Early Detection of Plant Faults by Using Machine Learning

To detect fault early is very important for safety. This paper proposes a method to detect plant faults early. For the early detections, to classify whether the signals is an early sign of faults or not in subtle signal difference is necessary. To the classification, this paper uses SVM (Support Vector Machine), which is one of a powerful classification technique of machine learning. Data in abnormal state of plants is obtained a few moments later after a fault occurs, hence in early stage of faults, the data of abnormal state is not available. Only measured data at normal state of plants is available in learning stage of the method. For a classification of such cases where only one side of data, which is from normal state, available, one class SVM is useful. In addition, to classify a very subtle signal as a signal of abnormal state, a high generalization ability is necessary for SVM. To get the high ability, a new kernel function, the generalized Gaussian function is used. To show the effectiveness of the method given in this paper, a simulation example of a water level control experimental plant is given.