Rotor Crack Depth Estimation Using Recurrence Quantification Analysis

Condition monitoring of machinery is gaining increasing importance as more and more complicated and interconnected systems are being developed. In such systems, it becomes critically important to forewarn about the impending faults that have developed in the components to ensure robust and safe operation. Faults that develop in rotating machinery are notoriously difficult to detect using conventional methods. These faults can exacerbate very quickly, which can cause irreparable damage, financial loss and potential casualties. Thus, development of robust techniques for damage detection is imperative for improving safety and reliability of machinery. These techniques use signals from widely deployed sensors, and once validated, can be used for continuous monitoring of the machinery systems, thereby aiding further automation. In this paper, we have developed a technique wherein Recurrence Quantification Analysis (RQA) is used in conjunction with Machine Learning to predict the degradation of a crack in a rotating shaft. Recurrence is a phenomenon that is widely exhibited in many nonlinear processes, and recurrence plot is a two-dimensional visualization of the higher dimensional dynamical systems and helps to understand the recurrence of states in phase space. RQA is an objective and quantitative method to analyze recurrence plots and yields features of the dynamical systems which we employ in conjunction with appropriate Machine Learning algorithms in order to develop a robust diagnostics algorithm. To obtain the dynamic response of the system for various fault conditions, a physics-based model of the cracked rotors is used, and in order to map the RQA features to fault space, a multi-layer perceptron neural network is used. The results of this study demonstrate that RQA has outstanding performance in crack depth estimation with minimal need for expert knowledge about the dynamic response of the system.