Research On Lightweight Anomaly Detection Method Of Oil Depot Industrial Control System Based On Machine Learning

Oil is an important part of the national energy structure,which is flammable,explosive,volatile and liquid.The safety of oil storage is related to the safety of people ’s lives and property and national security.The stable operation of the oil depot industrial control system is the fundamental guarantee.In recent years,the research of anomaly detection methods based on machine learning has become a research hotspot.However,with the development of the industry to the Industrial Internet of Things and Industry 4.0,the number of deployed industrial equipment is increasing.The limited computing resources and storage resources have become constraints and difficulties,which affect and limit the application of the algorithm.Aiming at the two basic elements of machine learning,data and model in anomaly detection algorithm,this paper focuses on data feature selection,model hyperparameter search and model compression,systematically studies the key technologies of lightweight anomaly detection in resourceconstrained scenarios,and conducts experimental research on public industrial control datasets and self-built oil depot datasets.The specific contents are as follows :(1)In order to reduce the data dimension to reduce the consumption of storage resources,slow down the pressure of the model to process large amounts of data,an unsupervised feature selection method based on time series feature fusion autoencoder is proposed.This method is based on the autoencoder.The encoder analyzes the underlying consistency of the data through the continuous relaxation technique of discrete data with improved Laplacian scores.The decoder extracts the timing information of the data through the series feature module and fuses it into the deconvolution layer to increase the series information.Experiments showed that this method can effectively achieve unsupervised data feature selection.For example,it was superior to other methods in reducing the 32.69% data dimension of the self-built dataset.(2)In order to reduce the time and memory consumption required for searching model hyperparameter,a model hyperparameter selection method based on improved multi-population genetic algorithm is proposed.Based on genetic algorithm,this method increases the genotype contained in the population through multi-population evolution and population communication strategy to enhance the population diversity,and further improves the search efficiency by combining simulated annealing algorithm and hash fitness storage strategy.Experiment showed that this method can achieve efficient model generation.And the accuracy of the generated model in experiments on three datasets,including the self-built dataset,was above 98%.(3)In order to reduce the amount of calculation and storage parameters of the deep learning model and reduce the consumption of memory resources and storage resources,a pruning method based on Poisson distribution is proposed.This method analyzes the filter weight of the deep learning model in the historical training process and calculates its mixed difference.The activity of the filter weight is obtained by the adaptive threshold and the redundant parameters are deleted according to the importance of each channel of the filter.Experiments showed that this method can effectively compress the model,and its performance is higher than other pruning methods in experiments on three datasets,including the self-built dataset.