Optimizing Neural Networks using Digital Signal Processing for Accelerometer Data

Sammanfattning

Rolling element bearings are critical components in rotating machinery, and monitoring their health is a vital to prevent catastrophic failures. Traditional deep learning models often struggle with mechanical noise in raw vibration data, leading to large computationally expensive models unsuitable for more constrained devices. This study uses the IMS bearing dataset to investigate the impact of digital signal processing on the performance and computational cost. Two different neural architectures (Convolutional Neural Network and Multi Layered Perceptron) were tested using precision, recall, inference speed and model parameters as criteria. With Digital Signal Processing using a Fast Fourier Transform as well as a fourth order Butterworth bandpass filter and targeted feature extraction on the models, it is possible to drastically reduce the input data, improving both models in all criteria. By going from a baseline CNN model with 65 506 to 2 402 while increasing the precision from 95.89% to 99.12%. Similarly the MLP baseline had 4 098 parameters with the optimized having only 38, while the precision goes from 65.93% to 97.66%. These results indicates the efficiency of using Digital Signal Processing to optimize neural networks for real world application.