CNN Hardware Accelerator for Real-Time Bearing Fault Diagnosis

Area of Science:

• Engineering
• Computer Science
• Artificial Intelligence

Background:

• Bearing condition monitoring is crucial for industrial machinery maintenance.
• Traditional methods often require data transfer to cloud servers, posing latency and security concerns.
• Developing efficient, edge-based diagnostic systems is essential for real-time fault detection.

Purpose of the Study:

• To introduce a novel one-dimensional convolutional neural network (CNN) hardware accelerator for real-time bearing condition assessment.
• To implement the accelerator on a cost-effective field-programmable gate array (FPGA) platform for edge computing.
• To enhance diagnostic accuracy and reduce memory footprint through specific techniques.

Main Methods:

• Development of a 1D CNN hardware accelerator.
• Implementation on an FPGA evaluation platform for on-site data processing.
• Utilization of signal down-sampling to extend the visible time span of signal images.
• Application of quaternary quantization to reduce model memory requirements.
• Achieving real-time fault diagnosis at a 100 MHz clock frequency with a 64 K samples/s data rate.

Main Results:

• The proposed CNN hardware accelerator enables real-time bearing condition diagnosis.
• The down-sampling technique improved diagnostic accuracy.
• Quaternary quantization reduced neural network model memory demand by 89%.
• The system achieved a fault diagnosis precision of 96.37%.
• The response time of the CNN hardware system was 0.28 seconds.

Conclusions:

• The developed 1D CNN hardware accelerator provides an efficient and accurate solution for real-time bearing condition monitoring.
• Edge-based processing on FPGA platforms negates the need for cloud data transfer, enhancing speed and security.
• The combination of down-sampling and quaternary quantization offers significant improvements in accuracy and memory efficiency.