The Research On The Method Of Fuzzy Pattern Recognition For Flaw Classification In Steels

Flaw recognition is one of the fundamental issues in nondestructive inspection. Flaw recognition technique is very important to improve quality of products and promote security production. It is a cross subject involving a lot of fields, such as in ultrasonic, electronics, digital signal processing, pattern recognition and artificial intelligence etc. Recently with the development of computer technique and digital signal processing, it has began to mature, and has application in the realm of modern engineering. Fuzzy pattern recognition is an important application branch of artificial intelligence. In application fields, it has many advantages, such as rational expression, simple algorithms etc. It has developed quickly within this decade and has important application feature in aviation, ship and oil, etc. The work of this thesis focuses on the method of ultrasonic fuzzy pattern recognition for steel defects, mainly researching on feature extraction and selection, clustering algorithms and fuzzy recognition. It is mainly summarized as follows: 1. Feature extraction and selection The relationship between ultrasonic signal characteristics and flaw classes is not straightforward. Thus the ultrasonic signals are usually preprocessed to enhance classification performance. The preprocessing for the determination of features from the raw signal by use of various digital processing techniques is called the features extraction. This paper analyzes various ultrasonic flaw features, including time-domain waveform features, frequency-domain spectrum features and bispectrum features. The bispectrum analysis, having both of the amplitude and the phase information, is different from the conventional power spectrum analysis, it can reveal some new information about the flaw echoes, which are not obtained by the conventional power spectrum analysis. So, this paper extracts feature of flaw echoes based on bispectrum analysis. The feature vector of the extremum-amplitude information based on bispectrum estimation is extracted, and is considered as the feature vector of flaw echoes in steels. Based the feature vector of the extremum-amplitude information, this paper analyze four flaw classes, including crake, gas porosity, porosity and slag. Analytical results indicate the method is feasible and effective. 2. Clustering algorithm In the clustering algorithms based target function, fuzzy c-means clustering algorithm is early proposed. FCM improves partition performance and reveals the classification of data more reasonably. Unfortunately, FCM exists some disadvantages, for instance, FCM easy plunges into partial minimum, the convergence speed of FCM is lower, and FCM can not process noise data, etc. In order to settle these problems, many researchers proposed modified algorithms. Suppressed fuzzy c-means clustering algorithm (S-FCM) is one of those modified algorithms, although its convergence speed is faster, but astringency of it can not be assured. The thesis analyzes this cause, and proposes a new algorithm: fast checked fuzzy c-means clustering algorithm (W-FCM). The new algorithm adopts the thinking of S-FCM, which magnifies the biggest membership degree and suppresses the subordinate factor. It initializes the membership matrix with the use of division matrix of HCM, suppresses the second biggest membership degree and the third biggest membership degree using different checked coefficients, that the suppress degree of the third one is greater than that of the second one, consequently makes its convergence speed faster than the one of FCM under the precondition of not decreasing partition performance. The results of experimentation show that W-FCM algorithm integrates the advantages of HCM algorithm and FCM algorithm, and it is an effective fuzzy c-means clustering algorithm. 3. Fuzzy recognition The thesis researches multilevel fuzzy recognition for ultrasonic flaw classification in steels. During fuzzy recognition, it classifies testing set by method of the biggest membership degree. When the membership degree of it belongs to its cluster is smaller than the one of it does not belongs to, it is possible to result in an error classification. To the problem, this thesis adopts multilevel fuzzy recognition...