A predictive maintenance model for internet of things devices using long short-term memory and one-dimensional dilated group convolution with residual connection

Abstract

Predictive Maintenance (PdM) of Internet of Things (IoT) devices to enhance their reliability is becoming increasingly crucial as the IoT develops. Loss caused by malfunction can be avoided or minimized if appropriate preparations are made in advance. By summarizing the relevant literature, the use of deep learning to establish the PdM model has become a current research hotspot, but few studies use the classification method. In the multivariate time-series data produced by IoT devices, each data has a temporal and spatial correlation, so how to obtain the two correlations should be given more attention. We propose to extract temporal correlation using the Short-Term Memory network (LSTM) because it can identify long-term dependencies. To obtain spatial correlation, we propose a one-dimensional dilated group convolution with residual connection (1DDGCR) block. The residual connection in it can avoid gradients vanishing as the network deepens. Then we propose a novel PdM model for IoT equipment combining LSTM and 1DDGCR. It is named LSTM_1DDGCR. We utilize the FD001 public dataset, which is a subset of the C-MAPSS dataset, to evaluate the performance of LSTM_1DDGCR. We compare it with two models proposed by previous researchers and find that the proposed LSTM_1DDGCR model shows better performance. In addition, LSTM_1DDGCR is applied to a real-world dataset, and it also shows good performance in practical applications.