Diagnosis of Mechanical and Electrical Faults in Electric Machines Using a Lightweight Frequency-Scaled Convolutional Neural Network

Optimizing computational efficiency while maintaining accuracy in electrical machine fault detection is a critical challenge. To address this, the Frequency-Scaled Convolutional Neural Network is proposed as a lightweight yet highly accurate model for detecting electrical machine faults. A key feature of this model is its initial layer, which is inspired by the effects of faults on frequency harmonics in rotating systems. This layer includes a trainable frequency-scaled convolutional layer, designed to optimally separate frequency features over time, hence, reducing the need for a more complex model to achieve high accuracy. Additionally, to further decrease model complexity, a partially connected 2D linear layer is developed in the final layer. The model's performance is evaluated through three case studies. First, the Case Western Reserve University bearing dataset, a well-established benchmark, is used. Despite having only 2,020 trainable parameters and 190,000 floating point operations per second, compared to other models in the literature with millions of parameters, the proposed model achieves 100% accuracy, significantly reducing computational burden while maintaining precision. The model is also applied to the inter-turn short circuit fault dataset for permanent magnet synchronous motors and a public dataset with various fault types, where it again achieves 100% accuracy.