Smoothing and Matrix Decomposition-Based Stacked Bidirectional GRU Model for Machine Downtime Forecasting

Abstract

The machine downtime occurring during routine production (MDT_RP) because of recessive disturbances (RecDs) can cause huge economic losses and slow down production. In modern industries, condition monitoring, prognosis, and maintenance policies are widely applied to minimize machine failures caused by dominant disturbances (DomDs). However, MDT_RP, because of RecD, has rarely been explored. RecD multivariate time series data faces the challenge of changing information with many noisy and abnormal data points, making it difficult for sequential methods (SMs) to forecast MDT_RP accurately. To address this gap, a novel smoothing and matrix decomposition (MD) based stacked bidirectional gated recurrent unit (STMD_SBiGRU) is proposed for MDT_RP forecasting. Existing SMs have disadvantages in that they are highly affected by noisy data, which significantly affects their feature information extraction capability. The generated error gets amplified during forward propagation, thus interfering with the parameter’s optimization. The proposed STMD_SBiGRU has the advantage of capturing the maximum variance in the dataset by using various MD methods, as well as reducing abnormalities by applying various smoothing factors. This dual innovation of integrating MD and smoothing facilitates the effective distribution of parameters across multiple stacked layers and directions in a proposed model, thus avoiding complexity and overfitting problems of conventional SMs while improving network generalization performance. The extensive experimental work demonstrates that STMD_SBiGRU can forecast MDT_RP with better performance and is highly robust to noisy data compared to other data-driven methods.