Detection of Nonstationary Events in Motor Signals Using Nonparametric Statistical Hypothesis Testing

This work presents a lightweight statistical method for detecting nonstationarities in motor current signals, aiming to improve the diagnosis of transient events such as sudden load changes or inverter disturbances. Unlike traditional MCSA techniques that rely on spectral analysis, the proposed framework segments the signals into non-overlapping windows and extracts global statistical features to capture distributional changes over time. These features are compared using nonparametric hypothesis tests with p-value fusion strategies, enhancing robustness, with an adaptive buffering scheme based on Wasserstein distance to allow the system to adjust to signal variability. Results indicates accuracy reaches values as high as 95.9% with the false positive and false negative rates go as low as 7.7% and 0% respectively for both Monte Carlo simulations and real-world tests (from 529 MCSA signals), demonstrating strong potential for real-time, embedded, and scalable motor health monitoring without requiring complex signal decomposition or machine learning models.