Scalable Predictive Maintenance Model for Manufacturing Systems Based on Recurrent Neural Networks

Abstract

Many manufacturing systems have started generating different types of time-series data from many places – such as machines, conveyors, robots and support stations – under various operating conditions, and scaled predictive maintenance is required. This paper presents a recurrent neural network-based maintenance model that learns equipment degradation from synchronous sensor streams, operating conditions and quality feedback, and is deployable in production cells with different sampling frequencies and asset numbers. The model is a gated recurrent encoder, a cross-asset parameter-sharing mechanism, a reliability-aware loss function, and an adaptive decision layer that converts failure probability and remaining useful life estimates into maintenance actions. A new experimental data set has been introduced to this paper, which contains vibration, current, temperature, acoustic emission, cycle load and quality deviation data from a large-scale production line. Based on numerical experiments, the new model reduced the mean absolute error of remaining useful life from 11.8 hours to 7.4 hours, increased the F1-score from 0.842 to 0.913, and lowered the simulated maintenance cost by 18.6% compared with the original model. Based on the above results, recurrent representation learning can be employed to support predictive maintenance decisions in a distributed manufacturing environment and optimise the trade-off between reliability, throughput and resource scheduling for maintenance objectives.