Quantized Deep Transfer Learning - Gearbox Fault Diagnosis on Edge Devices

Abstract

This study has designed and implemented a deep transfer learning (DTL) model-based framework that takes an input time series of gearbox vibration patterns, which are accelerometer readings. It classifies the gear’s damage type from a predefined catalog. Industrial gearboxes are often operated even after damage because damage detection is formidable. It causes a lot of wear and tear, which leads to more repair costs. With this proposed DTL model-based framework, at an early stage, gearbox damage can be detected so that gears can be replaced immediately with less repair cost. The proposed methodology involves training a convolutional neural network (CNN) model using a transfer learning technique on a predefined dataset of eight types of gearbox conditions. Then, using quantization, the size of the CNN model is reduced, leading to easy inference on edge and embedded devices. An accuracy of 99.49 % using transfer learning of the VGG16 model is achieved, pre-trained on the Imagenet dataset. Other models and architectures were also tested, but VGG16 emerged as the winner. The methodology also addresses the problem of deployment on edge/embedded devices, as in most cases, accurate models are too heavy to be used in the industry due to memory and computation power constraints in embedded devices. This is done with the help of quantization, enabling the proposed model to be deployed on devices like the Raspberry Pi, leading to inference on the go without the need for the internet and cloud computing. Consequently, the current methodology achieved a 4x reduction in model size with the help of INT8 Quantization.