Explainable fault diagnosis of gas-liquid separator based on fully convolutional neural network

Abstract

The diagnosis of slug flow is extremely important for the efficient operation of the gas-liquid separator. Traditional fault diagnosis based on the convolutional neural network has not involved the explainability of the convolutional neural network, which makes the model difficult to interpret from the perspective of physical meaning. An explainable diagnostic method based on a fully convolutional neural network is proposed. The class activation mapping, multivariate mutual information, global average pooling and t-distributed stochastic neighbor embedding are combined to analyze the diagnostic process of the network. The experimental results based on simulation data showed that the proposed method can be utilized to interpret the correlation degree between different operating conditions, the importance of each period in the measurement variable, and the engineering significance of the convolutional kernels of the last layer, which provides information supplement for fault reasoning of human experts.