Research On Intelligent Recognition And Prediction Methods For On-Road Vehicles With High Emissions

With the increasing level of the national economy,the number of domestic vehicles shows a trend of rapid growth year by year,resulting in an increasing proportion of onroad vehicle emissions in the overall air pollutants,and the main contributor to on-road vehicle emissions is high emission vehicles.How to accurately recognize high emission of on-road vehicles and effectively prevent the occurrence of high emission by vehicles is the key to controlling the pollution of on-road vehicles.Accelerating the resolution of this problem will contribute to China’s clean air program and the construction of beautiful China.In response to the above challenges,the thesis conducts research on intelligent recognition and prediction methods for on-road high-emitting vehicles.The main research contents are as follows:(1)To tackle the problem that the remote sensing system(RSS)sets a fixed cut-off threshold to judge the high emission of road vehicles,which is easy to cause error screening,and adopts the method of manually analyzing the statistical characteristics of RSS big data and regularly adjusting the cut-off threshold in various regions,which has high human supervision cost and is difficult to realize,an automatic and fastrecognition model for on-road high-emitting vehicles(AFR-OHV)based on unsupervised clustering learning is proposed,and it is the first application of machine learning,combined with big data analysis for remote sensing monitoring of on-road high-emitting vehicles.The model adopts automatic boundary detection,initial K center determination,normalized K-Medoids to adaptively constructs and updates a clustering database using real-time collections of emission datasets from an RSS.Then,new vehicles,that pass through the RSS,are recognized rapidly by the nearest neighbor classifier,which is guided by a real-time updated clustering database.Experimental results,based on real data of RSSs in Zibo,Shandong,Shijiazhuang,Hebei,and Hefei,Anhui,compared with the other three models,show that AFR-OHV has the best performance indicators of Davies-Bouldin index(DBI)and Dunn validity index(DVI)while the convergence speed is less than 5 seconds.The average value of DBI is 32.96 and the average value of DVI is 0.0045.The average area(AUC)under the receiver operating characteristic curve(ROC)of the model classifier is 0.9686.The rates of different classifications of excessive emissions and self-adaptive cut-points are calculated automatically for on-road high-emitting vehicles,detected by the RSS.The identification performance of the model for high-emission vehicles are better than the traditional recognition method with fixed thresholds,and the recall rate is 13.04%higher on average.(2)In order to solve the problem of the fusion of vehicle inspection data and RSS data,t-distributed stochastic neighbor embedding(t-SNE)is used to reduce the dimension of characteristic parameters,such as emission concentration of different vehicle exhaust components,opaque smoke,vehicle speed,acceleration and specific power,vehicle length,vehicle type,wind speed,wind direction,temperature and humidity,and a high-emission recognition model based on particle swarm optimizationsupport vector data description(PSO-SVDD)for the fusion of vehicle emission labels and RSS data is proposed.Applied to the RSSs in Zibo,Shandong,Ruzhou,Henan,and Anqing,Anhui,the experimental results show that the recognition performance of the PSO-SVDD model based on the Gaussian kernel function is better than that of the particle swarm optimization-one class-support vector machine(PSO-OC-SVM)model based on polynomial kernel function,and the rate of accuracy,precision and recall are 12.07%,22.76%and 2.02%higher respectively.Besides,the model shows high recognition accuracy for the real data collected by different RSSs in different time and space,which further verifies that the model owns good adaptability.The average training time of the model is less than 30 seconds,meeting the real-time requirements of RSS monitoring.(3)To deal with the problem of high emission prediction and early warning of onroad vehicles,a parallel attention-based long short-term memory(PA-LSTM)deep learning-based prediction and early warning model for on-road high-emitting vehicles is proposed.By constructing a multi-layer attention spatial coding mechanism,Vehicle emission data and parameters of portable emission measurement system(PEMS)and on-board diagnostics(OBD)are isolated spatially and input into the model in parallel,which removes the interference of attention weight between different types of data and speeds up the convergence of model training.Combined with the long-short-term memory neural network model and the time decoding mechanism,the high-emission prediction of on-road vehicle emissions is realized.Applied to the PEMS monitoring of vehicle emissions from on-road tests in Yanqing,Beijing and bench tests in the vehicle emission control laboratory of Tianjin Nankai University,the experimental results show that PA-LSTM model can achieve a more accurate prediction of vehicle emissions.Compared with the model with the best prediction performance among the seven popular models,the evaluation indexes RMSE,MAE and MAPE are reduced by 7.49%,4.36%and 0.89%respectively,and R2 is increased by 0.55%.The proposed model can eliminate outliers and restrain the offset of the zero levels in the PEMS data.The most significant thing is that the PA-LSTM model can foresee about the possible high vehicle emissions in the future and provide timely feed-back to the aftertreatment system of vehicle emissions,so as to make corresponding control measurements in time and avoid the occurrence of high emissions.From the perspective of practical engineering application,this thesis studies and designs the above models,which provides new ideas and practical methods for solving the key problems of intelligent recognition and prediction of on-road high-emitting vehicles.