The Study Of Machine Diagnosis Based On The Theory About Image Analysis

With the development of modern industry, science and technology, machine devices and control systems become more and more complicated and elaborate. Even more, the cost of them is very considerable. Thus if the systems are interrupted by faults, heave losses in manpower and material resources will be suffered. So large numbers of manpower, material resources and financial is devoted to the study of fault inspection and diagnosis in many country, which is one of the most important research domain. In this paper, the study of fault diagnosis for moving machine device is presented. The main contents are listed as follows:First, the image processing for machine device is studied. At first, the main noise in machine image is analyzed. Fuzzy technology and median filter have combined to a fuzzy adaptive median algorithm for removal of impulse noise while preserving image details. Then histogram equalization is used to enhance the image contrast. The winner filter is applied to removal'mosaic'phenomenon. Next, the popular method for edge detection is introduced. The study emphasis is a novel edge detection method based on gray distance in the 3×3 mask. Two pixel sets S0 and S1 in the mask are used to define an objective function. It is known that if the interset distance between S0 and S1 in the mask is large and the intraset distance of S0 and S1 are small, the compactness of sets S0 and S1 is high and edge intensity is large. Therefore, a new objective function that's contains the distance described above is presented in order to estimate the edge intensity. The non-maxima suppression is applied to extract the edge points. At last, the threshold segmentation is studied. An improved 2D adaptive thresholding is proposed which unite interclass and intraclass variance. By this, a new threshold objective function is defined to compute the optimal threshold.The second, the feature extraction and description based on image vision characteristic are studied. At first the usual feature extraction method is referred, such as principal components analysis, invariant moments and geometrical feature. Then the author put an emphasis on the RHT-LSM method to detect lines in image which is the combination of random Hough transform and least square method which can detect the bending line. At last the FINRT method for shape feature extraction is expatiate which is affine invariant. The projection of the objective image in different directions is computed by Radon transform (RT). In order to overcome the affine variant shortcoming for the RT, the integral, normalization, and Fourier transform are processed. Even more, the dimensionality of the FINRT is reduced by discarding a few high-frequency DFT coefficients which will accelerate the computation speed and robustness.The third, the recognition method for device components and fault using radial basis function (RBF) neural network is studied. RBF neural network provides a powerful technique for generating multivariate nonlinear mapping. Because of their simple topological structure, the training of RBF network is rapid. This provides a motivation for using RBF neural network to fault diagnosis in this research. At first, the author introduced the essential structure of RBF neural network, performance, and its common learning algorithm. In order to obtain the most salient and invariant features of the image, the Fisher's linear discriminant (FLD) is applied in the truncated Fourier domain. FLD considers not only between-class variance but also within-class variance, and optimizes the solution by maximizing the ratio of between-class scatter to within-class scatter. It is used to find a linear projection of the original vectors from high-dimensional space to an optimal low-dimensional subspace. After FCM and FLD, it is reasonable to choose the mean value of the training sample in every subclass as the RBF center. Then the ratio of the median distance from one center to the centers belonging to other classes to the overlap coefficient is estimated as width for the RBF. At last, the number of hidden nodes is selected by the cross validation method.The last, the proposed method is used in trouble of moving freight car detection system for diagnosing the moving train's fault and validating the feasibility of the method. The main objective of the system is to record the moving freight's image, recognize the important part, and fault diagnosis. The experimental results indicate that the performance of the presented method is correct and effect. The right recognition rate for the important components in the freight train is 90%.