Optimal Squared Wald Statistics-Based Methodology for On-Line Machine Condition Monitoring and Degradation Assessment

Condition-based maintenance necessitates a health index (HI) to depict machine health condition to support maintenance decision-making. Currently, construction of a HI for online machine condition monitoring and degradation assessment is still challenging. Firstly, most existing degradation models need historical abnormal and faulty data for training. Secondly, weights/parameters of degradation modeling lack physical interpretation for automatically picking up informative sensors and locating fault-related information. Thirdly, a single HI rather than a composite HI is normally constructed to reflect partial degradation information. To address these problems and challenges, an optimal squared Wald statistics based methodology for on-line machine condition monitoring and degradation assessment is proposed. To begin with, based on Wald test, Wald mean and trend statistics are respectively introduced to detect whether the mean and degradation trending of a HI are statistically changed. Then, by using an integration of two squared Wald statistics constructed from the Wald mean and trend statistics as a new composite objective function, a degradation assessment statistic optimization model is innovatively proposed to generate an optimal composite HI with physically interpretable fusion weights, which can locate informative frequency bands for nonprocess data in the frequency domain and can automatically pick up informative sensors for process data. Finally, an on-line version of the proposed optimal squared Wald statistics based methodology is developed to be independent of historically abnormal and faulty data. The effectiveness of the proposed method is verified by two significantly different process and nonprocess run-to-failure datasets. Results show that the proposed methodology is capable of promptly detecting incipient faults, and optimal fusion weights can identify informative sensors for process data and informative frequency bands for nonprocess data.