Data-driven Fault Detection of Electric Motors Using Novel Convolutional Neural Network Designs and Integration of Adaptive Signal Processing

Abstract

Fast and accurate fault detection of electric motors is crucial for ensuring reliable performance, minimizing downtime, and preventing further damage to the industrial and transportation systems. The growing demand for electric motors in different applications highlights the need for advanced fault detection methods capable of addressing different challenges. Traditional approaches often struggle with limitations such as poor adaptability to variable operating conditions, sensitivity to noise, and high computational requirements, making them less practical for real-time and resource-constrained applications. This thesis addresses these challenges by developing and evaluating several novel fault detection models, each designed for specific needs. The first model, Long Short-Term Memory Regulated Network (LSTM-RegNet), is designed to address the need for accurate fault feature extraction from different measurements in consistent-speed conditions. By introducing LSTM-regulated feature maps, this model incorporates temporal dependencies into the network, which helps the model extract complementary fault features and add them to the feature map. The novel integration of LSTM regulation within the feature map represents a significant improvement for data-driven fault detection and increases the accuracy and efficacy of the fault detection model. To address the need for lightweight and computationally efficient solutions, the Frequency-Scaled Convolutional Neural Network (FSCNN) is introduced. By introducing a convolutional layer with a scale-trainable wavelet kernel, the model is able to adaptively extract frequency-based fault features. Additionally, the combination of 1D and 2D convolutional layers enhances the processing of both temporal and spatial information, while maintaining a lightweight architecture. Besides, a partially-connected layer is introduced in the last layer of the network to further decrease the complexity of the model while maintaining accuracy. FSCNN is specifically designed for applications requiring resource-constrained hardware, such as embedded systems or edge devices. Its balance between computational efficiency and accuracy positions it as an effective solution for real-time fault detection in scenarios where computational resources are limited. For noisy environments, a model integrating residual Short-Time Fourier Transform (STFT) with a channel-wise regulated convolutional neural network is proposed. The residual STFT introduces a residual of three time-frequency representations with different window lengths, mitigating noise interference and preserving essential fault features. The channel-wise regulated network is designed to better extract the fault features from the residual representation, filter out redundant information, and amplify critical features for enhanced noise robustness. This model is particularly suitable for motors operating in harsh environments where measurement noise can obscure fault characteristics. Finally, the speed-adaptive model, combining Adaptive Window Short-Time Fourier Transform (AWSTFT) with a speed-weighted positional-embedding network, is developed specifically for real-time fault detection in variable-speed conditions. The transformed data using AWSTFT is fed to the speed-weighted positional-embedding network. This design enhances the model’s adaptability to varying operating conditions and localization of the fault by considering the frequency weights and incorporating the positional information. A meta-learning approach is employed to help the model adapt faster to the new tasks, such as unseen speed and load levels. This model is implemented on dSPACE MicroLabBox and TMS320F28379D for real-time Inter-Turn Short-Circuit (ITSC) fault detection at different speeds, which further demonstrates its practical effectiveness, making it an appropriate solution for dynamic operating environments. Each proposed model in this thesis is specifically designed to address critical gaps in existing fault detection methods, contributing to the development of accurate, efficient, and adaptable fault detection models for electric motors. The core contribution of this thesis include: • A novel design of a regulated network with novel features specifically developed for accurate fault detection. • The novel wavelet-fused convolutional neural network layer with trainable scales for improved fault feature extraction and reduced model complexity. • The novel residual STFT design along with a channel-wise regulated network to enhance the robustness of the fault detection model against noise. • An adaptive fault detection design based on the proposed AWSTFT and a lightweight speed-weighted positional-embedding network, trained through a novel meta-learning approach for better generalization across varying working conditions. Three of the proposed models are evaluated using four datasets. The first two datasets, Case Western Reverse University bearing data center and University of Ottawa datasets, include constant speed data with vibration measurements for induction motors. Both are benchmarks for evaluating robustness under real-world conditions, the former features bearing fault and the latter features different fault types. The third dataset is an ITSC fault in a 4-pole PMSM with current measurements. The final dataset, acquired specifically for this thesis, features ITSC faults in a 16-pole PMSM under variable-speed conditions. This dataset challenges the adaptability of fault detection models to dynamic operating scenarios. The last fault detection model is trained and tested by this dataset. Moreover, the model is implemented real-time on the corresponding test rig.