Data driven predictive maintenance applications for industrial systems with temporal convolutional networks. (July 2022)

• Record Type: Journal Article
• Authors: Kumar Sharma, Deepak; Brahmachari, Shikha; Singhal, Kartik; Gupta, Deepak
• Is Part Of: Computers & industrial engineering. Volume 169(2022)
• Publication Date: 2022-07
• Subjects: CPSs -- Temporal Convolutional Networks -- Machine Learning -- Predictive Maintenance -- Time-Series Forecasting -- Industrial Systems; Engineering -- Data processing -- Periodicals; Industrial engineering -- Periodicals; 620.00285
• DOI: 10.1016/j.cie.2022.108213
• Languages: English
• ISSNs: 0360-8352

Abstract

Highlights: This paper explores the use of TCNs for predicting failures in industrial machines. It presents comparative experiments between TCNs and CNN/LSTM networks. Models are evaluated under different conditions and their performances are presented. Results show that TCNs can outperform LSTMs/CNNs for long time sequence forecasting. Abstract: Cyber-physical systems (CPS) are an indispensable aspect of the modern age's data driven industrial systems. These systems can be controlled and monitored with the help of computer-oriented devices and software that are responsible for integrating the physical environment with cyber frameworks. Owing to the nature of operations in any physical process industry, it becomes imperative to deal with potential failures before they occur. To avoid downtime and losses, predictive maintenance is one relevant policy that utilizes prior information and domain knowledge to help in scheduling operations and maintenance. Predictive maintenance (PdM) in industrial applications is known to improve the efficiency, lifetime, and reliability of the machines and thereby reducing the maintenance cost. With the advances in machine learning approaches in cyber physical systems, reliable predictions can be performed to significantly reduce downtime and operational losses associated with the physical processes. In this paper, usefulness of Temporal Convolutional Networks (TCNs) is investigated with the aim of forecasting the remaining useful life (RUL) for Turbofan engines. This paper demonstrates the effectiveness of using TCNs for prognosis under various evaluation conditions and also provides comparison of their performance with hybrid architectures like CNN-LSTM networks and meta -heuristically optimized LSTM networks. The proposed methods were able to achieve upto 94.47% accuracy in case of binary classification tasks and upto 98.7% precision in case of multi-label classification. The cumulative results in accordance to elaborated test cases are presented with the conclusion of the study.