Robust and Lightweight Deep Learning Model for Industrial Fault Diagnosis in Low-Quality and Noisy Data
Jaegwang Shin, Suan Lee
Abstract
Machines in factories are typically operated 24 h a day to support production, which may result in malfunctions. Such mechanical malfunctions may disrupt factory output, resulting in financial losses or human casualties. Therefore, we investigate a deep learning model that can detect abnormalities in machines based on the operating noise. Various data preprocessing methods, including the discrete wavelet transform, the Hilbert transform, and short-time Fourier transform, were applied to extract characteristics from machine-operating noises. To create a model that can be used in factories, the environment of real factories was simulated by introducing noise and quality degradation to the sound dataset for Malfunctioning Industrial Machine Investigation and Inspection (MIMII). Thus, we proposed a lightweight model that runs reliably even in noisy and low-quality sound data environments, such as a real factory. We propose a Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM) model using Short-Time Fourier Transforms (STFTs), and the proposed model can be very effective in terms of application because it is a lightweight model that requires only about 6.6% of the number of parameters used in the underlying CNN, and has only a performance difference within 0.5%.