Effective Approaches to Extract Features and Classify Echoes in Long Ultrasound Signals from Metal shafts

Abstract

Applications of machine learning demand exploration of feature extraction methods and classifier types in order to obtain systems with reliable highest accuracy. The industrial application discussed in this thesis is the classification of ultrasonic echoes in an A-scan. The application is particularly challenging as A-scans are taken from the end of long large complex metal shafts. Although several pattern analysis and machine learning techniques have been used with success in analyzing A-scan data [43, 89], they are typically in the context of very short signals produced from machine parts like plate surfaces or pipe surfaces. Those cases are usually much simpler; in particular, the task reduces to detecting the existence of an echo (indicating a fault in the material). When signals for testing come from long shafts, however, a major problem of mode-converted reflection emerges. These reflections are echoes that do not correspond to real faults (cracks), nor to characteristics in the shaft. These mode-converted echoes may cause misjudgement of the position of cracks on shafts as some critical faint echoes from a cracked surface may lie somewhere among the multiple secondary echoes. Consequences of misclassification are catastrophic with enormous cost in downtime, consequential damage to associate equipment and potential injury to personnel [23]. The problem is then, to discriminate efficiently the different types of reflectors amongst the large volumes of digitalized ultrasonic shaft defect information. As the relationship between ultrasonic signal characteristics and flaw classes is not straightforward, we need to identify and extract informative sets of signal features from which classification might be performed more efficiently and accurately. Among various methods for extracting signal features, the Fast Fourier Transform (FFT) is a useful scheme for extracting frequency-domain signal features [23, 62]. This seems natural when dealing with ultrasound since the traditional representation of these types of signals is by mathematical Fourier series that identify physically meaningful features, like frequency and phase. But recent studies on the ultrasonic flaw classification employ the Discrete Wavelet Transform (DWT) as part of their feature extraction scheme. DWT provides effective signal compression and time-frequency presentation [69, 86]. Many researchers have compared these two feature extraction schemes (FFT and DWT), and most comparisons showed a superiority of DWT to FFT in discriminating the type of flaw (or its non-existence) [74, 78, 90]. However, these previous reported studies have compared the DWT based features with the FFT with limited feature components. Typically, short signals have been reported, with little attention to phase components of FFT sequences. This thesis is the first study analyzing feature extraction in more complex ultrasonic signals from shafts. In particular, we introduce a new FFT-based feature extraction scheme FFT_Magpha which effectively represents both magnitude and phase components of FFT sequences. By employing this state-of-the-art FFT feature extraction scheme, we have more extension and reliability in the investigation about the feasibility of FFT as a better feature extraction scheme than other types of feature extraction schemes such as DWT. On the other hand, the time-variance problem exhibited in DWT has resulted in reservations about its wide acceptance even though DWT coefficients provide effective time-frequency representation of non-stationary signals, and thus are considered useful features for input into classifiers. To solve this, we study a new preprocessing technique for time-domain A-scans, which offer consistent extraction of a segment of the signal from long signals that occur in the NDT of shafts. We compare the performance of this new echo-gating technique with other previously developed methods and investigate that we can use DWT more efficiently as a feature extraction scheme for ultrasonic signal classification by employing this new method in the preprocessing stage. In addition, our investigation in this thesis finds the potential of DWT to be a more reliable feature extraction scheme, through the more stable classification results in different runs of cross validation tests than the results produced in the tests using FFT-based feature extraction scheme. This potential is especially beneficial for the practical NDT for shafts, where we can train a classifier with arbitrary training data and then use the classifier for in-field ultrasonic shaft signal test. We also demonstrate the superiority of using DWT as the feature extraction scheme in the ultrasonic shaft signal classification involving not only ANN hut also SVM. These results dissipate any doubt that the DWT feature extraction methodology is too far suited for ANN which has been popularly employed previously in many similar experimental scenarios. Through these experimental comparisons employing various learning algorithms, we find a certain facility when specific classes of echoes are concerned with different combinations of feature extraction (FFT or DWT) and classifier (ANN or SVM), though DWT is superior to FFT and SVM is superior to ANN in terms of the overall classification accuracy. This finding leads into a hybrid classifier that will improve overall performance by giving more weight to the more trustworthy sub-classifier. Based on those experimental analysis, we design an Integrated SVM classifier (ISVM) which is a combined classification system efficiently employing benefits from each of two SVM classifiers using two feature extraction schemes, FFT aud DWT. The outcomes of a classifier based on FFT is not totally dismissed in this system although the DWT-based classifier has been shown to be superior. This property of ISVM enables us to combine classifiers considering the misclassification cost, to obtain a more informative classification for its application in the field. We also explore the diverse possibilities of heterogeneous and homogeneous ensembles by combining the classifiers along the dimension of feature extraction mechanism, along the dimension of combination method and along the dimension of type of classifier. The experimental result suggests guidelines for designing an integrated multi-classifier system for shaft test data by way of selectively employing the combining structure.