Optimization Techniques for Convolutional Neural Network Architectures Applied to PMSM Motor Diagnostics

Currently, techniques for diagnostics of the stator winding faults in Permanent Magnet Synchronous Motors (PMSMs) increasingly rely on deep learning models, particularly Convolutional Neural Networks (CNNs), due to their ability to directly process diagnostic data and detect patterns indicative of faults. However, many CNN architectures proposed in the literature are highly complex, with an excessive number of neuron connections that exceed the requirements of the specific fault detection tasks. This complexity can negatively impact the practical implementation and real-time application of such systems.

In this study, we present an optimization approach to reduce the number of connections in a CNN applied to PMSM fault detection and classification. The proposed optimization algorithm utilizes information about the correlation between automatically extracted fault symptoms at different layers of the network. By evaluating the statistical repeatability of features within the network’s architecture, the algorithm selectively eliminates redundant connections and neurons that do not contribute significantly to the fault detection process.

As a result, the optimized CNN requires fewer parameters while maintaining high classification accuracy. This reduction in network complexity also leads to improved response speed, which is crucial for real-time monitoring of PMSM motors. The proposed method ensures that the diagnostic system can quickly identify and classify faults in the stator, enabling more efficient maintenance and reducing the risk of motor failure. The results demonstrate the effectiveness of the optimization technique in both improving performance and enhancing the practical feasibility of CNN-based diagnostic systems for PMSM motors.