Machine Fault Diagnosis Method Using Lightweight 1-D Separable Convolution and WSNs With Sensor Computing

Compared with a wired machine fault diagnosis system, a wireless one based on wireless sensor networks (WSNs) has many inherent merits, such as low cost and ease of installation. However, the limited bandwidth and battery energy of WSNs will impede the high-speed data collection and transmission required in the machine fault diagnosis system. To address this challenge, this article proposes a novel machine fault diagnosis method based on separable convolution and WSNs with sensor computing. First, a lightweight 1-D separable convolution network fault diagnosis model is designed and embedded on the WSNs sensor node. Machine fault diagnosis is then completed on the sensor node, and only the diagnosis result is transmitted in the WSNs to reduce transmission data. A set of experiments have been conducted on the experimental setup to evaluate the proposed method. The results show that the accuracy of the proposed method can reach 98.3%, and the payload transmission data of the WSN decrease from 2048 bytes to 2 bytes, while the node’s energy consumption saves about 15%. Compared with traditional 1-D convolutional neural network (CNN) and 1-D residual network (ResNet), the proposed method is more suitable for the application on resource-constrained WSNs nodes.

I. Introduction

Efficient machine fault diagnosis technology can prevent unexpected device faults, decrease device downtime, and increase device operating efficiency. Over the last few decades, fault diagnosis technology is continually updated by using the latest technological advances in two main categories, i.e., fault diagnosis algorithm and fault diagnosis system. A wide variety of algorithms, such as overall level monitoring, frequency spectrum analysis, artificial neural networks (ANNs), and support vector machine (SVM), have been successfully developed and implemented for machine fault diagnosis [1].