Abstract: The rapid development of intelligent manufacturing brings great opportunities and challenges to security protection.Various kinds of security threats may cause serious losses or even disasters,which have become an urgent problem to be solved in the industrial Internet.A novel symmetric recurrent neural network that utilized the adaptability of neural network and the characteristic of gate circuit in Long Short Term Memory(LSTM) network was proposed.Aiming at the problem of low accuracy inanomaly detection for industrial Internet streaming data over time,the forget gate,memory gate and output gate were calculated by different weights and current input data.Then the prediction results were solved by sigmoid activation function,which were used as the input of Gated Recurrent Unit(GRU) network layer to promote the rapid fitting of the current network,so that the better parameters could be obtained in a short time.The last hidden state of LSTM was kept by combining the advantages of LSTM and GRU to take as the input of next layer for GRU,which made the neural network more smooth and maximum retention the parameters of LSTM.The proposed method greatly improved the accuracy of neural network,which could both efficiently and quickly detect the anomalies in industrial Internet.

Key words: recurrent neural network, long short term memory, gated recurrent unit, industrial Internet, anomaly detection

摘要： 智能制造的迅速发展给网络安全防护带来了巨大的机遇与挑战,各类安全威胁会造成严重的损失甚至灾难,已成为工业互联网亟待解决的问题。鉴于此,提出一种新的均衡循环神经网络,利用神经网络的适应性特点,采用长短期记忆网络(LSTM)的门电路特性,针对工业互联网流数据随着时间推移异常检测准确性较低的问题,通过不同权重与当前输入数据重构得出遗忘门控、输入门控和输出门控。随后通过sigmoid激活函数求得预测结果,并将该结果作为门控循环单元网络(GRU)的网络层输入,由GRU网络层促使当前网络快速拟合,从而较快地获得较优的参数。本方法结合LSTM和GRU的优势,保留LSTM最后时刻的隐藏状态,作为下一层网络GRU的输入,使网络层的连接更加平滑,最大程度地保留LSTM所学习到的参数,获取隐藏特征,既可提高神经网络的精度,又可高效、快速地检测工业互联网络的异常。

关键词: 循环神经网络, 长短期记忆, 门控循环单元, 工业互联网, 异常检测