Parallel Algorithm Design Of Error-Correcting Support Vector Machine And Application On Analog Circuit Fault Diagnosis

Whereas support vector machine (SVM) possesses the properties of non-linear and high-dimensional recognition, it has been widely used in large amount of areas, such as pattern recognition. In recent years, the development of multi-classification recognition methods is to expand its application ranges from the binary classification to larger areas. However, the bottle of SVM is the computing efficiency, how to accelerate the training and test speed is becoming a focusing hot spot in all walks of life.People have higher requirement for the reliability and practicability of analog circuit, for it has been widely used in life. Nevertheless, some analog circuit faults make it important to find a way diagnosing the fault circuit effectively. Analog circuit fault diagnosis is a way of pattern recognition, so it has the prominent advantages for applying SVM to it. Meanwhile, to accelerate diagnosing speed is to satisfy the practical desire. This thesis researches parallel SVM algorithm and applies to the analog circuit fault diagnosis, in order to improve the diagnosing efficiency.EC-SVM algorithm is a multi-classification SVM algorithm and merged with error correcting coding theory, it has the features of error tolerance and high recognizing accuracy. Additionally, thanks to the parallel feature of GPU and two-layer parallel architecture of CUDA, this thesis proposes a parallel scheme for the system, and proposes a parallel scheme for solo SVM, especially for the most computationally intensive part of kernel function and prominently improved the performance.This thesis also applies the parallel system to the analog circuit fault diagnosis. After the step of feature extraction by wavelet decomposition, the fault feature sample set is used in parallel EC-SVM system which is already designed, and makes a large number of experiments. The results of experiment demonstrate that parallel system is suitable for analog circuit fault diagnosis with the high recognition accuracy and dramatically reduced training and test time.