Application Research Of Deep Learning Based Intrusion Detection Algorithm In AMI

Smart grid integrates computer network with traditional power system,and realizes the intelligence of power grid.The Advanced Metering Infrastructure(AMI)connects the power system and the user side,realizes the interaction of data and information between the power supply side and the user side,then promotes the development of smart grid.Therefore,the safe operation of AMI is the key to the development of smart grid.With the increasing closer connection between smart grid and computer network,the network attacks on AMI system are increasing.At present,the security defense technology in AMI system is still the passive defense technology mainly such as encryption-based protection protocol,which cannot resist the invasion of unknown network attacks.Meanwhile,the equipment in AMI system is in the terminal position which is exposed,the storage space is limited,and the cost of deploying intrusion detection equipment is high.The active defense technology represented by intrusion detection has become an important barrier of AMI security defense.At present,the AMI intrusion detection algorithm based on machine learning is constantly proposed.Machine learning algorithm has better learning and classification ability for small sample data,However,in the face of a large amount of high-dimensional data information,the learning ability of machine learning algorithm is reduced,and the generalization ability is weakened.The main research work of this thesis is as follows:Firstly,aiming at the problem that traditional machine learning algorithms are not effective in detecting and identifying high-dimensional data,an improved online learning machine AMI intrusion detection algorithm DBN-OS-RKELM is proposed.Its extracts the features of the collected historical network log data through DBN,shows the high-dimensional data in the form of low-dimensional,retains the main features and eliminates the redundant features.The machine learning algorithm OS-ELM,has a good recognition rate for small data samples,and reduces the training time in the process of online batch data training.However,when the data dimension increases,the learning performance of OS-ELM algorithm decreases.At the same time,in the real-time online batch learning process,the detection results are unstable and the detection rate decreases.The regularization and kernel function are added to the OS-ELM algorithm to prevent the over fitting of OS-ELM in the process of batch training and improve the stability of OS-ELM in batch learning.Finally,the newly arrived data is added to the OS-RKELM network structure,the output weight is updated in real time,and intrusion detection classification is carried out in real time.Through simulation experiments,DBN-OS-RKELM improves the detection accuracy of the OS-ELM algorithm while maintaining a low training time.Through the simulation experiment,it shows that DBN-OS-RKELM has better generalization ability and faster learning rate to improve the accuracy of the algorithm.Secondly,in the process of AMI intrusion detection,the accuracy of DBN-OS-RKELM algorithm is not high when the data scale is small and the data characteristics are incomplete.Therefore,this thesis proposes a DBN-FOA-GRNN algorithm,which uses the generalized regression neural network(GRNN)with good nonlinear mapping ability and fast convergence speed,the algorithm is suitable for the classification of unstable data.The Fruit Fly Algorithm(FOA)is used to optimize the unique random parameters of GRNN,which reduces the probability of falling into local optimum.The experimental results show that the detection accuracy of DBN-FOA-GRNN algorithm is higher than that of DBN-OS-RKELM algorithm when the data feature information is less,which effectively solves the problem that the detection accuracy of DBN-OS-RKELM algorithm is not high when the data feature is incomplete.Compared with the traditional machine learning method,DBN-FOA-GRNN algorithm can also maintain better intrusion detection performance when there is less data.This thesis uses the public intrusion detection data set NSL-KDD to verify and analyze the algorithm.Compared with the traditional machine learning method,the proposed intrusion detection algorithms have a certain improvement in detection effect.Combining the two algorithms can solve the problem of intrusion detection at different data scales.It is suitable for the needs of AMI intrusion detection and has a certain application value.