Extraction of Main Causal Relationship Based on Transfer Entropy with Two-stage Filtering

Modern industrial processes are becoming more complex, creating greater challenges for industrial process control and industrial safety. At the same time, the application of a large number of sensors makes it possible to obtain detailed observation data. By analyzing the causal relationship between variables, the interaction between variables can be understood, which is of great value in risk assessment and root cause analysis. Transfer Entropy (TE) can detect causal relationships between variables without assuming a model, and has been widely used in many fields. However, TE is computationally expensive. In order to overcome this shortcoming and simplify the obtained causal relationship, a new workflow with two-stage filtering is proposed in this work. Before calculating TE, the first filtration is performed with mutual information to reduce the number of required calculations. After completing the hypothesis test, the difference between the TE true value and the hypothesis test maximum value is used to complete the simplification to get the main causal relationship. This effectively reduces the amount of computation while retaining the main causal relationship. The effectiveness of the method is verified by a chemical industrial process.