Autoencoder And Transfer Entropy Based One-Class Anomaly Detection And Causal Analysis

The detection and analysis of anomalies have positive implications for improving production safety,enhancing production efficiency,reducing production costs,and elevating product quality in industrial settings.However,obtaining anomaly samples requires adding disturbance to the system,and even causing damage,which results in a relative scarcity or absolute absence of anomaly samples.Therefore,there is a need to explore anomaly detection and analysis algorithms that have a wider applicability.Firstly,it is necessary to identify the occurrence of anomalies.This task is categorized as one-class classification,where anomaly detection is accomplished using only positive samples.In this paper,a reconstruction-based approach is proposed,which utilizes a mutual information minimization autoencoder that adds independent constraints to the reconstruction of samples.This approach endows the network with some excellent properties,thereby enhancing its ability to complete the task.The effectiveness of this proposed method is validated through a steel plate surface defect detection task,and network characteristic analysis is also performed.Furthermore,after detecting anomalies,it is necessary to identify the variables that are affected by the anomalies.In this paper,based on the principle of filtering for anomaly variable identification,the Kullback-Leibler divergence is transformed to obtain a new metric for anomaly variable identification.This metric is based on kernel estimation of the probability density function of samples and does not require a specific sample distribution,making it applicable to a wider range of scenarios.By using a multiphase flow case study,it is demonstrated that this method can effectively identify anomaly variables under different anomaly conditions.Finally,an analysis is conducted on the causal relationships between the identified variables in the entire system.To address the issues of high computational complexity,weak causal relationships,complex results,and low applicability of transfer entropy theory in the application of process industries,a hierarchical filtering mechanism is proposed,along with a new causal strength measurement.This mechanism achieves the identification of the main causal relationships while reducing computational complexity during the implementation process.A comprehensive experiment is conducted using an example of chemical production process.The anomaly detection is accomplished using mutual information minimization autoencoder,the abnormal variable identification is achieved based on the deformed Kullback-Leibler divergence,and the causal analysis is performed using transfer entropy and the hierarchical filtering mechanism.The specific implementation process of the hierarchical filtering mechanism is described in detail,and the effectiveness of the causal analysis method is verified by comparing the system composition and control strategies.