Research On Data-driven Methods Of Equipment State Assessment And Fault Diagnosis

Optical communication equipment is essential for power communication private networks.The optical communication equipment must be in good working order for the power information system to function well.Furthermore,with the digitalization of the power grid and the expansion of optical communication networks,extensive optical communication equipment has raised the bar for operation and maintenance.However,the existing operation and maintenance methodology,which is dependent on human examination of equipment alarm data,makes it difficult to quickly identify the equipment’s abnormal operation state and the root cause of the fault.By using a data-driven approach,intelligent operation and maintenance technology can learn the complicated mapping relationship between real-time monitoring data and healthy states from large-scale historical equipment monitoring information and build a mapping model between them.This study investigates the approaches of real-time health state assessment,alarm pattern anomaly detection,and fault rule extraction of optical communication equipment based on data mining and machine learning.(1)The goal is to reduce the ambiguity about the equipment’s health under various working conditions.A quantitative assessment of equipment real-time state is proposed in this research using density peak clustering and a weighted linear rectified model.Firstly,a sliding time window divides the historical alarms,and the alarm subset is described as an alarm pattern vector based on the type of each alarm in the subset.Then,the density peak cluster analysis is performed on the sample points in the healthy state to create the benchmark health state matrix of the equipment.Finally,the equipment’s health state is quantified by computing the difference between the real-time alarm pattern and the health baseline.The experimental results show that the W-Re LU function is more accurate and stable than existing similarity measuring approaches.(2)Due to the enormous number of alarms and complicated attribute information,it is difficult to find the abnormal alarm pattern in a timely manner.we propose an equipment state anomaly detection method based on alarm multi-feature extraction and the WOASVDD model.Firstly,a sliding window mechanism depending on the number of elements is designed to divide the alarm sequence set,and the multiple features are extracted from the divided alarm subsequence to fully utilize the alarm information.Then,based on feature domain division and importance,an unsupervised feature selection approach is developed.Finally,the WOA algorithm optimizes the model parameters,enables the automatic adjustment of the parameters,and determines the best decision boundary.Experiments demonstrate that this method is effective at anomaly detection,and the proposed feature selection method can improve the SVDD model’s anomaly detection performance.(3)To address the problem that it is difficult to locate the root cause of a fault in time based on the alarm information after an abnormality occurs in optical communication equipment,we present an alarm-fault rule extraction method based on the explainable decision tree.This method adopts random forest for fault feature selection under the condition of unbalanced data,and balances the data set by undersampling and adding Gaussian noise.The method’s main idea is to train a series of decision trees corresponding to all combinations of fault features,and select the decision tree with the best classification accuracy for specific types of faults for rule extraction.The experimental results verify that the extracted rules based on this method have high accuracy and explainable when the fault feature values obey a normal distribution.