Fault Detection for Non-Gaussian Processes Using Multiple Canonical Correlation Analysis Models and Box-Cox Transformation

Although the canonical correlation analysis (CCA)-based method has been successfully applied to Gaussian processes, the methods cannot obtain good results for the non-Gaussian processes. Meanwhile, due to the different operating conditions of the systems in actual operation, the data generated under different operating conditions is greatly different, which makes the fault detection more difficult. To deal with these problems, a fault detection method based on multiple canonical correlation analysis models and Box-Cox transformation is proposed. The method divides all conditions into some categories of conditions and builds the switching rules. Then the current condition is determined and switched in real time through real-time status data and switching rules. Since the systems are the non-Gaussian processes, the quadratic statistic established by the CCA for each condition is non-negative and non-Gaussian distributed variable. So, Box-Cox transformation is used to convert quadratic statistic into Gaussian distributed variables, such that the properties of Gaussian distribution can be employed to determine the ranges of the quadratic statistic corresponding to different health conditions. Finally, the fault detecting thresholds of the models are established based on a 99.7% confidence interval of the Gaussian distribution. The proposed method is applied to the process of monitoring of the suspension system. The application results indicate that the proposed approach can give higher fault detection rate and lower false positive rate and improves the detection ability.