Early Diagnosis Method For Slowly Varying Small Fault Based On Data Feature Extraction Technique

Nowadays, along with the dramatic development of the modern industrial technology, the structure of the control system is more complicated. Therefore, it can bring out a series of serious safety accidents once this kind of system generates some slight fault behaving slow changes by days and months multiplying. Traditional micro fault diagnosis methods detect the tiny faults by improving the noise ratio of the fault signal with the means of reducing the noise energy or accumulating the fault amplitude.In this paper, we take the PCA(Principal Component Analysis) and DCA(Designated Component Analysis) as an extraction tool of multivariate statistical feature to design an new method of multi variable data feature extraction, and present an early fault diagnosis method based on AA(Average Accumulative). Then, we combine this early fault diagnosis method with the CUSUM(Cumulative Sum) for better detecting the trend of early fault showing slight changes. In this paper, the specific work is as follows:(1) We put forward an early detection method for slowly tiny faults based on MF(Median Filter) and CUSUM. By the filtered data, namely the accumulation of SPE(Squared Prediction Error) statistics derived from PCA, we establish a MF-CUSUM-PCA model to realize the anomaly detection showing tiny changes in early stage, at the same time, we combine the method of data feature extraction based on the knowledge guide, and construct the MF-CUSUM-DCA model to realize the early fault diagnosis with the slowly tiny changes, reaching the goal of fault pattern recognition.(2) We put forward a method of feature extraction based on the average accumulation. It can not only reduce the noise energy, but also accumulate the fault signal amplitude, improving the noise ratio of the fault signal significantly. To the problem that the measured data is no longer independent and identically distributed after average cumulative, we establish a time-varying anomaly detection model based on AA-PCA to accomplish the early detection with slowly tiny faults. We also establish a time-varying model based on AA-DCA to realize the early fault diagnosis with multiple slowly tiny changes. At last, we combine the method based on AA and the method of CUSUM to detect the trend of slowly varying small fault in advance.(3) The algorithm we put forward is implemented with the GUI of MATLAB, providing a reference for engineers to make health maintenance decisions.