A graph-guided network with adaptive evaluation and improvement for disturbed sensors in fault-tolerant soft sensor modeling

Operating in harsh environments, sensors frequently encounter disturbances, causing prevalent deviations and drift in measured values from true values. The disturbed measurement brings extra difficulty for soft sensing, since the performance of most existing methods depends heavily on the assumption that the data is accurate and disturbance-free. Considering the above difficulty, this paper proposes a graph-guided network with adaptive evaluation and improvement (GAEI) to achieve fault-tolerant soft sensor modeling. First, an adaptive evaluation strategy is proposed to calculate sensor reliability, which is developed from two aspects. For instantaneous noise, a pointwise analysis considering the intra-variable temporal dependencies is performed. For continuous drift, the graph structure comparison that reflects the inter-variable dependencies is established, which can deal with additive deviation, static multiplicative deviation, and time-varying multiplicative deviation. Second, a specific message passing operator is designed within a graph neural network. It aims to effectively exploit information from trusted variables, thereby improving the quality of various deviations. Finally, the evaluation and improvement have an end-to-end structure, providing an adaptive solution to reduce the influence of disturbances. The effectiveness of GAEI is sufficiently demonstrated in a real cement production process.