Sensor fault identification based on Error-Correcting Output Codes method

In this paper, we proposed a method within the framework of Error-Correcting Output Codes (ECOC) to solve sensor fault feature extraction and online identification problem. Time and frequency domain signal features are selected as the initial fault characteristics, and we enhance the separability of initial characteristics by a nonlinear transformation based on the probabilistic confidence of the first ECOC outputs. Then construct the second ECOC to classify the fault state taking the feature extracted from the first ECOC as the input. We compared the accuracy and efficiency with this method between different ECOC coding strategies. The results indicate that some certain coding combination show better results on efficiency and accuracy, being able to meet the demand of sensor online fault identification. And generalization experiment on UCI Machine Learning Repository show that this method can also be promoted to other problems.