Fault diagnostic of machines under variable speed operating conditions using order tracking and novelty detection

Abstract

Condition-Based Monitoring (CBM) of rotating machinery is becoming increasingly important because it allows improving the machine performance. Nevertheless, most of the real-world machinery operate unique pieces, which are not suitable for inducing faults, and thus making unfeasible to collect useful data from undamaged machine conditions. To this end, novelty detection (ND) had been developed, modeling the normal state to detect machine faults. Therefore, the extraction of a representative feature set must be carried out accurately to represent the target class under different machine states. However, there can be several operating conditions that reflect the dynamic behavior of the machinery, often resulting in non-stationary signals. To improve the stochastic description of non-stationary operating conditions, we propose a CBM methodology that relies on a set of the time-varying narrow-band features that are extracted from the order tracking approach, aiming to encode the time-varying behavior of the acquired vibration signals. With the goal of modeling the target machine condition, the key point here is conceiving the order components like dynamic features, and then, estimating several statistical and similarity parameters for those features to characterize each narrow-band component. Afterward, the multi-dimensional outlier detection problem is solved using both distance-and distribution-based data description classifiers. The ND scheme is tested on test-rig databases holding different types of machine faults when the machine operates under variable speed. As a result, the proposed methodology improves the classification rates compared with the state-of-the-art features and allows characterizing the machine state under its actual operating conditions.