Engine Fault Diagnosis Based On Optimal Model Of Support Vector Machine

With the improvement of people's living standard, the automobile has become indispensable for daily transportation, and our automotive industry has also been rapidly developed. Fierce competition of market has motivated automobile manufacturers to significantly improve their levels of development and manufacture, which causes structure and function of the vehicle to be more diverse and complex. Meanwhile, the category of vehicle failure increases significantly, and fault diagnosis and condition monitoring of the vehicle also faces an increasing number of difficult problems. In this paper, the engine, key parts of the vehicle, is used as research and test object to mainly research some excellent methods which could build a suitable model that reflects the relationship between fault conditions and their corresponding fault features. On the basis of fully absorbing existing research results, a support vector machine model, which makes model structures and parameters optimized, is put forward, and then the method has been applied to engine fault diagnosis successfully. Main research contents and results are as follows:(1) Based on the current research in which the model structure is seldom considered in model optimization of support vector machine, the factors of SVM model structure have been analyzed systematically, such as SVM's binary-class frames, multi-class patterns and kernel forms, etc. On this base, the paper has set up an optimization model whose objective function is the error rate of cross validation on the training sample, and whose optimal design variables include not only the model parameters such as kernel function's parameters, SVM's penalty parameters, but also the above-mentioned factors of model structure. The model that optimizes model parameters and structure includes more factors influencing on SVM's classification accuracy and it belongs to hybrid optimization including discrete and continuous variables. The experiment result shows that, whether it is the benchmark data set or engine fault diagnosis's data set, its error rate of classification of cross-validation in the predetermined model structure demonstrates a kind of non-monotonic and multi-peak distribution. Therefore, there are two key points to solving the optimization model. One is how to avoid premature convergence of optimization algorithm and the other is how to settle the problem of above-mentioned hybrid optimization.(2) This paper uses two PSO algorithms to optimize the parameters of SVM model. Firstly, the control parameters of PSO algorithms have been optimized by Taguchi experiment design method, and then a dynamic neighborhood PSO algorithm, DNPSO algorithm, is proposed, which, through the increasing size of particle neighbors, allows the swarm to have a better exploring ability in the early and middle period of iteration and a better exploiting ability in the middle and later period. As a result, the algorithm is superior to the standard PSO algorithm. Above all, some adaptive control parameters of PSO algorithm, such as inertia weight and acceleration coefficient, which are selected through Taguchi method, lay a foundation for the further improvement of the PSO algorithm.(3) By analyzing the causes of premature convergence on PSO algorithms, we have recognized that only the reasonable control parameters can't prevent PSO algorithm getting into the local extremum. To overcome the premature convergence problem, it is necessary to apply disturbance variation to amend the swarm's structure. Through analyzing the omni-directional disturbance based on restructuring strategy, the paper puts forward a kind of parallel directional turbulence PSO algorithm, i.e. HSPO-PDT algorithm, which adopts directional information matrix to decide particles' mutation directions, and can fully absorb the existing optimization results. According to the experiments of benchmark test functions and compared with the existing algorithms, the results shows that the algorithm has been greatly improved and fully able to deal with multimodal optimization problems.(4) Although the PSO algorithm can achieve the mixed discrete/continuous optimization. the amount and physical significance of design variables on the support vector machine structure and parameter optimization model are also changed when the structures of kernel function and the binary-classification support vector machine are changed, therefore, the single PSO algorithm can't realize the optimization of the model. In this paper, a substep optimization method of Taguchi method selecting model structure and improved particle swarm optimization algorithms optimizing model parameters is employed so that each group of sample data sets can eventually get a more appropriate model of support vector machine. Since the kernel function is an important influencing factor of support vector machine, this paper analyzes the structure conditions of kernel function and builds two kinds of combination kernel function based on RBF kernel. By testing the standard data set, the kernel functions show better results for some data sets in the model training.(5) After studying and analyzing the structure of the engine fault diagnosis system based on CAN BUS, this paper sets up corresponding models of fault diagnosis based on different fault feature extraction methods. In order to reduce the computational complexity, it is suggested that the support vector machine that predetermines model structure should be firstly adopted to diagnose the engine faults. Secondly, a step-by-step modeling method of all-around optimizing support vector machine's model structures and parameters should be adopted when the diagnosis isn't satisfying. By testing the engine cylinder-piston clearance, intake and exhaust system, the diagnosis of multi-position fault based on operational parameters, the above method is proved feasible.