Research On Anomaly Detection Of Oil Storage Control System With Lightweight Convolutional Network And Transfer Learning

Abnormal detection of the operation data of the oil storage control system is one of the important means to ensure the safe production of the system.However,current researches face many problems.(1)The edge-cloud collaborative architecture is more and more widely used in modern industrial control system,but limited by the fact that the edge side is generally an industrial computer with weak computing power,the anomaly detection algorithms based on deep learning need a lot of computing resources,and are difficult to be directly used on the edge side.(2)The traditional anomaly detection algorithm only gives the classification results of whether anomaly occurs,but it is difficult to locate which points are attacked,and it is difficult to troubleshoot on-site.(3)In the industrial control system of different sites of the same business,there are differences in the number and types of points.A classifier suitable for one site cannot be directly used in other sites.Improving the transfer learning ability of the classifier is an urgent problem to be solved.On the basis of previous researches,this paper analyzes the characteristics of the underlying business data of the oil storage control system,and carries out the following work.1.Firstly,an anomaly detection algorithm "Single Net" based on a lightweight convolutional network is proposed,which is named "Single Net".In order to avoid the interference of adjacent irrelevant points,the network structure of single-column convolution and associative calculation is designed,which improves the detection accuracy and training speed,saves computing resources,and the training and testing of the algorithm can be completed on the edge side.2.Then,on the basis of the lightweight network,a point data attack identification method based on attention model is proposed.This method draws on the encoding and decoding process in the attention model,designs horizontal template convolution,dynamically calculates the weighted average of the point similarity,and can locate the attacked point under the condition of ensuring the accuracy.3.Finally,this paper uses deep transfer learning to solve the problem of scene transfer.Different from the traditional methods,in order to learn the shape of point timing curve,a Laplace regular term is added to the deep transfer learning network.This method improves the recognition rate of targets and improves the transfer problem between scenes.