Research On Fault Diagnosis Model And Algorithm Based On Data-driven

With progress of industrial technology, the process scale of modern industrial production is growing, and the complexity is also getting higer and higer. Once the complicated produce process device fails, not only huge economic losses may be caused but also casualties and damage to the ecology can be resulted in. Through monitoring the running states of production process, fault diagnosis can determine reasons for the system function abnormally as well as the location where the fault happened and the extent of the fault. Moreover, fault diagnosis can forecast the developing trend of fault status and present the countermeasures to eliminate faults, thus the safety in production is guaranteed and the quality of products is improved.Accurate mathematical model of complicated produce process is usually difficult to get, and this will cause the limitions in application of tranditional fault diagnosis methods based on models. However, as the computer technology is widely used in industry field, we can acquire vast mass process data. This creates favourable conditions for development of fault diagnosis strategy based on data analysis. The fault diagnosis method based on data-driven finds the connotative information in the mass process data through various data processing and analyzing methods, and completes the fault detection, isolation, evaluation and decision-making for a complex process, thus the process supervisory ability is enhanced. In this paper, the typical chemical production process with high complexity, strong coupling, nonlinearity and uncertainty is stressd and some novel fault diagnosis methods of complex industrial process based on data-driven are studied.First, the problem about selection of Kernal Function in the KPCA algrithm has been studied. A kind of new Kernal Function is constructed through analyzing the typical Kernal Function; furthermore, a modified KPCA fault diagnosis algorithm is presented. Simulation experiments prove that this algorithm is not only sensitive to the local information but also has strong generalization ability, so the fault information can be found precisely.Second, the feature selection based on infromation entropy and the problem of fault diagnosis are discussed, and some methods based on mutual information are studied. To solve the problem that mutual information is difficult to be calculated in high dimensionality space, a new feature selection algrithm is presented based on second-order mutul information on the assumption that feature information does not depart from uniform distribution seriously. This algrithm can adaptively estimate the redundant information about output category between the candidate feature and selected features without manually presetting the parameters relative to the redundancy degree. Anexact evaluation criterion of features is provided, and a high diagnostic rate and model adaptability are achieved.Third, Kernal method is introduced into Partial Least Square Algorithm (PLS), and an improved model of fault diagnosis based on Tracking Recursive Least Squares (TRPLS) is set up. In this method, the process data are mapped into high dimension sapce from low-dimensional input space through non-linear mapping, and then the non-linear relationship of variables is linearized. The prediction accuracy and the ability to track time-varying of TRPLS algorithm are improved by introducing a tracking factor.And then, a kind of fault diagnosis modelling approach is presented using G-K fuzzy clustering analysis in the product space of input and output. By introducing the fuzzy concept of degree of fault, different shapes and directions of fault mode in data sets can be identified with strong capability of noise treatment.Finally, experiments on typical industrial process proved the aforementioned fault diagnosis models and algorithms.