A texture based industrial fault diagnosis model using Gammatone Filter Bank and transfer learning

Abstract

Rapid industrial growth has increased the vulnerability of systems to malfunction and permanent damage. Fault detection systems have been installed to prevent such occurrences. In order to eliminate potential life-threatening dangers or unforeseen obstacles that may jeopardize the manufacturing process, early fault detection has become an essential aspect of modern industry. Because artificial intelligence has become increasingly successful across numerous different domains, many researchers have employed deep learning models to classify faults and are always trying to find faster, more accurate ones. In this paper, we present a deep transfer learning architecture that consists of long short-term memory (LSTM) layers of Recurrent Neural Network to extract features enhanced by gammatone like spectrogram. For the dataset, we have used malfunctioning industrial machine investigation and inspection (MIMII) and ToyADMOS datasets. Our experimented results show that the proposed model detect the different faults with precision. Also, our modified gammatone fast fourier method outperforms traditional gammatone accurate method with consistent performance across all environments.