A One-Dimensional Depthwise Separable Convolutional Neural Network for Bearing Fault Diagnosis Implemented on FPGA

Area of Science:

• Electrical Engineering
• Computer Science
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) are powerful tools for signal processing and fault diagnosis.
• High computational and memory demands of traditional CNNs limit their application in resource-constrained hardware.
• Depthwise Separable Convolution (DSC) offers a computationally efficient alternative to standard convolutions.

Purpose of the Study:

• To present a hardware implementation of a 1D CNN utilizing DSC for rolling bearing fault diagnosis.
• To optimize the CNN model for reduced computational complexity and memory footprint.
• To evaluate the performance of the hardware implementation in terms of accuracy, power consumption, and resource utilization.

Main Methods:

• Hardware implementation of a 1D CNN on a VC707 FPGA development board.
• Utilizing Depthwise Separable Convolution (DSC) to reduce computational load and parameters.
• Applying DoReFa-Net quantization to compress network parameters and activation functions.
• Processing the Paderborn University (PU) rolling bearing current signal dataset with minimal preprocessing.

Main Results:

• The quantized DSC model requires only 22 KB of storage.
• The model performs 1,203,128 floating-point operations.
• Achieved a fault diagnosis accuracy of 96.12%.
• Demonstrated a power consumption of 527 mW at 50 MHz clock frequency.

Conclusions:

• Hardware implementation of DSC-based CNNs is feasible and efficient for rolling bearing fault diagnosis.
• The proposed method significantly reduces computational complexity and memory requirements.
• The system achieves high accuracy and low power consumption, suitable for embedded applications.