基于改进CNN-LSTM的综合传动装置异常检测方法  

Detection method of integrated transmission system anomaly based on improved CNN-LSTM

作　　者：姚海洋 陈涛[1] 贾然[1] 邢培鑫 孙光新 YAO Haiyang;CHEN Tao;JIA Ran;XING Peixin;SUN Guangxin(Key Laboratory of Modern Measurement and Control Technology,Ministry of Education,Beijing Information Science&Technology University,Beijing 100192,China;Assembly and Commissioning Room,Beijing Machinery and Equipment Research Institute,Beijing 100854,China)

机构地区：[1]北京信息科技大学现代测控技术教育部重点实验室,北京100192 [2]北京机械设备研究所总装调试室,北京100854

出　　处：《机电工程》2024年第11期1986-1994,共9页Journal of Mechanical & Electrical Engineering

基　　金：国家重点基础研究发展计划项目(MKF20210009)。

摘　　要：针对履带装甲车辆工作环境恶劣、工况复杂多变,导致综合传动装置数据异常、检测困难等问题,提出了一种基于改进的卷积神经网络(CNN)与长短期记忆网络(LSTM)复合模型的异常检测方法。首先,使用主成分分析法(PCA)和滑动窗口对综合传动监测数据进行了降维和序列划分,提升了异常检测所用的数据质量;然后,使用改进型CNN-LSTM提取了序列数据中的空间特征,利用全连接层输出了数据类别,该方法将CNN和LSTM并联,并引入了残差连接结构,以提高网络对综合传动数据的学习能力;最后,搭建了综合传动装置的异常检测实验台,布置了多种传感器采集综合传动装置的状态数据,并对改进CNN-LSTM方法的有效性进行了验证。研究结果表明:采用残差连接结构改进的并联型CNN-LSTM复合模型,在综合传动系统漏油实验数据的测试集上的异常检测准确率达到了92.7%,并且其接收者操作特性(ROC)曲线下的面积(AUC)达到了0.982,相比于传统CNN-LSTM提升了0.034。改进CNN-LSTM模型具有较强的鲁棒性和泛化能力,能够为综合传动装置的数据异常检测提供一种较为可行的新方法。To address the significant challenges posed by the harsh and unpredictable operating environments of tracked armored vehicles,leading to the anomaly detection of data in complex comprehensive transmission systems,an innovative anomaly detection method based on an enhanced convolutional neural network(CNN)and long short-term memory(LSTM)hybrid model was proposed.Firstly,the advanced principal component analysis(PCA)and precise sliding window techniques were skillfully utilized to effectively reduce dimensionality and accurately segment the comprehensive transmission monitoring data,thereby substantially enhancing the quality of the data used for critical anomaly detection.Then,an optimized CNN-LSTM model was employed to meticulously extract spatial features from the sequence data,with the data categories outputted through a robust fully connected layer.The CNN and LSTM in parallel were ingeniously combined by the sophisticated approach,and a novel residual connection structure was introduced to significantly enhance the networks learning capability for intricate comprehensive transmission data.Finally,a meticulously constructed experimental platform for comprehensive transmission system anomaly detection was established,and various sensors were strategically arranged to collect the intricate state data of the transmission system,validating the effectiveness of the improved CNN-LSTM method.The research results impressively shows that the refined CNN-LSTM composite model with innovative residual connections achieves an outstanding anomaly detection accuracy of 92.7%on the test set of comprehensive transmission system oil leakage experimental data,with the area under the receiver operating characteristic(ROC)curve(AUC)reaching 0.982,marking a notable improvement of 0.034 compared to the conventional CNN-LSTM.The proposed enhanced CNN-LSTM model demonstrates exceptional robustness and generalization ability,offering a highly viable new approach for data anomaly detection in sophisticated comprehensive transmission dev

关 键 词：机械传动装置 卷积神经网络 长短时记忆网络 残差连接 主成分分析法 接收者操作特性 曲线下的面积 

分 类 号：TH132[机械工程—机械制造及自动化] U463.2[机械工程—车辆工程] TJ810.1[交通运输工程—载运工具运用工程]