Analog electronic system multiple fault diagnosis based on minimizing-deviation LS-SVR

Aiming at multi-duplicated samples training and excessive training models in the process of multi-classification with standard support vector machine, and insuring high integral and partial diagnosis accuracy for analog electronic system, a multiple fault diagnosis method based on minimizing-deviation least squares support vector regression (MDLS-SVR) was proposed. With the square of deviation between the dimension fitting error and the average fitting error of the sample introduced to mimize the spacing between each dimension of the output variable, a new multiple-output least squares support vector regression was finally obtained which had high resolution for prediction results. Then through multiplying the model output by pre-set multi fault modes, the corresponding multiple fault mode to which the maximum value in the result set mapped was just the final diagnosis result. The simulation results show that, the new method simplifies the training process and could keep high integral and partial diagnosis accuracy under small sample set.