Addressing sensor degradation : raindrop-inspired fault-tolerant graph neural network for soft sensing in industrial processes

Abstract

Physical sensors usually suffer from faults or degradation due to long-term use in the harsh industrial environment, which remains highly challenging for existing learning-based soft sensing models. Recently, graph neural network-based fault-tolerant soft sensing has started to receive attention. However, previous studies still face two key challenges: first, how to construct the robust graph structure under sensor degradation, and second, how to eliminate the disturbance of error propagation during node aggregation. To tackle these underexplored issues, we draw inspiration from the natural phenomenon of raindrops falling and propose a new graph-based fault-tolerant soft sensing framework called RIFTG, which can extract the comprehensive and discriminative features and alleviate the impact of sensor degradation. First, we build a dynamic robust copula graph structure learning method, in which copula-based graph, dynamic graph, and robust graph are used to capture long-term static, short-term time-varying, and fault-tolerant correlations among process variables, respectively. In addition, RIFTG introduces an attention fusion mechanism that adaptively combines the three kinds of sensor graphs, improving the handling ability of fault sensor data. Motivated by the scientific discovery of ripples formed by raindrops, we design a hierarchical message passing operator to correct the feature distribution of biased nodes. Specifically, we divide all nodes into normal and abnormal nodes and devise new message function and updating function to complete feature aggregation for fault sensors. Extensive experimental results on two public industrial datasets verify that RIFTG outperforms state-of-the-art fault-tolerant soft sensing models.