Predictive maintenance for wire drawing machine using MiniRocket and GA-based ensemble method

Predictive maintenance has gained increasing importance recently due to the potential for significant losses resulting from unexpected machine failures. Additionally, the lengthy process of manufacturing and shipping components for repairs exacerbates the situation. Therefore, this study aims to introduce a method for enhancing time series classification performance by integrating MINImally RandOm Convolutional KErnel Transform (MiniRocket) and a genetic algorithm (GA)-based ensemble method for time series classification in the context of predictive maintenance. This study lies in the utilization of MiniRocket for feature extraction from time series data, which significantly reduces computational time compared to traditional methods. Additionally, the application of GA to ensemble learning of homogenous tree-based models introduces a unique strategy for improving classification accuracy and stability. In addition to using benchmark datasets, this study addresses a real-world problem provided by a prominent wire and cable company in Taiwan. Based on experimental results involving a wire drawing machine, it is evident that employing the GA within the ensemble learning method yields superior performance compared to using individual algorithms. This improvement is consistent across both benchmark and case study datasets, in terms of four performance indicators.