Impact of window sizes and sensor quality on MCSA for misalignment fault detection

Abstract

To realize predictive maintenance in various manufacturing sectors, accelerometers are frequently used to detect misalignment, imbalance, and/or bearing faults in rotating machinery such as motors, pumps, fans, and turbines. However, Motor Current Signature Analysis (MCSA) presents a promising, non-intrusive alternative for maintenance needs. This study therefore examines the effectiveness of MCSA for the detection of misalignment faults. To do so, we conducted experiments using a state-of-the-art fault emulating drivetrain, capable of inducing varying degrees of misalignment faults at two different motor speeds. Current data was collected using both high-quality and low-quality sensors to allow assessing the impact of sensor quality on predictive performance. Utilizing this dataset, we propose a traditional machine learning pipeline for predicting the degree of misalignment and explore the influence of the analysis window on the performance. An ablation study was conducted to assess performance across various window sizes and sensor qualities. Our results highlight that prediction window size has a significantly larger impact on predictive performance than the quality of the current sensor. For instance, increasing the window size from 800 samples to 12800 samples with a similar feature subset can reduce the MAE score by a substantial margin (65%), while the difference between low and high quality sensors at a fixed window size only has a marginal reduction (17%). The evaluation results and obtained insights, considering the crucial aspects of current sensor quality and analysis window size, together with the open code and dataset, establish the necessary foundation for further advancements in predictive maintenance.