Explainable fault diagnosis of oil-gas treatment station based on transfer learning

Fault diagnosis is crucial for safe operation of the oil-gas treatment station. With the rapid-increasing volume of the data collected in oil-gas fields, more attention has been paid to data-driven diagnosis method. It is difficult for the traditional neural network to learn data features thoroughly without sufficient data samples, which makes transfer learning an effective solution to this problem. However, the existing diagnosis researches based on transfer learning do not involve the explainability analysis, resulting in the black-box nature of diagnosis results. This makes the model difficult to be trusted when deployed in the application scenario. An explainable diagnosis method based on transfer learning is proposed. The two-dimensional class activation map algorithm and multi-dimensional dynamic time warping theory are utilized to explain the diagnosis process of the deep residual network. Through the data collected at the oil-gas treatment station, the process of transfer diagnosis of four abnormal conditions is explained in detail. The experimental results show that this method can be applied to effectively analyze the regional similarity of samples and sample regions attentioned by diagnosis model. This can significantly improve the confidence of the diagnosis model and provide powerful auxiliary tools for fault reasoning and decision-making of human experts.