Research on sensor condition monitoring and signal reconstruction based on self-correcting anomaly diagnosis model

Abstract

Condition monitoring is essential in industrial processes to ensure safe and efficient operations. Sensor signals, which accurately reflect the state of industrial systems, play a central role in this monitoring. However, the harsh conditions in many industrial environments, especially in nuclear power plants, increase the likelihood of sensor failures. Condition monitoring systems detect anomalies by reconstructing input data, with high reconstruction errors indicating the presence of anomalies. The Multivariate State Estimation Technique (MSET) is a widely used nonlinear, non-parametric model for condition monitoring. Traditional nonlinear models assume that training and test data come from the same distribution. This assumption can lead to significant errors when the model encounters anomalies, making it challenging to detect and reconstruct sensor states. To address these challenges, this paper introduces a self-correcting anomaly diagnosis model. Unlike traditional methods, this model establishes a dedicated data structure to store normal sensor patterns and generates a dynamic memory matrix that adapts to changes in industrial processes; The proposed method combines penalized offset projection with multi-scale estimation to mitigate the impact of anomalies on estimation results. Additionally, a variable correlation analysis method is developed to optimize input feature selection for the model. The new approach self-corrects anomalous data in a transformed signal space, achieving accurate reconstruction of sensor states. The model's performance is validated using real sensor data from a nuclear power plant system. Results demonstrate that the proposed model significantly enhances signal reconstruction and anomaly detection capabilities, even under more severe simulated conditions. Compared to traditional nonlinear models, the new method improves the metric for reducing anomaly interference by an order of magnitude. However, we did not change the calculation method of the higher-order kernel in the original method, which still faces the problem of matrix inversion.