Application Research Of Support Vector Data Description In Industrial Process Fault Detection

With the rapid development of technology,the structure of modern industrial production equipment has become particularly complicated,and the structural complexity of the system has become higher and higher.Once these production systems fail in the industrial production process,not only the normal production is not guaranteed,but also the incalculable loss of the enterprise,and even extremely serious catastrophic accident.If we can timely and accurately monitor and diagnose the fault in the production process,we can find out the causes of the faults and timely repair the system to ensure the smooth production,effectively reduce the cost of production material,and increase the company's economic profit.Under this demand,data-driven process diagnosis and system fault detection technology has quickly become a hot issue in the field of process monitoring.Data-driven fault diagnosis and detection technology can effectively improve the safety and reliability of production system,timely and accurately monitor and detect process fault,and increase the safety of process operation.However,in real life,due to the strong coupling and structural complexity of industrial production processe,industrial production process data has high dimensional,non-Gaussian,multimodal,nonlinear,and time series related characteristic.Different characteristic of these production data need to be studied to solve new problems and strategies.The main research work and contributions of this paper are as follows:(1)Aiming at the problem of large amount of production data,long fault detection time and low detection efficiency in continuou production process,a Principal Component Analysis(PCA)and Support Vector Data Description(SVDD)was proposed.SVDD combined fault diagnosis and detection method.The method firstly reduced the data dimension of the original production data collected by the principal component analysis method,obtained the principal dimension spatial data and the residual spatial data with simple dimension;and then scored the principal space and the residual space on this basis.The matrix used the support vector data description method to build the model and obtain the threshold.Finally,the new test data was put into the established model,and the previously established model was used for fault detection.The effectiveness of the algorithm is studied by a numerical simulation example and TE process data.The experimental result show that the proposed method has the advantages of saving time and reducing missed detection rate.(2)A multi-modal industrial process fault diagnosis method(NNDSVDD)based on the combination of Nearest Neighbors Difference(NND)algorithm and SVDD algorithm was proposed for the multi-mode operating environment of modern industry.First,the NND was used to preprocess the multi-modal data to eliminate the multi-modal data structure.Then,the SVDD was used to determine the control limit of the statistic on the obtained differential data set.Finally,thenew test data was brought into the model for calculation.The statistical value of the data compare it to the control limit to determine the state of the test data.The proximity differential algorithm can eliminate the multi-mode structure of data to provide a good data modeling basis for SVDD.Applying SVDD on the basis of differential data set improves the detection ability of traditional SVDD for multi-modal process fault.NNDSVDD was applied to numerical simulation example and semiconductor production processe for simulation testing.Simulation result show that NNDSVDD solve the problem of multi-modal data processing and improve the fault detection rate compared with the traditional SVDD.(3)The strong correlation and multi-modality problems between the data collected in the chemical production process.A weighted dynamic SVDD multi-modal fault diagnosis method based on neighborhood difference(NND-DWSVDD)was proposed.First,the data multi-modal structure was eliminated by NND to ensure that the process data obeys the single-peak distribution.Next,the dynamic method and the weighting factor was added to the differentially processed single-mode data,which is important to eliminate the correlation between data.The information was highlighted;finally,the model was built used the SVDD method to achieve online detection.NND-DWSVDD not only solve the problem that the traditional SVDD has poor detection effect due to neglecting the sequence correlation between data,but also realize the purpose of detecting the multi-mode fault by SVDD,which greatly improve the detection efficiency and detection capability of the algorithm.