Fault Detection And Location Of Industrial Processes Based On Data-driven

With the increase of the scale and complexity of industrial systems,if there are some faults in these systems,it will not only affect the normal production,but also cause certain economic losses and even casualties.The fault detection and diagnosis technology based on data-driven can effectively improve the reliability of the system.In order to improve the reliability of fault diagnosis for non-linear and multimodal processes,the fault detection and isolation algorithms are studied in this paper.The main work and contributions of this paper are as follows:Aiming at the nonlinear characteristics of chemical production process and the selection of kernel parameters in kernel locality preserving projections(KLPP),a new fault detection method based on ensemble kernel locality preserving projections(EKLPP)is proposed.Firstly,a series of kernels with different parameters are selected to project the nonlinear data into high-dimensional space,and the nonlinear information of the data is extracted.The projection matrix is obtained,and a series of sub-KLPP models are established.Secondly,the kernel matrix of the data to be detected is calculated and projected onto the KLPP model.The test results of each sub-model are obtained by using statistics.The Bayesian decision is used to transform the detection results into the form of probability.Finally,the ensemble learning method is used to combine the test results.The method is applied to TE process.The simulation results show that the method has better detection effect for nonlinear data than the traditional method.According to the multimodal characteristics of industrial production process,a new fault detection method based on weighted differential principal component analysis(WDPCA)is proposed.Firstly,the nearest neighbor of the original data sample and its first k nearest neighbors are selected.The mean and the weight of the first j nearest neighbors of the nearest neighbor are calculated respectively.The weighted difference method is used to preprocess the original data and eliminate the multimodal and nonlinear characteristics.Secondly,the load matrix and the control limits of SPE and T~2 are calculated by principal component analysis.The PCA model is established.Then the fault detection is carried out.The method is applied to a numerical examples and semiconductor production process.The simulation results show that the method has better detection effect for multimodal data than the traditional method.A new fault detection method based on improved local entropy locality preserving projections(ILELPP)is proposed for the multimodal characteristics of industrial production process.Firstly,the improved local entropy method is used to pre-process the original data,eliminating the multimodal and non-Gaussian characteristics of the original data.Secondly,the projection matrix of LPP and the control limits of SPE and T~2 detection indexes are calculated.The LPP model is established.Finally,the data to be detected are projected onto the LPP model after pre-processed by improved local entropy method,and fault detection is carried out.The method is applied to the simulation of a numerical example and semiconductor manufacturing process.The results show that this method has better detection effect for multimodal data compared with the traditional method.In order to solve the problem that kNN has poor detection effect on weak fault points in modes with less dispersion degree when the dispersion degree of each mode is large,a new multimodal fault detection method based on probability density kNN(PD-kNN)is proposed.Firstly,the kNN model of each mode is established.Secondly,the probability density of the sample to be tested and each training sample is calculated.The mode of the sample to be tested is determined.Finally,the model of its mode is used to detect the fault of the sample to be tested.The method is applied to a numerical example and semiconductor industrial process data to verify the effectiveness.Simulation results show that this method has better detection results compared with traditional methods.Aiming at the problem that traditional fault location methods are often affected by smearing effect,the application of kNN in fault location of chemical production process is studied,and a new variable contribution(VCkNN)location index based on kNN is defined.For the detected fault,the variable contribution by kNN of training data is calculated.Secondly,the contribution of variables of test data to location index is calculated.Finally,the fault location map is drawn.The method is applied to a numerical example and TE process.The simulation results show that the method can locate faults accurately and verify the effectiveness of the method.