Multi-sensor anomaly detection using Scalograms and Kurtograms fusion on the Paderborn University dataset

Accurate and timely fault detection in rotating machinery is essential for industrial safety and operational efficiency. The predominant reliance on single-sensor systems, typically vibrationbased, limits insights into machinery health and is vulnerable to sensor-specific noise, malfunctions, and the challenge of subtle fault signatures in dynamic environments. Furthermore, traditional supervised classification methods struggle with unseen faults and data scarcity. This study overcomes these limitations by conducting a comprehensive evaluation of oneclass anomaly detection (AD) methods combined with multi-sensor feature fusion, using the Paderborn University bearing dataset, which includes synchronous vibration, current, and torque measurements. We systematically fuse sensor signals at the feature level, applying four AD algorithms: Isolation Forest, One-Class SVM, Autoencoders (AE), and Variational AE, across 15 sensor combinations and three feature sets (scalogram, kurtogram, and merged). Results from 30,450 experiments per feature set and per AD algorithm demonstrate that merging scalograms and kurtograms features consistently enhances AD, with several multi-sensor combinations achieving F1 and AUC scores above 0.95. Sensor fusion amplifies these effects, and notably, merging features leads to substantial performance gains. By systematically evaluating every possible combination of sensor types across four AD algorithms and three feature sets, this study eliminates arbitrary fusion strategies and rigorously identifies the most effective sensor and feature integrations for machinery fault detection. These findings confirm that integrating complementary feature domains through multi-sensor fusion produces significantly improved, generalizable, and practical solutions for machinery health monitoring.