deep learning approaches to induction motor condition monitoring

Abstract

The field of Industry 4.0 has seen a significant increase in demand for efficient and effective methods of data-driven analysis, particularly in the domain of condition monitoring for machinery. This thesis explores the use of deep learning techniques to address the challenges faced in this field, focusing on the development of energyefficient and generalised approaches for Induction Motor fault detection and classification. The first part of the thesis introduces the Multi-Channel LSTM-Capsule Autoencoder, a novel Neural Network (NN) architecture designed to tackle issues such as generalisation ability, the need for large volumes of labelled data, and understanding spatial context in multivariate time series data from a single data source. Experimental results demonstrate the architecture’s resilience to overfitting, improved training efficiency, and state-of-the-art performance in outlier detection. Building upon the LSTM-Capsule Autoencoder, the second part presents the Dataset Fusion algorithm, a novel dataset composition method for fusing periodic signals from multiple homogeneous datasets into a single dataset while retaining unique features. The proposed approach, tested on a case study of 3-phase current data from Induction Motor fault datasets, significantly outperforms conventional training approaches and effectively generalises across all datasets. The algorithm’s effectiveness under non-ideal conditions and its computational efficiency, in line with the principles of Green AI, highlight its potential for practical use in real-world applications. The final part introduces the Order Domain Transformer (ODT), a pre-processing algorithm designed to standardise and align the frequency components of signals from different motors, enabling the fusion of multiple heterogeneous datasets in the frequency domain. Experimental results indicate that using ODT maintains performance on data from the same motors but results in a substantial improvement in cross-motor generalisation and model performance. The ODT approach demonstrates the potential to train a single model for multiple motors, optimising the utilisation of available labelled data and reducing the computational resources required for training. The proposed methods in this thesis progressively address the challenges of working with single data sources, multiple homogeneous data sources, and multiple heterogeneous datasets, providing a comprehensive framework for data-driven fault detection and classification in industrial settings.