Hardware Accelerators for a Convolutional Neural Network in Condition Monitoring of CNC Machines

Wind turbines are a vital component as a renewable source of power in the global energy infrastructure. However, the occurrence of bearing defects significantly contributes to wind turbine downtime, which often arise due to the utilization of worn-out tools in computerized numerical control (CNC) milling machines during the fabrication process. To accurately detect tool's level of wear, an analysis of vibration data is performed using a convolutional neural network (CNN) with Fourier transformation-based preprocessing. Within this study, an efficiently implementable CNN is developed and trained, with a focus on hardware implementation. Enhancing the inference speed compared to a software-based execution, a hardware realization of the CNN is explored using high-level synthesis (HLS). The resulting accelerator is integrated as a coprocessor with a RISC-V-based microcontroller within a field-programmable gate array (FPGA). The results obtained are promising, with the CNN exhibiting a root mean square error (RMSE) of
$4.27\ \mu \mathrm{m}$
, indicating a suitable level of accuracy. The integrated accelerators significantly reduce inference runtime by over 99 %, while utilizing just 57.4 % of the available lookup tables (LUT) on the selected FPGA. This advancement facilitates the ongoing analysis of vibration data during the machine's operation.