Spatial Density Clustering-based Anomaly Detection Method For Steel Energy Data

The accuracy of data in the steel industry’s energy systems directly affects the effectiveness of the decisions of dispatch operators.Due to the complexity of the production environment,the data acquisition and transmission equipment are greatly interfered with,resulting in high noise or even abnormal points in the collected data,which seriously interferes with the dispatcher’s judgment of the future operation trend of the system.Therefore,it is particularly important to study the anomaly detection problem of steel energy data.Considering the occurrence and consumption of energy in the steel production process and the data characteristics of the storage unit,the data to be measured is classified as a singledimensional time series and a multidimensional time series.Aiming at a one-dimensional time series with cyclical characteristics,an anomaly data detection method based on multi-criteria evaluation is proposed.Considering the characteristics of uncertain length of the sequence subperiod,the periodic component is decomposed based on the robust local weighted regression of the feature,and the original sequence is divided periodically according to this,and the morphological characteristics of each sub-periodic sequence are evaluated based on the five criteria of kurtosis,coefficient of variation,oscillation coefficient,waveform similarity and abnormal peak,and then the abnormal sequence with an outlier degree greater than the set threshold is identified based on the adaptive fuzzy C mean cluster.Aiming at the multidimensional time series with correlation characteristics between variables,a method for detecting anomalous data based on variable-scale HCA-DBSCAN combined manufacturing behavior is proposed.In order to reduce the complexity of its parameter selection,an adaptive determination architecture of search radius based on probability density estimation is designed,and a search strategy based on the partition is proposed according to this radius,so as to efficiently identify the local suspected anomaly point set.Aiming at the problem that the false detection rate of boundary points of the partition block is too high,an evaluation method based on local outliers is proposed,and the approximate upper bound of the outlier is used instead of the upper bound of the outlier to calculate itself,which improves the efficiency of the algorithm.In addition,the abnormal data is identified in combination with the production signal,and the rapid and efficient detection of abnormal data in the multidimensional time series is realized.In order to verify the effectiveness of the proposed method,the actual production data in the energy system of a large domestic steel enterprise were selected for experiments.The proposed method is compared with the distance-based detection method and the density-based detection method,and the detection results are comprehensively evaluated by the three indicators of false detection rate,missed detection rate and detection time,and the results show that the detection accuracy and efficiency of the proposed algorithm are significantly improved.