Rotating Machinery Fault Diagnosis Based On Improved Support Vector Machine And Texture Image Analysis

People’s demand for operational reliability and safety of the mechanical equipment was higher and higher with the development of the mechanical equipment towards high speed, heavy load and complication. It was very significant for equipment’s efficient operation and manufacture safety to implement high precision intelligent fault diagnosis with effective pattern recognition methods. Taking the bearing and tool as the study object, Research was carried out around the two core problem of fault classification and fault feature extraction. The key problem were the fault classification methods based on the improved support vector machine and the fault feature extraction methods based on texture analysis, and the engineering application of these methods.The content and results in this paper is as follows:1) Fault diagnosis methods based on support vector machine was the research hotspot. To overcome the adverse effects of randomicity of model parameters, a parameter optimized support vector machine was proposed, which was based on artificial bee colony algorithm, and it is applied to bearing fault diagnosis. In this method, the inverse of classification error rate is used as fitness value, and the artificial bee colony algorithm is used to optimize the penalty factor and kernel parameter of support vector machine. Through the test of UCI dataset, it is proved that the proposed method, taking both the local optimal solution search and global optimal solution search into account, overcomes the defect that traditional optimization method tends to stuck in local optimal solution, and high recognition rate is acquired. The cost time of searching optimized parameters of small number classification problem is also reduced. At last, the proposed method is used in bearing fault diagnosis experiment, and high recognition rate is acquired.2) Tool wear state recognition was a pattern recognition problem under the small samples condition. Only a small quantity of training samples could be acquired under the specific processing condition. To solve the problem, A tool wear state recognition method based on improved hyper-sphere support vector machine was proposed after research. In this method, features were extracted from cutting force signal and vibration acceleration signal, and through correlation analysis the ultima feature vectors were composed of the mean value, RMS, the energy value and the approximate entropy of low frequency band obtained from wavelet transform. In the aspect of classification algorithm, considering the difference of samples’distribution, gravitation method was used to improve the decision function of hyper-sphere support vector machine for acquiring the optimized classification formula. The improved hyper-sphere support vector machine was adopted as classifier to implement tool wear state automatic recognition. Prove by experiment, the proposed method based on hyper-sphere support vector machine had great generalization and learning ability, and high recognition rate could be achieved.3) As the basis of multi-dimension fault information feature extraction method, the texture analysis method based on support vector machine was researched. A support vector machine half-supervised machine learning method based on fuzzy C-mean algorithm objective to feature vector automatic acquirement was proposed and used in texture image segmentation. In this method, an improved Laws energy measure method was adopted as feature extraction method. Feature image was segmented into several blocks and fuzzy C-mean algorithm was adopted to classify the pixels’feature vector in smooth block and to acquire the class marks. The feature vectors and class marks were treated as the training samples of fuzzy support vector machine to implement automation of training sample acquirement. Then the trained support vector machine was used to classify the feature vectors in unsmooth block to implement the high precision texture segmentation. The final classification image was composed of the classification marks of fuzzy C-mean algorithm and fuzzy support vector machine. Some texture image from Brodatz set was chosen to test the proposed method, and high precision rate was achieved. The theoretical basis of time-frequency distribution map feature extraction algorithm in next chapter was also provided.4) Feature extraction played an important role in fault diagnosis classification. The time-frequency contour map by S transform and Hilbert spectrum map by Hilbert-Huang transform contained rich two-dimension information. Base on the research of texture image segmentation method in previous chapter, the fault feature extraction problem, which was about how to apply the texture analysis method in image processing field to one-dimension signal’s time-frequency distribution map feature extraction, was discussed in this chapter. Firstly, two-dimension discrete wavelet transform was applied to time-frequency contour gray image by S transform and Hilbert spectrum gray image by Hilbert-Huang transform, then Laws energy measure method was used to calculate the mean square deviation of wavelet image of every channel as the feature, then the features were composed into feature vector. As a result, an effective mapping model corresponding one-dimension signal to two-dimension texture feature was built. At last, using the support vector machine as classifier, the feature extraction method was approved effective and feasible through bearing fault diagnosis experiment.5) Directing towards the function requirement of condition monitoring and fault diagnosis on high-grade numerical control machine, research and development on networking numerical control machine overall unit intelligent monitoring and diagnosis experiment platform was made. The author’s main work of signal acquisition technology on the strong coupling condition monitoring unit, software function of remote numerical control machine condition monitoring and fault diagnosis system and TDNC-Connect transport protocol for seamless information interaction were expounded. The purpose was constructing the networking numerical control machine overall unit intelligent monitoring and diagnosis experiment platform with support vector machine fault classifier. Through the platform the function of condition display, performance forecast, fault diagnosis and remote monitoring aiming at numerical control machine were implemented. The reliability and effectiveness of this platform was proved through engineering application.