Structural Drift Detection in Industrial Vibration Signals Using a Non-Supervised Structural Similarity Score (SISSI)

This repository provides an evidence-based technical note documenting the application of a non-supervised structural similarity score (SISSI) to industrial vibration data for early detection of structural drift. The study is conducted on the publicly available IMS Bearing Dataset (Set 1), a widely used run-to-failure benchmark collected on an industrial test rig. Each measurement consists of one second of vibration data sampled at 20 kHz, acquired at regular intervals throughout the operational lifetime of rolling element bearings until failure. The work focuses on the following questions: whether a structural similarity score between one-dimensional signals can capture meaningful information about gradual degradation, why direct thresholding of such scores is insufficient in noisy industrial signals, and how standard statistical process control techniques can transform a noisy similarity score into a robust early-warning indicator. The analysis demonstrates that: Self-comparison tests yield similarity values close to unity, confirming deterministic and consistent score behavior. Comparisons between consecutive vibration measurements produce moderate similarity values, as expected for real stochastic mechanical signals. Baseline-referenced similarity scores exhibit a clear long-term downward trend over the run-to-failure experiment, indicating progressive structural drift. Simple threshold-based decision rules fail to reliably detect this drift due to short-term fluctuations. A one-sided CUSUM accumulator applied to the similarity score successfully converts gradual degradation into a statistically justified early-warning signal. The contribution of this repository is methodological rather than algorithmic. The core structural similarity computation (SISSI) is treated as a deterministic black-box score and is not disclosed in full. Instead, the experimental protocol, decision logic, and aggregated results are provided to allow independent assessment of the conclusions while preserving proprietary implementation details. No supervised learning, fault labels, or domain-specific feature engineering are used. The results illustrate how a structural similarity score, when combined with standard accumulation techniques, can serve as a generic health indicator for noisy industrial time-series data. This work does not claim fault classification or precise time-to-failure prediction. Its scope is limited to early detection of sustained structural change, positioning the method as a general-purpose component for industrial condition monitoring and predictive maintenance systems.