A hybrid machine learning approach for predicting short production stoppages

Abstract

Short production stoppages are unanticipated events that adversely impact a business's productivity and profitability. Frequent occurrences of short stoppages disrupt workflow, reduce productivity, undermine competitive advantage, and contribute to increased operational costs. This study introduces a predictive maintenance model to manage short production stoppages proactively. The proposed hybrid model combines both prognostic and diagnostic approaches to predictive maintenance. The diagnostic component identifies the causes of short stoppages and is modeled using a feedforward artificial neural network (FFNN). The prognostic component forecasts the timing of stoppages, utilizing long short-term memory (LSTM) models. The hybrid approach capitalizes on the FFNN's learning capabilities and the LSTM's strength in capturing long-term dependencies to deliver accurate diagnostic and prognostic predictions. The model is applied in the textile industry, where FFNN and LSTM models are integrated to analyze historical performance and operational data from circular knitting machines. In the diagnostic phase, the FFNN model identified the causes of stoppages with 98.05% accuracy, while in the prognostic phase, the LSTM model predicted the time between stoppages with a strong coefficient of determination (R2) of 0.95472. These results demonstrate that the integrated hybrid model effectively predicts short and instantaneous production stoppages.