Low-power analog circuit-based physical neural networks for motor fault diagnosis

Deep learning methods are widely applied in motor fault diagnosis of new energy electric vehicles to analyze complex data and detect subtle faults. However, implementing deep learning models on digital processors requires significant computational power, resulting in high energy consumption and costs. Therefore, this study investigates a low-power analog circuit-based physical neural network (ACPNN) framework for motor fault diagnosis. The approach begins with the design of the basic electronic circuit and classification model for ACPNN, followed by the use of a physics-aware training method to optimize the network's performance. The framework is validated on two motors with different electrical and mechanical faults, demonstrating its effectiveness in diagnosing various fault types. Additionally, a power consumption analysis compares the analog circuit model with traditional digital processors. Results show ACPNN's energy efficiency, positioning it as a promising alternative for real-time motor fault diagnosis in low-power settings. APCNN's low-energy feature suits IoT applications with power-constrained devices. This study highlights physical neural networks' potential in energy-efficient smart system solutions, offering a sustainable and cost-effective approach to motor monitoring and fault detection.