Abstract
The safe operation of the fluid machine is greatly affected by fault states. With the development of data collection technology in process industrial systems, data-based methods are widely applied in fault diagnosis. The observed data of the fluid machine belongs to typical multivariable time series, so the Euclidean features related to the observed timestamps and the non-Euclidean features related to the observed variables need to be extracted simultaneously by the fault diagnosis method. However, the existing diagnostic studies do not involve the explainability analysis of the diagnostic process, which makes it hard to evaluate the contribution of these features to the accurate diagnosis. An explainable diagnosis model based on temporal and graph convolutional neural network is proposed. The class activation map algorithm is improved to perform explainability analysis of the diagnosis process related to the observed variables and timestamps. Using the simulation data of the fluid machine, features related to observed variables and timestamps of six operating states are fully extracted. Through data experiments, this method can be utilized to achieve high-precision fault diagnosis, and can intuitively display the contribution of each observed variable and its each timestamp to network decision-making. This helps to trace system faults and has significant benefits for process safety assurance.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call