Fault Detection And Prediction Based On Subspace Fusion And Relevance Vector Machine Algorithm

With the rapid development of modern industrial technology,industrial process becomes more complex and the probability of failure also increases.However,occurrence of failure will reduce the reliability of the system,and even cause major accidents.Therefore,the role of fault detection and prediction is important for ensuring the safety of the production process.In addition,the complex system model is difficult to establish directly.As the wide application of computer technology and intelligent measuring instruments in production practice,process data has also become ever more abundant and accessible.Under this background,the data-driven fault detection and prediction technology has been developed and applied rapidly,which can realize early fault detection and remaining useful life(RUL)prediction according to existing data.Maintenance management can be carried out in advance to prevent the occurrence of serious dangerous accidents,reduce the maintenance and operation cost,and significantly improve the safety and reliability of the industrial system.Based on the existing research results,this paper proposes an improved fault detection and prediction method for the shortcomings of the existing methods.The main contents of paper are as follows:(1)Due to the large scale and multi-mode characteristics of process data,the traditional centralized fault detection methods have certain limitations.Therefore,this paper proposed a fault detection strategy based on Bayesian inference(BI)fusion of multiple characteristics sub spaces.Through the Jarque-Bera test and Augmented Dickey Fuller test methods,the overall process variables are divided into three subspaces,named Gaussian stationary subspace,non-Gaussian stationary subspace and nonstationary subspace.In addition,the dynamic principal component analysis(DPCA),dynamic independent component analysis(DICA)and cointegration analysis(CA)were used to detect the fault in each subspace.After that,local monitoring results are fused to obtain the global monitoring statistics by BI.(2)Since the poor learning ability and generalization ability of the single kernel model,it is difficult to take into account the global and local characteristics at the same time.Therefore,the mixed kernel is used to improve the performance of RVM model.It is worth noting that the parameters and weights of mixed kernel functions are difficult to determine directly.The grid search method is adapted to select the parameters and weights of the kernel function.Finally,the optimized multiple kernel relevance vector machine(OMKRVM)model achieve more accurate prediction capability.(3)Considering the accuracy of single prediction methods is not ideal,a hybrid prediction model combined with unscented particle filter(UPF)and OMKRVM is proposed in this paper.Firstly,the UPF model is applied to get the initial estimation and the error between the initial estimated value and the real value.Since the error data has noise,hence,complementary ensemble empirical mode decomposition(CEEMD)is adopted to reconstruct error series to reduce noise.The OMKRVM model is used to predict the evolutionary trend of the error data.Then,the initial estimation is corrected by the error predicted value,so as to eliminate the accumulated deviation and achieve a higher precision prediction.