Multisensor Deep Learning for Predictive Maintenance in Industrial Manufacturing Systems

Abstract

Unexpected equipment failures can significantly affect production efficiency and operational reliability in modern manufacturing environments. This paper presents a multisensor deep learning framework for predictive maintenance that integrates vibration, temperature, acoustic, and operating-condition measurements to identify early signs of mechanical deterioration. The proposed architecture combines temporal convolutional networks with cross-sensor attention mechanisms to learn equipment degradation patterns under varying operating conditions. The experimental design considers approximately 37 million sensor observations collected from 2,400 industrial motors across 16 manufacturing facilities over an 18-month period. In the illustrative evaluation, the proposed framework achieves an F1-score of 0.921, compared with 0.869 for the strongest conventional baseline. The model also increases the average failure-warning lead time from 29 to 46 hours while reducing false maintenance alerts by 27.3%. Additional robustness analyses indicate that incorporating multiple sensing modalities improves prediction consistency under varying equipment loads and environmental conditions. The findings highlight the potential of intelligent condition monitoring to support more reliable and efficient industrial maintenance decisions.