车联网数据的PCA-LVQ行驶工况识别方法与测试

Typical drivingcondition identification method and verification based on driver’s networking data and PCA-LVQ algorithm

提出了基于主成分分析-学习向量量化(PCA-LVQ)神经网络智能算法的行驶工况的识别方法。基于用户车联网数据,通过运动学片段划分后,首先对速度、刹车频次、驾驶时间等多维度特征参数进行主成分分析(principal component analysis,PCA),实现输入信息降维处理,避免冗余信息带来的识别误差。其次将降维后的信息输入到LVQ神经网络模型中进行训练,并将模型用于用户典型驾驶工况的识别,分别对模型识别的影响因素进行研究。结果表明:基于PCA-LVQ智能算法的行驶工况识别方法能够有效进行工况识别,工况识别的精度与运动学片段长度相关,还受训练样本量和识别量的影响,但不受工况顺序影响。

In machine learning,Principal Components Analysis(PCA)and Learning Vector Quantization(LVQ)are two common intelligent algorithms and usually used for data dimension reduction and classification,respectively.The PCA is an effective method that adoptsan orthogonal transformation to transforms the original set of variables into a smaller set of linear combinations that account for most of the variance in the original set.And the LVQ is an adaptive method for data classification based on training data with the desired class information.The combination of both algorithms(PCA-LVQ)can identify key information in a large amount of complex information and then classification.The PCA-LVQ algorithm could greatly improve the classification accuracy compared to the LVQ algorithm.a driving conditionidentification method based on the PCA-LVQ intelligent algorithm is proposedin this study.The typical driving conditions during the driving process of the car have a significant impact on the durability performance of the vehicle.The percentage of typical driving conditions of customer used in the development or revision of vehicle durability specifications directly determines whether the specifications can reflect the actual driving conditions of vehicle users.Therefore,the determination of typical driving conditions is particularly important for the development of durability specifications.For the traditional vehicle durability specification development and revision,the typical driving conditions are often obtained through the vehicle GPS signal.However,the typical road information contained in GPS depends heavily on the map holder’s back-end data,and the identification of the typical driving conditions of the vehicle appears to be difficult in the absence of the relevant database.However,with the emergence of artificial intelligence algorithm,the typical driving working condition recognition method based on the vehicle network data is gradually attracted attention.And the PCA-LVQ algorithm describedin this paper is the very artificial intelligence algorithm that can be used for accurate classification.To realize the dimensionality reduction processing of the input signal information and avoid the recognition errors caused by redundant information,the PCA algorithm is firstly conducted to the multi-dimensional feature parameters,such as speed,braking frequency and driving time after the kinematic segmentation of driver’s networking data.Secondly,the reduced-dimensional information is input into the LVQ neural network model for training,and then employed to the recognition of typical driving conditions of users.The influencing factors of model recognition are investigated finally.The accuracy of driving condition recognition is related to the length of kinematic segments,but the length of the optimal block driving condition,specifically,needs to be analyzed and judged based on the kinematic segments of the measured data.The accuracy of typical driving condition recognition is also influenced by the volume of training data.When the recognition sample size is comparable to the volume of training data,the cumulative error of the PCA-LVQ identification model recognition model tends to be stable.However,mixing and crossing different working conditions with each other,the recognition accuracy of the typical driving condition recognition model based on PCA-LVQ,and the blocks of misidentified working conditions remain unchanged,which shows that the model is not affected by the order of working conditions.The accuracy of the algorithm can be further improved by modifying the algorithm to input the correct driving conditions identified as the training set into the model.