Automatic driving emission interpretable prediction method considering road geometric condition influence

Abstract

The invention provides an automatic driving emission interpretable prediction method considering the influence of road geometric conditions. The method comprises the following steps: S1, recording data of an automatic driving vehicle and related information of a control strategy through an on-site real vehicle test experiment, carding an automatic driving track and emission data of the automatic driving vehicle, preprocessing the data, and constructing a data set; s2, calculating related information of an automatic driving operation condition and road geometry according to the constructed data set; s3, implementing feature engineering by using a machine learning method; s4, screening a machine learning model with good interpretability from the classic machine learning model library for training; step S5, performing evaluation based on model prediction performance and interpretability, and determining an optimal automatic driving emission interpretable prediction model; according to the method, the influence of different road geometry and control strategies on automatic driving emission can be considered, and the emission prediction precision of automatic driving in a real road scene is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of traffic carbon emission prediction, and in particular to an explainable prediction method for autonomous driving emissions taking into account the influence of road geometry conditions. Background Art

[0002] With the surge in the number of vehicles around the world, the energy consumption and greenhouse gas emissions caused by road traffic have become increasingly serious, and reducing road traffic carbon emissions has become an urgent global issue. As a rapidly developing emerging technology in the contemporary transportation field, autonomous driving technology carries the expectation of improving energy efficiency and reducing exhaust emissions. As the main carrier for vehicle travel, the geometric characteristics of the road have a significant impact on the operating status of both manually driven vehicles and autonomous driving vehicles. Different driving modes will inevitably lead to different operating states of vehicles when adapting to changes in road geometry, resulting in differences in exhaust emissions. However, the emission prediction model built based on manually driven vehicle data is difficult to accurately predict and explain autonomous driving emissions. Therefore, it is particularly important to accurately predict and explain the exhaust emissions of autonomous driving vehicles with different control strategies under the influence of road geometry.

[0003] In response to this, existing research in this field has used methods such as real vehicle tests, virtual simulation tests, and theoretical analysis, mainly based on data from manually driven vehicles to explore the impact of different road geometry conditions on exhaust emissions and to build emission prediction models that consider the impact of road geometry conditions. Few studies have taken autonomous vehicles as research objects and simultaneously considered the road curvature characteristics and road slope characteristics to build interpretable emission prediction models. Therefore, accurately predicting and explaining the exhaust emissions of autonomous vehicles with different control strategies under the influence of road geometry has become a key issue that needs to be solved urgently and has important practical significance. Summary of the invention

[0004] The present invention proposes an explainable prediction method for autonomous driving emissions that takes into account the influence of road geometry conditions. It can consider the impact of different road geometries and control strategies on autonomous driving emissions, and effectively improve the emission prediction accuracy of autonomous driving in real road scenarios.

[0005] The present invention adopts the following technical solutions.

[0006] The method for interpretable prediction of autonomous driving emissions considering the influence of road geometry conditions includes the following steps: Step S1: Record the data of the autonomous driving vehicle and the relevant information of the control strategy through on-site real vehicle test experiments, sort out its autonomous driving trajectory and emission data and perform preprocessing to construct a data set; Step S2, using the constructed data set to infer information related to the autonomous driving operating conditions and road geometry; Step S3: Implement feature engineering using machine learning methods; Step S4: Select a machine learning model with good interpretability from a classic machine learning model library for training; Step S5: Evaluate the model prediction performance and interpretability to determine the optimal interpretable prediction model for autonomous driving emissions.

[0007] In step S1, the vehicles selected in the experiment are: commercial self-driving vehicles of different brands and specifications on the existing market; The test site is selected as: a protected closed test site where the terrain slope and curvature of the road fluctuate widely; The experimental design is as follows: autonomous vehicles of different brands and specifications use different control strategies and set speeds to drive on the test field, and use a portable emission measurement system PEMS (Portable Emission Measurement System, PEMS) to collect emission data, and use the global positioning system GPS (Global Positioning System, GPS) and the on-board automatic diagnostic system OBDⅡ (On-Board Diagnostics Ⅱ, OBDⅡ) to record vehicle trajectory data; In the experiment, the relevant information recorded about the autonomous driving vehicle includes: fuel type, total load mass, maximum power, engine displacement, emission standard, injection type, vehicle type classification, passenger seats and maximum speed; The recorded information related to the automatic driving control strategy includes: Adaptive Cruise Control ACC (ACC), headway, ACC set speed (30-120km / h), acceleration control strategy and equipped auxiliary driving system; The collected emission data include: instantaneous emission rates of pollutants such as carbon dioxide, carbon monoxide, hydrocarbons, nitric oxide, and nitrogen dioxide; The converted trajectory data includes: the horizontal coordinate of the vehicle x , vertical coordinate y , vertical coordinate z ,speed V and driving distance S Second-by-second data; The road geometry data includes: road slope G , road curvature C , Road slope change rate GCR , Road curvature change rate CCR ; The operating condition data include: acceleration a , acceleration rate, vehicle specific power VSP ; In step S1, during data combing, the trajectory data is converted into a three-dimensional rectangular coordinate system.

[0008] Step S1 includes the following steps: Step S11, preprocessing the acquired trajectory and emission data, including missing value processing, duplicate value processing, and outlier detection and processing; The missing value processing method is to use the third-order spline interpolation method to fill the missing values; The duplicate value processing method is to find and delete identical duplicate data; The outlier detection and processing method is to use the quartile method IQR to detect outliers and remove them; Step S12: Calculate road geometry data, including road slope G , road curvature C , Road slope change rate GCR , Road curvature change rate CCR ; The road slope G The calculation formula is: ; The road curvature C The calculation formula is: ; The road slope change rate GCR The calculation formula is: ; The road curvature change rate CCR The calculation formula is: ; In the formula, z t for t The vertical coordinate of the moment; ΔS t for t- 1 to t Change in driving distance at any moment; Δx t for t- 1 to t The change of the horizontal axis at a certain moment; Δy t for t- 1 to t The change of vertical coordinate at a certain moment; G t for t The slope of the moment; C t for t The curvature of the moment; Step S13: Calculate the automatic driving operation data, including accelerationa , acceleration rate, vehicle specific power VSP wait; The vehicle specific power VSP The calculation formula is: ; In the formula, V t for t Time speed; a t for t Time acceleration; G t for t The slope of the road at any time.

[0009] Step S2 comprises the following steps: Step S21, encoding the category features related to the autonomous driving vehicle and the control strategy; The features to be encoded include ACC headway, acceleration control strategy, vehicle sensors, vehicle auxiliary driving systems, fuel type, emission standards, vehicle model classification, and injection type, which together with the above recorded and inferred data constitute the original training data set; Step S22: using the geometric path of the road, the trajectory of the autonomous driving vehicle and related information of the control strategy, and the autonomous driving operating conditions as input features to form model input data, using the autonomous driving emissions as output results, and constructing and training multiple machine learning models with embedded feature importance, the machine learning models including XGBoost, lightweight gradient boosting machine, random forest, decision tree, and extreme random tree; Step S23: Use the above machine learning model to output the basic feature importance, and select the machine learning model with the largest proportion of road geometry and autonomous driving control strategy related feature importance as the base model sensitive to the influence of road geometry conditions; Step S24: Based on the model selected in S23, a recursive feature elimination method is used to select the best feature combination with the root mean square error as the cross-validation evaluation index.

[0010] Step S21 encodes the rules for the category features related to the autonomous driving vehicle and the control strategy as follows: If the feature name is ACC headway, the digital coding rule is 1 for short time interval, 2 for medium time interval, and 3 for long time interval. If the feature name is acceleration control strategy, the digital coding rule is 1 for constant headway strategy; 2 for safety distance model strategy; 3 for comfort optimization strategy; 4 for energy consumption optimization strategy; If the feature name is Lane Keeping Assist, the digital coding rule is 1 if not installed and 2 if installed. If the feature name is Lane Centering Control, the digital coding rule is 1 if not equipped and 2 if equipped. If the feature name is automatic emergency braking, the digital coding rule is 1 if not installed and 2 if installed. If the feature name is Traffic Congestion Assist, the digital coding rule is 1 if not equipped and 2 if equipped. If the feature name is blind spot monitoring and assistance, the digital coding rule is 1 if not equipped and 2 if equipped. If the feature name is high-speed assisted driving, the digital coding rule is 1 if not equipped and 2 if equipped. If the feature name is LiDAR, the digital coding rule is 1 if not equipped and 2 if equipped. If the feature name is millimeter wave radar, the digital coding rule is 1 if not equipped and 2 if equipped. If the feature name is camera, the digital coding rule is 1 if not installed and 2 if installed. If the feature name is ultrasonic sensor, the digital coding rule is 1 for unassembled and 2 for assembled. If the feature name is inertial measurement unit, the digital coding rule is 1 if not assembled and 2 if assembled. If the feature name is Global Navigation Satellite System, the digital coding rule is 1 if not equipped and 2 if equipped. If the feature name is fuel type, the digital coding rule is: diesel takes 1; gasoline takes 2; CNG takes 3; LNG takes 4; LPG takes 5; hybrid takes 6; If the characteristic name is emission standard, the digital coding rule is 1 for National I, 2 for National II, and 3 for National III. Country IV takes 4; Country V takes 5; Country VI takes 6; If the feature name is vehicle type classification, the digital coding rule is 1 for light gasoline vehicles and 2 for light diesel vehicles; For heavy-duty gasoline vehicles, take 1; for heavy-duty diesel vehicles, take 2; If the feature name is injection type, the digital coding rule is 1 for carburetor, 2 for electronic injection, 3 for compression ignition, 4 for indirect injection, and 5 for direct injection.

[0011] Step S3 comprises the following steps: Step S31: Using the feature combination selected in step S2, based on the vehicle specific power VSP , road slope G and road curvature C Categorize the data; Step S32: using road geometry, information related to the autonomous driving vehicle and control strategy, and autonomous driving operating conditions as input variables, and autonomous driving emissions as output variables, normalizing the samples consisting of the input variables and the output variables, and randomly dividing the samples according to a preset ratio to form a model training set and a test set; Step S33: select a variety of machine learning models with good interpretability from the classic machine learning model library, and perform parameter optimization through Bayesian optimization and then perform training and testing; The machine learning models include XGBoost, lightweight gradient boosting machine, random forest, decision tree, extreme random tree, support vector machine, deep forest, and ladder boosting tree.

[0012] In step S32, the samples are randomly divided into a ratio of 7:3 as a model training set and a test set respectively.

[0013] Step S4 comprises the following steps: Step S41, constructing a hierarchical structure model, including a target layer and an indicator layer; The target layer is: model interpretability; The indicator layer includes evaluation indicators of feature importance distribution consistency, feature interpretation stability, feature interaction interpretability, global interpretation clarity, and local interpretation stability; The feature importance distribution consistency is: the degree of conformity between the feature importance distribution output by the model and the expert's expectations; The feature explanation stability is: the stability of the distribution of SHAP (Shapley Additive Explanations) values ​​of different samples; The feature interaction interpretability is: the clarity of the model in reflecting the interaction between features; The global explanation clarity is: the clarity of the global explanation of the model SHAP; The local explanation accuracy is: the reasonableness of the SHAP value of a single sample to explain the prediction result; Step S42: construct an importance judgment matrix based on expert scores and perform normalization processing; The importance judgment matrix is ​​formed by comparing and scoring each evaluation index in pairs; The matrix form is: ; The normalization processing formula of the importance judgment matrix is: ; In the formula, the matrix elements d ij For the i The evaluation index is relative toj The importance of the evaluation indicators d norm,ij For the i The evaluation index is relative to j The importance normalization results of the evaluation indicators; Step S43, using the eigenvector method to calculate the weights and perform a consistency check. If the consistency check passes, the weights of each evaluation index are output; if the consistency check fails, the process returns to step S42 to adjust the importance judgment matrix; The weights are calculated using the eigenvector method as follows: Solve for the maximum eigenvalue of the importance judgment matrix λ max And the corresponding eigenvector, normalize the eigenvector to obtain the weight vector W , among which i Elements W i For the i The weight of each evaluation indicator; The consistency test is to calculate the consistency index: CI and consistency ratio CR ,like CR <0.1, the importance judgment matrix has acceptable consistency, otherwise the importance judgment matrix needs to be adjusted; The consistency index CI The calculation formula is: ; The consistency ratio CR The calculation formula is: ; In the formula, λ max is the maximum eigenvalue of the importance judgment matrix; W is the weight vector; W i For the i The weight of each evaluation indicator; CI is the consistency indicator; n is the dimension of the judgment matrix; CR is the consistency ratio; RI is a random consistency index that depends on the dimensions of the importance judgment matrix; Step S44: perform interpretability analysis on each machine learning model based on the SHAP method, invite experts to score the performance of each machine learning model on the above five evaluation indicators based on the interpretability analysis results of each machine learning model, and then calculate the weighted average value as the comprehensive score according to the weight of each evaluation indicator, and sort the interpretability of the model in descending order based on the comprehensive score; The interpretability analysis includes: feature importance analysis, feature interaction analysis, bee swarm diagram analysis, force diagram analysis, decision path analysis, and waterfall diagram analysis; Step S45: sort the models in ascending order based on the prediction performance evaluation index of the single machine learning model, and comprehensively sort the prediction performance of the models according to the average of the sorting; The model prediction performance evaluation indicators include root mean square error, mean absolute error and mean absolute percentage error; Step S46: Select the model with the smallest average values ​​of prediction performance and explainability as the optimal emission prediction model.

[0014] In step S42, the importance judgment matrix is ​​obtained by using a 1-9 scale method, inviting experts to compare and score each evaluation index in pairs, and forming an importance judgment matrix.

[0015] The present invention records the information related to the autonomous driving vehicle and control strategy through on-site real vehicle testing, obtains the autonomous driving trajectory and emission data, and infers the information related to the autonomous driving operating conditions and road geometry; implements feature engineering using machine learning methods; selects a variety of machine learning models with good interpretability from the classic machine learning model library for training; and finally determines the optimal autonomous driving emission interpretable prediction model based on the model prediction performance and interpretability. Compared with the prior art, the present invention has the following beneficial effects: (1) Using an autonomous vehicle equipped with PEMS to collect trajectory and emission data, and inferring road geometry data based on trajectory data, data collection and model construction can be performed on actual road sections, more realistically reflecting the impact of road geometry on autonomous driving emissions; (2) It can consider the impact of different road geometries and control strategies on autonomous driving emissions, effectively improving the emission prediction accuracy of autonomous driving in real road scenarios; (3) It can improve prediction accuracy while taking into account interpretability. It can explain the contribution of road geometry-related characteristics to autonomous driving emissions, and provide a reference for improving autonomous driving control and optimizing road geometry design in the future. It makes up for the deficiency that existing technical solutions are only targeted at straight roads and are difficult to reduce autonomous driving emissions from the perspective of adapting to road geometry conditions.

[0016] In summary, the present invention designs an automatic driving emission prediction method that considers the influence of road geometry and control strategy. The automatic driving emission prediction method that considers the influence of road geometry and control strategy of the present invention uses the information related to the automatic driving vehicle and the control strategy, the automatic driving operating conditions and the road geometry information as the model input, adopts a variety of machine learning models to predict the automatic driving emissions, and comprehensively considers the model prediction performance and interpretability to determine the optimal model, thereby providing an effective technical means for improving the accuracy of automatic driving emission prediction; through the design method of the present invention, data collection and model construction can be realized on any actual road section, which more realistically reflects the influence of road geometry conditions on automatic driving emissions; it can consider the influence of different road geometries and control strategies on automatic driving emissions, and effectively improve the emission prediction accuracy of automatic driving in real road scenes; while improving the prediction accuracy, it also takes into account the interpretability, and can explain the contribution of road geometry related characteristics to the emissions of automatic driving vehicles, providing a reference for improving automatic driving control and optimizing road geometry design in the future, making up for the deficiency that the existing technical solutions are only for straight sections and it is difficult to reduce automatic driving emissions from the perspective of adapting to road geometry conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments: Attached Figure 1 is a flow chart of an explainable prediction method for autonomous driving emissions taking into account the influence of road geometry conditions provided by an embodiment of the present invention; Attached Figure 2 It is a flow chart of an embodiment of the present invention for obtaining emission and trajectory data through real vehicle testing, and calculating road geometry and information related to autonomous driving operating conditions; Attached Figure 3 It is a flowchart of implementing feature engineering using a machine learning method in an embodiment of the present invention; Attached Figure 4 It is a schematic diagram of a process of selecting a plurality of machine learning models with good interpretability for training according to an embodiment of the present invention; Attached Figure 5 It is a schematic diagram of a process of determining an optimal explainable prediction model for autonomous driving emissions based on model prediction performance and explainability of the present invention. DETAILED DESCRIPTION

[0018] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings.

[0019] In order to make the features and advantages of the present patent application more clearly understood, the following embodiments are specifically described in detail as follows: It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0020] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0021] like Figure 1 As shown, the explainable prediction method of autonomous driving emissions considering the influence of road geometric conditions includes the following steps: Step S1: record the data and control strategy related information of the autonomous driving vehicle through on-site real vehicle test experiments, sort out its autonomous driving trajectory and emission data and perform preprocessing to build a data set; that is, record the autonomous driving self-vehicle (self-vehicle is an autonomous driving vehicle) and control strategy related information through on-site real vehicle tests, obtain the autonomous driving trajectory and emission data, and infer the autonomous driving operating conditions and road geometry related information; the flow chart of this step is as follows Figure 2 As shown; Step S2, using the constructed data set to infer information related to the autonomous driving operating conditions and road geometry; Step S3: Implement feature engineering using machine learning methods; that is, implement feature engineering using machine learning methods based on the acquired data; the flow chart of this step is as follows Figure 3 As shown; Step S4: Select a machine learning model with good interpretability from the classic machine learning model library for training; the flowchart of this step is as follows: Figure 4 As shown; Step S5: Evaluate the model prediction performance and interpretability to determine the optimal interpretable prediction model for autonomous driving emissions. The flowchart of this step is as follows: Figure 5 shown.

[0022] In step S1, the vehicles selected in the experiment are: commercial self-driving vehicles of different brands and specifications on the existing market; The test site is selected as: a protected closed test site where the terrain slope and curvature of the road fluctuate widely; The experimental design is as follows: autonomous driving vehicles of different brands and specifications use different control strategies and set speeds to drive on the test field, and use a portable emission measurement system PEMS (Portable Emission Measurement System, PEMS) to collect emission data, and use the global positioning system GPS (Global Positioning System, GPS) and the on-board automatic diagnostic system OBDⅡ (On-Board Diagnostics Ⅱ, OBDⅡ) to record vehicle trajectory data; In the experiment, the relevant information recorded about the autonomous driving vehicle includes: fuel type, total load mass, maximum power, engine displacement, emission standard, injection type, vehicle type classification, passenger seats and maximum speed; The recorded information related to the automatic driving control strategy includes: Adaptive Cruise Control ACC (ACC), headway, ACC set speed (30-120km / h), acceleration control strategy and equipped auxiliary driving system; The collected emission data include: instantaneous emission rates of pollutants such as carbon dioxide, carbon monoxide, hydrocarbons, nitric oxide, and nitrogen dioxide; The converted trajectory data includes: the horizontal coordinate of the vehicle x , vertical coordinate y , vertical coordinate z ,speed V and driving distance S Second-by-second data; The road geometry data includes: road slope G , road curvature C , Road slope change rate GCR , Road curvature change rate CCR ; The operating condition data include: acceleration a , acceleration rate, vehicle specific power VSP ; In step S1, during data combing, the trajectory data is converted into a three-dimensional rectangular coordinate system.

[0023] Step S1 includes the following steps: Step S11, preprocessing the acquired trajectory and emission data, including missing value processing, duplicate value processing, and outlier detection and processing; The missing value processing method is to use the third-order spline interpolation method to fill the missing values; The duplicate value processing method is to find and delete identical duplicate data; The outlier detection and processing method is to use the quartile method IQR to detect outliers and remove them; Step S12: Calculate road geometry data, including road slope G , road curvature C , Road slope change rate GCR , Road curvature change rate CCR ; The road slope G The calculation formula is: ; The road curvature C The calculation formula is: ; The road slope change rate GCR The calculation formula is: ; The road curvature change rate CCR The calculation formula is: ; In the formula, z t for t The vertical coordinate of the moment; ΔS t for t- 1 to t Change in driving distance at any moment; Δx t for t- 1 to t The change of the horizontal axis at a certain moment; Δy t for t- 1 to t The change of vertical coordinate at a certain moment; G t for t The slope of the moment; C t for t The curvature of the moment; Step S13: Calculate the automatic driving operation data, including acceleration a , acceleration rate, vehicle specific power VSP wait; The vehicle specific power VSP The calculation formula is: ; In the formula, V t for t Time speed; a t for t Time acceleration;G t for t The road slope at any time.

[0024] Step S2 comprises the following steps: Step S21, encoding the category features related to the autonomous driving vehicle and the control strategy; The features to be encoded include ACC headway, acceleration control strategy, vehicle sensors, vehicle auxiliary driving systems, fuel type, emission standards, vehicle model classification, and injection type, which together with the above-mentioned recorded and inferred data constitute the original training data set; see Table 1 below for details.

[0025] Table 1: Step S22: using the geometric path of the road, the trajectory of the autonomous driving vehicle and related information of the control strategy, and the autonomous driving operating conditions as input features to form model input data, using the autonomous driving emissions as output results, and constructing and training multiple machine learning models with embedded feature importance, the machine learning models including XGBoost, lightweight gradient boosting machine, random forest, decision tree, and extreme random tree; Step S23: Use the above machine learning model to output the basic feature importance, and select the machine learning model with the largest proportion of road geometry and autonomous driving control strategy related feature importance as the base model sensitive to the influence of road geometry conditions; Step S24: Based on the model selected in S23, a recursive feature elimination method is used to select the best feature combination with the root mean square error as the cross-validation evaluation index.

[0026] Step S21 encodes the rules for the category features related to the autonomous driving vehicle and the control strategy as follows: If the feature name is ACC headway, the digital coding rule is 1 for short time interval, 2 for medium time interval, and 3 for long time interval. If the feature name is acceleration control strategy, the digital coding rule is 1 for constant headway strategy; 2 for safety distance model strategy; 3 for comfort optimization strategy; 4 for energy consumption optimization strategy; If the feature name is Lane Keeping Assist, the digital coding rule is 1 if not installed and 2 if installed. If the feature name is Lane Centering Control, the digital coding rule is 1 if not equipped and 2 if equipped. If the feature name is automatic emergency braking, the digital coding rule is 1 if not installed and 2 if installed. If the feature name is Traffic Congestion Assist, the digital coding rule is 1 if not equipped and 2 if equipped. If the feature name is blind spot monitoring and assistance, the digital coding rule is 1 if not equipped and 2 if equipped. If the feature name is high-speed assisted driving, the digital coding rule is 1 if not equipped and 2 if equipped. If the feature name is LiDAR, the digital coding rule is 1 if not equipped and 2 if equipped. If the feature name is millimeter wave radar, the digital coding rule is 1 if not equipped and 2 if equipped. If the feature name is camera, the digital coding rule is 1 if not installed and 2 if installed. If the feature name is ultrasonic sensor, the digital coding rule is 1 for unassembled and 2 for assembled. If the feature name is inertial measurement unit, the digital coding rule is 1 if not assembled and 2 if assembled. If the feature name is Global Navigation Satellite System, the digital coding rule is 1 if not equipped and 2 if equipped. If the feature name is fuel type, the digital coding rule is: diesel takes 1; gasoline takes 2; CNG takes 3; LNG takes 4; LPG takes 5; hybrid takes 6; If the characteristic name is emission standard, the digital coding rule is 1 for National I, 2 for National II, and 3 for National III. Country IV takes 4; Country V takes 5; Country VI takes 6; If the feature name is vehicle type classification, the digital coding rule is 1 for light gasoline vehicles and 2 for light diesel vehicles; For heavy-duty gasoline vehicles, take 1; for heavy-duty diesel vehicles, take 2; If the feature name is injection type, the digital coding rule is 1 for carburetor, 2 for electronic injection, 3 for compression ignition, 4 for indirect injection, and 5 for direct injection.

[0027] Step S3 comprises the following steps: Step S31: Using the feature combination selected in step S2, based on the vehicle specific power VSP , road slope G and road curvature C Categorize the data; Step S32: using road geometry, information related to the autonomous driving vehicle and control strategy, and autonomous driving operating conditions as input variables, and autonomous driving emissions as output variables, normalizing the samples consisting of the input variables and the output variables, and randomly dividing the samples according to a preset ratio to form a model training set and a test set; Step S33: select a variety of machine learning models with good interpretability from the classic machine learning model library, and perform parameter optimization through Bayesian optimization and then perform training and testing; The machine learning models include XGBoost, lightweight gradient boosting machine, random forest, decision tree, extreme random tree, support vector machine, deep forest, and ladder boosting tree.

[0028] In step S32, the samples are randomly divided into a ratio of 7:3 as a model training set and a test set respectively.

[0029] Step S4 comprises the following steps: Step S41, constructing a hierarchical structure model, including a target layer and an indicator layer; The target layer is: model interpretability; The indicator layer includes evaluation indicators of feature importance distribution consistency, feature interpretation stability, feature interaction interpretability, global interpretation clarity, and local interpretation stability; The feature importance distribution consistency is: the degree of conformity between the feature importance distribution output by the model and the expert's expectations; The feature explanation stability is: the stability of the distribution of SHAP (Shapley Additive Explanations) values ​​of different samples; The feature interaction interpretability is: the clarity of the model in reflecting the interaction between features; The global explanation clarity is: the clarity of the global explanation of the model SHAP; The local explanation accuracy is: the reasonableness of the SHAP value of a single sample to explain the prediction result; Step S42: construct an importance judgment matrix based on expert scores and perform normalization processing; The importance judgment matrix is ​​formed by comparing and scoring each evaluation index in pairs; The matrix form is: ; The normalization processing formula of the importance judgment matrix is: ; In the formula, the matrix elements d ij For the i The evaluation index is relative to j The importance of the evaluation indicators d norm,ij For the i The evaluation index is relative to j The importance normalization results of the evaluation indicators; Step S43, using the eigenvector method to calculate the weights and perform a consistency check. If the consistency check passes, the weights of each evaluation index are output; if the consistency check fails, the process returns to step S42 to adjust the importance judgment matrix; The weights are calculated using the eigenvector method as follows: Solve for the maximum eigenvalue of the importance judgment matrix λ maxAnd the corresponding eigenvector, normalize the eigenvector to obtain the weight vector W , among which i Elements W i For the i The weight of each evaluation indicator; The consistency test is to calculate the consistency index: CI and consistency ratio CR ,like CR <0.1, the importance judgment matrix has acceptable consistency, otherwise the importance judgment matrix needs to be adjusted; The consistency index CI The calculation formula is: ; The consistency ratio CR The calculation formula is: ; In the formula, λ max is the maximum eigenvalue of the importance judgment matrix; W is the weight vector; W i For the i The weight of each evaluation indicator; CI is the consistency indicator; n is the dimension of the judgment matrix; CR is the consistency ratio; RI is a random consistency index that depends on the dimensions of the importance judgment matrix; Step S44: perform interpretability analysis on each machine learning model based on the SHAP method, invite experts to score the performance of each machine learning model on the above five evaluation indicators based on the interpretability analysis results of each machine learning model, and then calculate the weighted average value as the comprehensive score according to the weight of each evaluation indicator, and sort the interpretability of the model in descending order based on the comprehensive score; The interpretability analysis includes: feature importance analysis, feature interaction analysis, bee swarm diagram analysis, force diagram analysis, decision path analysis, and waterfall diagram analysis; Step S45: sort the models in ascending order based on the prediction performance evaluation index of the single machine learning model, and comprehensively sort the prediction performance of the models according to the average of the sorting; The model prediction performance evaluation indicators include root mean square error, mean absolute error and mean absolute percentage error; Step S46: Select the model with the smallest average values ​​of prediction performance and explainability as the optimal emission prediction model.

[0030] In step S42, the importance judgment matrix is ​​obtained by using a 1-9 scale method, inviting experts to compare and score each evaluation index in pairs, and forming an importance judgment matrix.

[0031] The autonomous driving in this example is L2 level autonomous driving. It targets the low-level autonomous driving vehicles currently in use, including electric and fuel-powered vehicles, and predicts the emission problems of low-level autonomous driving through the data collected by the on-board exhaust gas detection equipment PMES.

[0032] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0033] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0034] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0035] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0036] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

[0037] This patent is not limited to the above-mentioned optimal implementation method. Anyone can derive other forms of energy-saving and emission reduction evaluation methods for pure electric vehicles under complex road linear conditions under the inspiration of this patent. All equal changes and modifications made according to the scope of the patent application of this invention should be covered by this patent.

Claims

1. An explainable prediction method for autonomous driving emissions considering the influence of road geometry conditions, characterized by: The steps include: Step S1: Record the data of the autonomous driving vehicle and the relevant information of the control strategy through on-site real vehicle test experiments, sort out its autonomous driving trajectory and emission data and perform preprocessing to construct a data set; Step S2, using the constructed data set to infer information related to the autonomous driving operating conditions and road geometry; Step S3: Implement feature engineering using machine learning methods; Step S4: Select a machine learning model with good interpretability from a classic machine learning model library for training; Step S5: Evaluate the model prediction performance and interpretability to determine the optimal interpretable prediction model for autonomous driving emissions.

2. The explainable prediction method for autonomous driving emissions considering the influence of road geometry conditions according to claim 1, characterized in that: In step S1, the vehicles selected in the experiment are: commercial self-driving vehicles of different brands and specifications on the existing market; The test site is selected as: a protected closed test site where the terrain slope and curvature of the road fluctuate in a wide range; The experimental design is as follows: autonomous driving vehicles of different brands and specifications use different control strategies and set speeds to drive on the test field, use a portable emission measurement system PEMS to collect emission data, and use the global positioning system GPS and the on-board automatic diagnostic system OBDⅡ to record vehicle trajectory data; In the experiment, the relevant information recorded about the autonomous driving vehicle includes: fuel type, total load mass, maximum power, engine displacement, emission standard, injection type, vehicle type classification, passenger seats and maximum speed; The recorded information related to the automatic driving control strategy includes: adaptive cruise control ACC, headway, ACC set speed, acceleration control strategy and equipped auxiliary driving system; The collected emission data include: instantaneous emission rates of pollutants such as carbon dioxide, carbon monoxide, hydrocarbons, nitric oxide, and nitrogen dioxide; The trajectory data includes: the horizontal coordinate of the vehicle x , vertical coordinate y , vertical coordinate z ,speed V and driving distance S Second-by-second data; The road geometry data includes: road slope G , road curvature C , Road slope change rate GCR , Road curvature change rate CCR ; The operating condition data include: acceleration a , acceleration rate, vehicle specific power VSP .

3. The explainable prediction method for autonomous driving emissions considering the influence of road geometry conditions according to claim 2, characterized in that: In step S1, during data combing, the trajectory data is converted into a three-dimensional rectangular coordinate system.

4. The method for interpretable prediction of autonomous driving emissions considering the influence of road geometry conditions according to claim 2, characterized in that: Step S1 includes the following steps: Step S11, preprocessing the acquired trajectory and emission data, including missing value processing, duplicate value processing, and outlier detection and processing; The missing value processing method is to use the third-order spline interpolation method to fill the missing values; The duplicate value processing method is to find and delete identical duplicate data; The outlier detection and processing method is to use the quartile method IQR to detect outliers and remove them; Step S12: Calculate road geometry data, including road slope G , road curvature C , Road slope change rate GCR , Road curvature change rate CCR ; The road slope G The calculation formula is: ； The road curvature C The calculation formula is: ； The road slope change rate GCR The calculation formula is: ； The road curvature change rate CCR The calculation formula is: ； In the formula, z t for t The vertical coordinate of the moment; ΔS t for t- 1 to t Change in driving distance at any moment; Δx t for t- 1 to t The change of the horizontal axis at a certain moment; Δy t for t- 1 to t The change of vertical coordinate at a certain moment; G t for t The slope of the moment; C t for t The curvature of the moment; Step S13: Calculate the automatic driving operation data, including acceleration a , acceleration rate, vehicle specific power VSP ; The vehicle specific power VSP The calculation formula is: ； In the formula, V t for t Time speed; a t for t Time acceleration; G t for t The road slope at any time.

5. The explainable prediction method for autonomous driving emissions considering the influence of road geometry conditions according to claim 2, characterized in that: Step S2 comprises the following steps: Step S21, encoding the category features related to the autonomous driving vehicle and the control strategy; The features to be encoded include ACC headway, acceleration control strategy, vehicle sensors, vehicle auxiliary driving systems, fuel type, emission standards, vehicle model classification, and injection type, which together with the above recorded and inferred data constitute the original training data set; Step S22: using the geometric path of the road, the trajectory of the autonomous driving vehicle and related information of the control strategy, and the autonomous driving operating conditions as input features to form model input data, using the autonomous driving emissions as output results, and constructing and training multiple machine learning models with embedded feature importance, the machine learning models including XGBoost, lightweight gradient boosting machine, random forest, decision tree, and extreme random tree; Step S23: Use the above machine learning model to output the basic feature importance, and select the machine learning model with the largest proportion of road geometry and autonomous driving control strategy related feature importance as the base model sensitive to the influence of road geometry conditions; Step S24: Based on the model selected in S23, a recursive feature elimination method is used to select the best feature combination with the root mean square error as the cross-validation evaluation index.

6. The explainable prediction method for autonomous driving emissions considering the influence of road geometry conditions according to claim 5, characterized in that: Step S21 encodes the rules for the category features related to the autonomous driving vehicle and the control strategy as follows: If the feature name is ACC headway, the digital coding rule is 1 for short time interval, 2 for medium time interval, and 3 for long time interval. If the feature name is acceleration control strategy, the digital coding rule is 1 for constant headway strategy; 2 for safety distance model strategy; 3 for comfort optimization strategy; 4 for energy consumption optimization strategy; If the feature name is Lane Keeping Assist, the digital coding rule is 1 if not installed and 2 if installed. If the feature name is Lane Centering Control, the digital coding rule is 1 if not equipped and 2 if equipped. If the feature name is automatic emergency braking, the digital coding rule is 1 if not installed and 2 if installed. If the feature name is Traffic Congestion Assist, the digital coding rule is 1 if not equipped and 2 if equipped. If the feature name is blind spot monitoring and assistance, the digital coding rule is 1 if not equipped and 2 if equipped. If the feature name is high-speed assisted driving, the digital coding rule is 1 if not equipped and 2 if equipped. If the feature name is LiDAR, the digital coding rule is 1 if not equipped and 2 if equipped. If the feature name is millimeter wave radar, the digital coding rule is 1 if not equipped and 2 if equipped. If the feature name is camera, the digital coding rule is 1 if not installed and 2 if installed. If the feature name is ultrasonic sensor, the digital coding rule is 1 for unassembled and 2 for assembled. If the feature name is inertial measurement unit, the digital coding rule is 1 if not assembled and 2 if assembled. If the feature name is Global Navigation Satellite System, the digital coding rule is 1 if not equipped and 2 if equipped. If the feature name is fuel type, the digital coding rule is: diesel takes 1; gasoline takes 2; CNG takes 3; LNG takes 4; LPG takes 5; hybrid takes 6; If the characteristic name is emission standard, the digital coding rule is 1 for National I, 2 for National II, and 3 for National III. Country IV takes 4; Country V takes 5; Country VI takes 6; If the feature name is vehicle type classification, the digital coding rule is 1 for light gasoline vehicles and 2 for light diesel vehicles; For heavy-duty gasoline vehicles, take 1; for heavy-duty diesel vehicles, take 2; If the feature name is injection type, the digital coding rule is 1 for carburetor, 2 for electronic injection, 3 for compression ignition, 4 for indirect injection, and 5 for direct injection.

7. The explainable prediction method for autonomous driving emissions considering the influence of road geometry conditions according to claim 2, characterized in that: Step S3 comprises the following steps: Step S31: Using the feature combination selected in step S2, based on the vehicle specific power VSP , road slope G and road curvature C Categorize the data; Step S32: using road geometry, information related to the autonomous driving vehicle and control strategy, and autonomous driving operating conditions as input variables, and autonomous driving emissions as output variables, normalizing the samples consisting of the input variables and the output variables, and randomly dividing the samples according to a preset ratio to form a model training set and a test set; Step S33: select a variety of machine learning models with good interpretability from the classic machine learning model library, and perform parameter optimization through Bayesian optimization and then perform training and testing; The machine learning models include XGBoost, lightweight gradient boosting machine, random forest, decision tree, extreme random tree, support vector machine, deep forest, and ladder boosting tree.

8. The explainable prediction method for autonomous driving emissions considering the influence of road geometry conditions according to claim 7, characterized in that: In step S32, the samples are randomly divided into a ratio of 7:3 as a model training set and a test set respectively.

9. The explainable prediction method for autonomous driving emissions considering the influence of road geometry conditions according to claim 7, characterized in that: Step S4 comprises the following steps: Step S41, constructing a hierarchical structure model, including a target layer and an indicator layer; The target layer is: model interpretability; The indicator layer includes evaluation indicators of feature importance distribution consistency, feature interpretation stability, feature interaction interpretability, global interpretation clarity, and local interpretation stability; The feature importance distribution consistency is: the degree of conformity between the feature importance distribution output by the model and the expert's expectations; The feature interpretation stability is: the stability of the distribution of SHAP values ​​of different samples; The feature interaction interpretability is: the clarity of the model in reflecting the interaction between features; The global explanation clarity is: the clarity of the global explanation of the model SHAP; The local explanation accuracy is: the reasonableness of the SHAP value of a single sample to explain the prediction result; Step S42: construct an importance judgment matrix based on expert scores and perform normalization processing; The importance judgment matrix is ​​formed by comparing and scoring each evaluation index in pairs; The matrix form is: ； The normalization processing formula of the importance judgment matrix is: ； In the formula, the matrix elements d ij For the i The evaluation index is relative to j The importance of the evaluation indicators d norm,ij For the i The evaluation index is relative to j The importance normalization results of the evaluation indicators; Step S43, using the eigenvector method to calculate the weights and perform a consistency check. If the consistency check passes, the weights of each evaluation index are output; if the consistency check fails, the process returns to step S42 to adjust the importance judgment matrix; The weights are calculated using the eigenvector method as follows: Solve for the maximum eigenvalue of the importance judgment matrix λ max And the corresponding eigenvector, normalize the eigenvector to obtain the weight vector W , among which i Elements W i For the i The weight of each evaluation indicator; The consistency test is to calculate the consistency index: CI and consistency ratio CR ,like CR <0.1, the importance judgment matrix has acceptable consistency, otherwise the importance judgment matrix needs to be adjusted; The consistency index CI The calculation formula is: ； The consistency ratio CR The calculation formula is: ； In the formula, λ max is the maximum eigenvalue of the importance judgment matrix; W is the weight vector; W i For the i The weight of each evaluation indicator; CI is the consistency indicator; n is the dimension of the judgment matrix; CR is the consistency ratio; RI is a random consistency index that depends on the dimensions of the importance judgment matrix; Step S44: perform interpretability analysis on each machine learning model based on the SHAP method, invite experts to score the performance of each machine learning model on the above five evaluation indicators based on the interpretability analysis results of each machine learning model, and then calculate the weighted average value as the comprehensive score according to the weight of each evaluation indicator, and sort the interpretability of the model in descending order based on the comprehensive score; The interpretability analysis includes: feature importance analysis, feature interaction analysis, bee swarm diagram analysis, force diagram analysis, decision path analysis, and waterfall diagram analysis; Step S45: sort the models in ascending order based on the prediction performance evaluation index of the single machine learning model, and comprehensively sort the prediction performance of the models according to the average of the sorting; The model prediction performance evaluation indicators include root mean square error, mean absolute error and mean absolute percentage error; Step S46: Select the model with the smallest average values ​​of prediction performance and explainability as the optimal emission prediction model.

10. The explainable prediction method for autonomous driving emissions considering the influence of road geometry conditions according to claim 9, characterized in that: In step S42, the importance judgment matrix is ​​obtained by using a 1-9 scale method, inviting experts to compare and score each evaluation index in pairs, and forming an importance judgment matrix.

Images