Multi-channel, heterogeneous data source detection of anomalies in Tainter valve machinery systems using a hierarchical Bayesian approach

Ensuring the dependable operation of lock systems is essential for maintaining continuous waterway traffic. However, unforeseen mechanical failures in lock fill/empty Tainter valves, which regulate locking operations, create substantial management challenges. To address this challenge, an operational lock was equipped with various sensors spanning components from the three-phase motor to the final gear shaft. During certain valve openings, unexpected impulses were observed in the jack shaft and sector gear bearing block accelerometer signals, aligning with abrupt changes in sector gear angular displacement. To identify these anomalies rigorously, a heterogeneous multi-channel hierarchical Bayesian framework was implemented, utilizing sector gear bearing block acceleration and gear angular displacement signals. A first-order time differencing method was applied to the angular displacement signal to enhance anomaly visibility. Given the non-stationary nature of the signals and multi-channel heterogeneity, optimal windowing and normalization techniques were introduced to improve statistical reliability and computational efficiency. A binary hierarchical Bayesian hypothesis testing approach was developed, assuming Gaussian noise versus Gaussian-distributed signals conditioned on anomaly mean, standard deviation, and correlation coefficient between channels. This method treats anomaly mean, dispersion, and correlation as random variables, effectively capturing event-to-event variability. The findings provide valuable insights into the operational health of the Tainter valve system, enabling more robust diagnostics and predictive maintenance of lock machinery. The primary aim of this study is to develop and evaluate a hierarchical Bayesian framework for detecting target signals under a binary hypothesis test. The performance of this method is compared against classical non-Bayesian methods, including Neyman–Pearson detection and threshold-based outlier detection. This aim will be demonstrated on data obtained from a Tainter valve mechanical drive system. While developed for this specific application, the methodology has broad applicability for anomaly detection in time-series data across various engineering systems.