Fault Prognosis Based On Kalman Filter And Reconstruction Algorithm

With the development of industrial system,the requirements of stability and safety are gradually increasing.In order to ensure its long-term stable operation,fault prognosis methods have been used to monitor and predict their health status in real time.The faults of the industrial system are mainly caused by abnormal fluctuations in the measurement data,such as the mean value increases and the variable covariance changes,and so on.The fault prognosis method mainly identifies these fluctuations and predicts its evolution process.The application of these methods can effectively reduce the loss caused by the fault,and arrange the maintenance strategy reasonably to realize the condition based maintenance.Currently,the data-driven fault prognosis methods typically include two types,namely,detect the incipient faults and predict of remaining useful life(RUL).The detection of the incipient fault is mainly used to find the fault factor at the moment when the industrial process just shows the fault characteristic.The RUL prediction is used to predict the fault evolution process.In this paper,the application of Kalman filter and fault reconstruction method in fault prognosis of industrial systems are studied.Firstly,because existing industrial systems such as marine diesel engines often have complex operation environments and varied work patterns,this paper proposes an enhanced intermittent unknown input Kalman filter(EIIKF)method.This method combines the unknown input Kalman filter(UIKF)and the intermittent unknown input Kalman filter(IIKF)to processing the changes in measurement data caused by work patterns transitions based on estimates of unknown inputs.The improved sequential probability ratio test method is used to realize residual processing residual and find incipient faults.Secondly,considering that the monitoring statistics of T2 and SPE of measurement data will increase and exceed the control limit when the fault occurrence,the monitoring-statistics-based fault subspace decomposition method is combined with the control limit and the iterative method to extract the fault directions in the PCA space.Besides,this paper also selects the fault-related variables in the original monitoring space,which can reduce the noise and computational complexity of the fault evolution process.The VAR method is used to predict the fault evolution process.Finally,considering that the fault characteristics of vibration data are not conducive to be direct expressed,this paper extracts feature signals in the time domain and frequency domain.The fault reconstruction method is used to select the fault directions according to the fault correlation and the magnitude robustness of the fault evolution process.The incipient faults data is used to training VAR model,which can be used to predict the complete fault evolution process.