A driver driving style judgment model construction system based on physical constraints

Abstract

This application provides a system for constructing a driver driving style determination model based on physical constraints. The system includes a data acquisition module configured to acquire actual road driving data; a data processing module configured to perform K-means clustering calculations based on driving aggression using the actual road driving data to obtain a target clustering result; the K-means clustering calculation process based on driving aggression includes a clustering result correction mechanism; and a model construction module configured to train a driver driving style determination model based on the actual road driving data and the target clustering result to obtain a target driver driving style determination model. The driver driving style determination model is constructed based on an improved loss function with an adaptive load weighting factor. This system solves the problems of models not considering dynamic changes in load ratio, insufficient accuracy in determining multiple load conditions, and difficulty in adapting to different load scenarios.

Description

Technical Field

[0001] This invention relates to the field of transportation technology, and in particular to a driver driving style determination model construction system based on physical constraints. Background Technology

[0002] Driver style, as a core characteristic parameter, plays an indispensable role in areas such as the formulation of energy-saving control strategies for new energy commercial vehicles, optimization of vehicle energy management, efficient operation of logistics fleets, and driver safety evaluation. Accurate driver style determination helps logistics companies develop personalized energy-saving guidance plans, optimize vehicle power control logic, effectively reduce operating energy consumption and battery degradation costs, and improve the economic efficiency and operational safety of logistics transportation. It is a key supporting technology for the intelligent and energy-efficient development of new energy commercial vehicles.

[0003] Current mainstream driving style determination models in the industry are all trained based on real vehicle driving data to achieve intelligent classification, but they generally suffer from two major flaws: First, existing models rely solely on conventional driving parameters such as vehicle speed, pedal opening changes, and acceleration / deceleration characteristics for judgment, without considering the essential impact of changes in commercial vehicle load on driving behavior. Wide-ranging dynamic fluctuations in load are the norm for the operation of new energy commercial vehicles. The same driving style often exhibits differentiated characteristic parameters under different loads, and the same operation has fundamentally different power response and driving intensity under different loads, directly leading to a significant decrease in the accuracy of traditional models in judging different load conditions.

[0004] Secondly, existing models, in terms of algorithm design and training mechanisms, have not been specifically adapted to the core characteristics of dynamic changes in the load of commercial vehicles. They lack systematic modeling of the differences in driving behavior and dynamic response patterns under different loads, and have not integrated load-related constraints into the core training process. As a result, the models are no longer applicable in real-world scenarios such as trunk logistics and mountain freight where loads change frequently. Their judgment accuracy is significantly insufficient, and they cannot meet the judgment requirements under various load conditions, thus failing to meet the standards for practical application. Summary of the Invention

[0005] This application provides a driver driving style determination model construction system based on physical constraints to solve the problems of existing models not considering dynamic changes in load ratio, insufficient accuracy in determining multiple load conditions, and difficulty in adapting to different load scenarios.

[0006] The system includes: The data acquisition module is configured to acquire actual road driving data; the actual road driving data is the driving data of the vehicle in actual driving based on several drivers with different driving styles, taking into account the vehicle load ratio. The data processing module is configured to perform K-means clustering calculation based on driving aggression based on the actual road driving data to obtain the target clustering result; the K-means clustering calculation based on driving aggression has a clustering result correction mechanism. The model building module is configured to train a driver driving style determination model based on the actual road driving data and the target clustering results to obtain a target driver driving style determination model; the driver driving style determination model is built based on an improved loss function, which includes a load adaptive weighting factor.

[0007] Preferably, the actual road driving data includes time, vehicle speed, accelerator pedal travel value, brake pedal status, drive motor speed, drive motor torque, vehicle load capacity, and vehicle rated full load capacity.

[0008] Preferably, the data processing module includes: A feature extraction unit is configured to extract features based on the actual road driving data to obtain actual road driving features. A clustering calculation unit is configured to perform K-means clustering calculation based on driving aggression according to the actual road driving characteristics to obtain clustering results; The result correction unit is configured to correct the clustering results based on the clustering results to obtain the target clustering result; the target clustering result is the driving style corresponding to the actual road driving data.

[0009] Preferably, the feature extraction unit includes: A data segmentation component, configured to segment the actual road driving data into several trip segments according to a fixed duration; A feature calculation component is configured to calculate feature parameters in each of the trip segments to obtain the actual driving characteristics of the vehicle and the driving aggression. A feature standardization component is configured to standardize the actual road driving features and the driving aggression using the Z-score method to obtain the actual road driving features.

[0010] Preferably, the actual driving characteristics of the vehicle include: The following parameters are included: accelerator pedal travel rate of change, brake pedal travel rate of change, longitudinal acceleration, motor speed response rate, average accelerator pedal opening, average brake pedal opening, standard deviation of longitudinal acceleration, standard deviation of vehicle speed, number of rapid accelerations, number of rapid brakings, average motor speed, standard deviation of motor speed, and vehicle load ratio; the vehicle load ratio is the ratio of the actual weight of the vehicle to its rated full load weight in each travel segment. The driving aggression is the product of the average vehicle speed, the absolute value of the average acceleration, and the vehicle load ratio for each travel segment.

[0011] Preferably, the clustering calculation unit includes: Principal component extraction component, configured to perform PCA principal component analysis based on the actual driving characteristics of the vehicle after standardization, to obtain multidimensional principal component features; A clustering calculation component is configured to perform K-means clustering calculation based on the multidimensional principal component features and the standardized driving aggression to obtain several cluster centers; A style segmentation component is configured to segment the driving style of the corresponding trip segments based on the parameter features of all the cluster centers to obtain the clustering result; the clustering result includes the driving style corresponding to each trip segment, and the driving style includes aggressive driving style, stable driving style and conservative driving style.

[0012] Preferably, the result correction unit is further configured as follows: A first threshold for the driving aggression of the aggressive driving style and a second threshold for the driving aggression of the conservative driving style are set respectively; the first threshold is greater than the second threshold, the first threshold is taken as the 90th percentile of the cumulative distribution of driving aggression, and the second threshold is taken as the 10th percentile of the cumulative distribution of driving aggression. In the trip segments where the driving style is either the stable driving style or the conservative driving style, the trip segments where the driving aggression is greater than the first threshold are classified as the aggressive driving style; in the trip segments where the driving style is either the aggressive driving style or the stable driving style, the trip segments where the driving aggression is less than the second threshold are classified as the conservative driving style.

[0013] Preferably, the model building module is further configured as follows: Construct the driver driving style determination model; The driver's driving style determination model is trained based on the actual road driving data, and the predicted driving style is output. The loss function is calculated based on the predicted driving style, the actual driving style in the target clustering results, and the load adaptive weighting factor until the loss value meets the preset requirements, and then the target driver driving style determination model is output.

[0014] Preferably, training the driver's driving style determination model based on the actual road driving data includes: The actual road driving data is divided into a training dataset and a test dataset according to a preset ratio; The driver driving style determination model is trained using the training dataset.

[0015] Preferably, the model building module is further configured as follows: When the loss value meets the preset requirements, the model accuracy of the target driver driving style determination model is verified using the test dataset; when the model accuracy reaches the target requirements, the model construction is completed; when the model accuracy does not reach the target requirements, the target driver driving style determination model is trained again using the training dataset.

[0016] As described above, this application provides a driver driving style determination model construction system based on physical constraints. The system includes a data acquisition module configured to acquire actual road driving data, which considers vehicle load ratio and is based on the driving data of several drivers with different driving styles during actual driving. A data processing module is configured to perform K-means clustering calculation based on driving aggression using the actual road driving data to obtain a target clustering result; the K-means clustering calculation based on driving aggression includes a clustering result correction mechanism. A model construction module is configured to train a driver driving style determination model based on the actual road driving data and the target clustering result to obtain a target driver driving style determination model. The driver driving style determination model is constructed based on an improved loss function, which includes a load-adaptive weighting factor. This application solves the problems of existing models not considering dynamic changes in load ratio, insufficient accuracy in determining multiple load conditions, and difficulty in adapting to different load scenarios through the above system. Attached Figure Description

[0017] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1This is a schematic diagram of a driver driving style determination model construction system based on physical constraints according to this application; Figure 2 This is a schematic diagram of the data processing module in a driver driving style determination model construction system based on physical constraints according to this application; Figure 3 This is a schematic diagram of a feature extraction unit in a driver driving style determination model construction system based on physical constraints according to this application; Figure 4 This is a schematic diagram of a clustering calculation unit in a driver driving style determination model construction system based on physical constraints according to this application. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] It should be noted that the brief descriptions of terms in this application are only for the convenience of understanding the embodiments described below, and are not intended to limit the embodiments of this application. Unless otherwise stated, these terms should be understood in their ordinary and common meaning.

[0021] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0022] Figure 1 This is a schematic diagram of a driver driving style determination model construction system based on physical constraints according to this application.

[0023] Figure 2 This is a schematic diagram of the data processing module in a driver driving style determination model construction system based on physical constraints according to this application.

[0024] Figure 3 This is a schematic diagram of the feature extraction unit in the driver driving style determination model construction system based on physical constraints of this application.

[0025] Figure 4 This is a schematic diagram of a clustering calculation unit in a driver driving style determination model construction system based on physical constraints according to this application.

[0026] See Figures 1 to 4 As can be seen, this embodiment provides a driver driving style determination model construction system based on physical constraints, the system comprising: The data acquisition module is configured to acquire actual road driving data; the actual road driving data is the driving data of the vehicle in actual driving based on several drivers with different driving styles, taking into account the vehicle load ratio.

[0027] Specifically, in this embodiment, actual road driving data of multiple drivers with different driving styles are collected. The data parameters include: time, vehicle speed, accelerator pedal travel value, brake pedal status, drive motor speed, drive motor torque, vehicle load capacity, vehicle rated full load capacity, etc. The data sampling frequency is at least 1Hz.

[0028] The system also includes: The data processing module is configured to perform K-means clustering calculation based on driving aggression based on the actual road driving data to obtain the target clustering result; the K-means clustering calculation based on driving aggression has a clustering result correction mechanism.

[0029] Specifically, in this embodiment, the data processing module includes: The feature extraction unit is configured to extract features based on the actual road driving data to obtain actual road driving features.

[0030] The feature extraction unit includes: A data segmentation component, configured to segment the actual road driving data into several trip segments according to a fixed duration; A feature calculation component is configured to calculate feature parameters in each of the trip segments to obtain the actual driving characteristics of the vehicle and the driving aggression. A feature standardization component is configured to standardize the actual road driving features and the driving aggression using the Z-score method to obtain the actual road driving features.

[0031] Based on time sequence, the collected data is segmented into several travel segments according to a fixed duration, and multi-dimensional feature parameters are calculated. These feature parameters include: accelerator pedal travel rate of change, brake pedal travel rate of change, longitudinal acceleration, motor speed response rate, average accelerator pedal opening, average brake pedal opening, standard deviation of longitudinal acceleration, standard deviation of vehicle speed, number of rapid accelerations, number of rapid brakings, average motor speed, standard deviation of motor speed, and vehicle load ratio. The vehicle load ratio is calculated as the ratio of the vehicle's actual weight to its rated full load weight in each travel segment.

[0032] Building upon this, a driving aggression parameter is introduced to account for changes in driver intensity caused by vehicle load ratio. Specifically, it is calculated as the product of the average speed, the absolute value of average acceleration, and the vehicle load ratio for each trip segment. The driving aggression parameter represents the driver's level of intensity under different load conditions, providing a direct indication of the driver's driving style. Specifically, with the same average speed and average acceleration, a higher load indicates greater driver intensity and a more aggressive driving style. The driving aggression parameter is standardized using the Z-score method, consistent with typical driving behavior characteristics, to eliminate dimensional differences.

[0033] Using the methods described above, a raw fragment database was formed, containing several travel fragments and their characteristics.

[0034] The data processing module further includes: A clustering calculation unit is configured to perform K-means clustering calculation based on driving aggression according to the actual road driving characteristics to obtain clustering results.

[0035] The clustering calculation unit includes: Principal component extraction component, configured to perform PCA principal component analysis based on the actual driving characteristics of the vehicle after standardization, to obtain multidimensional principal component features; A clustering calculation component is configured to perform K-means clustering calculation based on the multidimensional principal component features and the standardized driving aggression to obtain several cluster centers; A style segmentation component is configured to segment the driving style of the corresponding trip segments based on the parameter features of all the cluster centers to obtain the clustering result; the clustering result includes the driving style corresponding to each trip segment, and the driving style includes aggressive driving style, stable driving style and conservative driving style.

[0036] The input data consists of all journey segments from the original segment database, with a default cluster center count of 3. Principal component analysis (PCA) is used to select and retain segments based on a cumulative variance contribution rate ≥ 80%. Principal component features are retained to eliminate differences in feature dimensions and redundant information; the driving aggression feature of vehicle load ratio is not included in PCA dimensionality reduction, but is retained independently as a separate weighting dimension.

[0037] An innovative approach is taken to calculate the distance between pre-selected cluster centers and each sample point during the clustering process. The calculation formula is as follows: ; In the formula, and These are the weight coefficients of the principal component features of conventional driving behavior and the weight coefficients of the driving aggression feature fused with the vehicle load ratio, respectively, and both satisfy the constraints. ; For the first The sample of the first travel segment and the first The weighted Euclidean distance between the cluster centers; is the sample number, with a value range of 1 to M, where M is the total number of travel segment samples in the original segment database; This is the cluster center number, with values ​​ranging from 1 to 3, corresponding to the cluster centers for the three driving styles; This represents the dimension number of the PCA principal component features, with a value ranging from 1 to... , The number of principal component feature dimensions retained after PCA dimensionality reduction; For the first The first sample 3D PCA principal component eigenvalues; For the first The first cluster center PCA principal component characteristic baseline value; For the first The standardized driving aggression value of each sample is an indicator that incorporates vehicle load ratio information; For the first The baseline value for the driving aggression dimension of each cluster center.

[0038] After multiple rounds of iterative calculations, the three cluster centers that minimize the sum of global distances are selected as the final clustering result.

[0039] The calculation process for the three cluster centers here strengthens the influence weight of the driving aggression feature of the vehicle load ratio through weight coefficients. It fully considers the differences in the intensity of driving caused by the dynamic changes in the vehicle load ratio, and avoids the core problem of the traditional K-means clustering algorithm that equally weights all features and does not consider the essential influence of the load ratio on driving behavior, which leads to a serious deviation between the clustering results of driving style under heavy load conditions and actual driving behavior, and misclassification of high-risk aggressive driving behavior.

[0040] Finally, based on the parameter characteristics of each cluster center in the clustering results, the corresponding travel segments were divided into three categories and labeled as aggressive driving style, moderate driving style and conservative driving style, respectively.

[0041] The data processing module further includes: The result correction unit is configured to correct the clustering results based on the clustering results to obtain the target clustering result; the target clustering result is the driving style corresponding to the actual road driving data.

[0042] The overlapping of driving aggression intervals in the clustering results stems from the objective characteristics of continuous and gradual changes in driving behavior and the absence of absolute rigid boundaries between different driving styles. This also aligns with the dynamic fluctuations in the load ratio of commercial vehicles, conforms to the characteristics of real driving scenarios, and possesses sufficient engineering rationality, thus requiring no forced elimination through algorithms. This mechanism only corrects extreme misclassified samples by setting a preset driving aggression threshold, fully preserving the reasonable overlapping intervals of the clustering results.

[0043] First, set driving aggression thresholds for aggressive and conservative driving styles respectively. and ,here The threshold is set using the quantile method: Take the 90th percentile of the cumulative distribution of driving aggression. Take the 10th percentile of the cumulative distribution of driving aggression.

[0044] Secondly, among all the trip segments belonging to the conservative and moderate driving styles, identify those where the driving aggression exceeds the driving aggression threshold. The selected segments of the journey were re-labeled as aggressive driving style; among all the segments belonging to both aggressive and conservative driving styles, those with a driving aggression level below a certain threshold were identified. The segment of the journey was relabeled as a conservative driving style.

[0045] Finally, the re-formulated and corrected clustering results are used as data for constructing a driver driving style determination model that takes into account vehicle load ratio.

[0046] Here, the level of aggressiveness among the three driving styles—aggressive, moderate, and conservative—is allowed to overlap, but the level of aggressiveness must exceed the threshold for aggressiveness. The trip segment must be driven in an aggressive style; the level of aggressive driving must be less than the aggressive driving threshold. The driving style must be conservative for the entire journey.

[0047] The system also includes: The model building module is configured to train a driver driving style determination model based on the actual road driving data and the target clustering results to obtain a target driver driving style determination model; the driver driving style determination model is built based on an improved loss function, which includes a load adaptive weighting factor.

[0048] Specifically, in this embodiment, firstly, an LSTM neural network is used to build a driver driving style determination model, which includes a 3-layer LSTM neural network structure.

[0049] The data acquired for the driver style determination model considering vehicle load ratio was randomly divided into training and testing datasets in a 7:3 ratio. The model was trained using the training dataset and its accuracy was verified using the testing dataset.

[0050] The model's input is the 13-dimensional raw feature parameters of each trip segment after standardization, and the model's output is the driving style label for each trip segment.

[0051] Traditional LSTM models use a cross-entropy loss function with fixed weights for all samples, failing to reflect the differences in driving styles under varying load conditions. To address this, this invention introduces a load-adaptive weighting factor, constructing a load-weighted cross-entropy loss function. This dynamically increases the loss penalty weight for heavily loaded samples during training iterations, allowing the model to pay closer attention to changes in driver style caused by vehicle load. The innovative loss function is as follows: ; In the formula, The global total loss value is the core indicator for measuring the training effect of the model. The smaller the value, the smaller the deviation of the model's prediction results and the higher the judgment accuracy. The total number of training samples; The load-adaptive weighting factor is calculated as follows: ( This is the weighting adjustment coefficient; Basic weights; For the first The core innovation of this loss function is the ratio of vehicle load to weight for each sample (ranging from 0 to 1). This ratio is positively correlated with vehicle load; that is, the higher the vehicle load, the greater the increase in load. The higher the value, the stronger the corresponding penalty; This is the single-sample cross-entropy loss, used to quantify the deviation between the model's predictions and the true labels. For the first Real driving style labels for each sample The loss value is higher for the driving style label predicted by the model and the greater the deviation between the two.

[0052] This design allows the model to focus on prediction accuracy in heavy-load scenarios, prioritize the accuracy of judgment in high-risk scenarios, and also take into account the rationality of empty vehicle scenarios, thus meeting the needs of actual engineering applications.

[0053] Based on the above method, a driver driving style determination model is constructed, and the model accuracy is verified using a test dataset. If the target requirements are met, the model construction is completed; otherwise, the model is retrained. Finally, a driver driving style determination model based on the physical constraint of vehicle load ratio is completed.

[0054] For example: This embodiment provides a driver driving style determination model construction system based on physical constraints. This system strictly adheres to the requirements of the patented technical solution, achieving high-precision determination of driving style under different load conditions for new energy commercial vehicles through a complete process of data collection, feature extraction, clustering and labeling, correction and adjustment, and model construction and verification. The specific implementation steps are as follows: 1. Obtain data for constructing a driver driving style determination model that considers vehicle load ratio: This embodiment uses a mainstream 4.5t pure electric van as the implementation vehicle. This model is widely used in urban logistics distribution, suburban trunk line transportation and other scenarios. The core parameters are as follows: the vehicle's rated full load capacity is 2t, the peak power of the drive motor is 120kW, the rated torque is 800N·m, the rated voltage of the power battery is 540V and the capacity is 180Ah. It is completely consistent with the actual operation configuration of new energy commercial vehicles, ensuring that the implementation process is highly consistent with the actual project. The collected data has strong practicality and promotion value. Based on the daily operating load characteristics of commercial vehicles, the full load range of vehicles is divided into three core operating conditions: empty / light load, medium load, and heavy load. Among them, empty / light load corresponds to a load ratio of 0~0.3 (actual load 0~0.6t), medium load corresponds to a load ratio of 0.3~0.7 (actual load 0.6~1.4t), and heavy load corresponds to a load ratio of 0.7~1.0 (actual load 1.4~2.0t), which fully covers the full load operating range of commercial vehicles when empty, half-loaded, and fully loaded, and fully considers the normal operating conditions of dynamic fluctuations in the load of commercial vehicles.

[0055] 1.1 Collect real-world driving data of vehicles with different driving styles: To ensure comprehensive data coverage of different driving styles, this embodiment selected 50 professional drivers with over 3 years of experience in logistics and transportation as data collection subjects. These drivers ranged in age from 25 to 55, with driving experience spanning 3 to 20 years, encompassing both young and middle-aged individuals, while also considering differences in driving habits. Initial labeling was achieved through a dual approach: a driving behavior questionnaire survey and a simulated driving pre-experiment. Drivers were pre-categorized into three types: aggressive (18 drivers), moderate (20 drivers), and conservative (12 drivers), comprehensively covering different driving style groups and effectively avoiding sample distribution bias. This laid the foundation for the rationality of subsequent clustering and labeling.

[0056] The data collection period is set for 7 consecutive days, taking into account the differences in traffic flow at different times of weekdays and weekends. The driving scenarios comprehensively cover four typical operating scenarios of new energy commercial vehicles: urban congested roads, suburban smooth roads, mountainous undulating roads, and highway trunk roads. Three load conditions are set simultaneously, with each load condition having a cumulative effective mileage of no less than 800 kilometers and a total effective mileage of over 2,400 kilometers. This fully covers the dynamic fluctuations of commercial vehicle load and ensures that the collected data can truly reflect the driving behavior characteristics under different scenarios and loads.

[0057] Data acquisition utilizes the vehicle's existing onboard CAN bus terminal, eliminating the need for additional high-cost sensors and offering a low-cost deployment advantage. The sampling frequency is set to 1Hz, meeting both the minimum sampling standard of the patented technology and compatibility with the conventional acquisition capabilities of mainstream domestic onboard terminals. This avoids data redundancy and increased costs caused by excessively high sampling frequencies, while accurately capturing short-term dynamic changes in driving behavior. The core data parameters collected include: time, vehicle speed, accelerator pedal travel, brake pedal status, drive motor speed, drive motor torque, actual vehicle load, and rated full-load load. These are all key indicators reflecting driving behavior and vehicle load status, providing comprehensive support for subsequent feature extraction.

[0058] After the raw data collection is completed, in order to ensure data quality, system preprocessing work is carried out: invalid parking data such as vehicle engine shutdown, CAN signal loss, and vehicle speed continuously at 0 are removed to avoid interfering with subsequent analysis; the 3σ criterion is used to remove outliers and eliminate invalid samples caused by signal interference to ensure data accuracy and reliability; finally, data cleaning and regularization are completed to form a standardized raw dataset, providing a high-quality foundation for subsequent segmentation and feature extraction.

[0059] 1.2 Fragment segmentation and feature extraction: Based on time sequence, the cleaned data was segmented into several travel segments with a fixed duration of 5 seconds. This duration, verified multiple times, not only matches the short-term dynamic characteristics of driving behavior and accurately captures the driver's instantaneous operating habits, but also adapts to the policy refresh cycle of commercial vehicle onboard control units, demonstrating sufficient engineering rationality. After segmentation, invalid segments with fewer than 5 valid data points were removed to avoid inaccurate feature extraction, ultimately yielding 5281 sets of valid travel segments, which formed the original travel segment library. Among them, there were 1826 sets of samples for unloaded / lightly loaded conditions, 1752 sets for medium-loaded conditions, and 1703 sets for heavy-loaded conditions. The balanced distribution of samples across the three load conditions avoids data bias affecting model training and ensures that the model fully learns the driving behavior characteristics under different load conditions.

[0060] For each effective travel segment, strictly following the patented technical solution, 13 core feature parameters are calculated and extracted, including: accelerator pedal travel change rate, brake pedal travel change rate, longitudinal acceleration, motor speed response rate, average accelerator pedal opening, average brake pedal opening, standard deviation of longitudinal acceleration, standard deviation of vehicle speed, number of rapid accelerations, number of rapid brakings, average motor speed, standard deviation of motor speed, and vehicle load ratio.

[0061] The vehicle load ratio is calculated as the ratio of the actual load to the rated full load (2t) within each travel segment. Therefore, the load ratios for unloaded / lightly loaded conditions are set to 0~0.3, medium load to 0.3~0.7, and heavy load to 0.7~1.0, accurately quantifying the real-time load status and providing core parameters for calculating driving aggression. Based on this, a driving aggression index integrating the vehicle load ratio is introduced to quantify the intensity of driving under different loads. The calculation method is the product of the average vehicle speed, the absolute value of the average acceleration, and the vehicle load ratio for each travel segment. This index intuitively reflects that under the same operation, a higher load ratio indicates higher driving aggression and a more aggressive driving style, aligning with the essential impact of commercial vehicle load on driving energy consumption. It can accurately distinguish between high-risk aggressive driving and conventional operation under heavy load conditions, compensating for the shortcomings of traditional indicators that do not consider the impact of load. The driving aggression index is standardized using the Z-score method to eliminate dimensional differences and obtain standardized values. This provides a unified standard for subsequent cluster analysis, ultimately forming an original fragment database containing 5281 samples and 13-dimensional features, supporting subsequent cluster annotation work.

[0062] 1.3. Based on the driving aggression considering the vehicle load ratio, K-means clustering is performed on all trip segments in the original segment database: In this embodiment, the number of cluster centers is set to 3, corresponding to three types of driving styles: aggressive, moderate, and conservative. This is consistent with the industry's general classification standards and the results of previous pre-experiment calibration, ensuring that the clustering results accurately correspond to the actual driving style type.

[0063] Before clustering calculation, Z-score standardization is performed on the 12-dimensional conventional driving features (except for driving aggression) to eliminate the influence of dimensions and numerical magnitude, ensure that the analysis weights of each feature are balanced, and avoid any one feature dominating the clustering results. The standardization formula is as follows: ; In the formula, For the original value of the feature, The sample mean. The standard deviation of the sample is 1. These are the standardized feature values. After standardization, all common features fall within the same numerical range, providing a reliable basis for subsequent PCA dimensionality reduction and cluster analysis.

[0064] After standardization, 12-dimensional conventional driving features were dimensionality reduced using Principal Component Analysis (PCA) to eliminate multicollinearity and dimensional redundancy, reducing computational load while retaining core features. Calculations showed that the cumulative variance contribution rate of the first six principal components reached 80.37%, meeting the technical requirement of ≥80%, and thus the features were ultimately retained. =6-dimensional principal component features, replacing the original 12-dimensional conventional features, to ensure that core driving features are not lost. Among them, the driving aggression feature does not participate in PCA dimensionality reduction, but is used as an independent weighting dimension to strengthen the constraint effect of load-related features on the clustering results, which is in line with the actual needs of commercial vehicle load affecting driving style.

[0065] The weighted Euclidean distance calculation formula proposed in this invention strengthens the influence of load characteristics through weight coefficients, thus solving the clustering bias problem caused by the traditional K-means equal weighting for medium and heavy loads. The specific formula is as follows: ; In the formula, after optimization and verification using a combination of grid search and contour coefficient method, the following is determined: =0.4 (conventional feature weights) =0.6 (driving aggression weight), which satisfies ; For the first The sample of the first travel segment and the first The weighted Euclidean distance between cluster centers is used; the smaller the distance, the higher the similarity. This is the sample sequence number (ranging from 1 to 5281). This is the cluster center number, with values ​​ranging from 1 to 3, corresponding to the cluster centers for the three driving styles; is the dimension number of the principal component features in PCA, with a value range of 1 to 6, where 6 is the number of principal component features retained after PCA dimensionality reduction; For the first The first sample 3D PCA principal component eigenvalues; For the first The first cluster center PCA principal component characteristic baseline value; For the first The standardized driving aggression value of each sample is an indicator that incorporates vehicle load ratio information; For the first The baseline value for the driving aggression dimension of each cluster center.

[0066] Clustering iteration is based on "cluster center movement distance < 10". -6 The process terminates with either "or a maximum of 1000 iterations" as the termination condition. After multiple iterations, the three cluster centers with the smallest total global distance are obtained as the final result. Based on the cluster center parameters, the samples are divided into three categories and labeled. The clustering results show that out of 5281 samples, 1901 are aggressive (36.0%), 2112 are robust (40.0%), and 1268 are conservative (24.0%), which perfectly matches the previous pre-classification results, confirming the rationality of the clustering.

[0067] The misclassification rate is calculated as: (Number of misclassified samples of a certain class / Total number of samples of that class) × 100%. Compared with traditional algorithms, this invention, after optimization, reduces the misclassification rate from 8.7% to 2.1% for unloaded / lightly loaded aggressive driving, from 12.6% to 2.8% for medium load, and from 21.2% to 3.5% for heavy load, effectively solving the clustering bias problem under medium and heavy load conditions and providing high-quality labeled data for model training.

[0068] 1.4 Establish a clustering result correction mechanism based on driving aggression to verify and correct the clustering results in 1.3: The correction mechanism in this embodiment only corrects extreme misclassified samples, retains reasonable overlap between driving styles, conforms to the continuous and gradual changes in driving behavior and the fluctuation pattern of commercial vehicle load, avoids the disconnect between labeling and actual driving behavior, and has engineering rationality.

[0069] Based on the aggression distribution of the three driving styles in the clustering results, and combined with the differences in load conditions, the threshold is scientifically set using the quantile method: =1.72 (90th percentile of radicalism, threshold for radicalism) =-1.58 (10th percentile for aggressiveness, conservative threshold), satisfies It can effectively distinguish extreme driving behaviors.

[0070] Correction measures: Relabel samples with an aggression level >1.72 in the robust and conservative categories as aggressive to avoid misclassification of extremely aggressive samples; relabel samples with an aggression level <-1.58 in the aggressive and robust categories as conservative to avoid misclassification of extremely conservative samples.

[0071] Ultimately, 87 sets of samples with extreme misclassification were adjusted. Among them, 32 robust sets were corrected to aggressive sets (24 sets from heavy-load conditions, confirming the necessity of strengthening the load weight), and 55 robust sets were corrected to conservative sets (31 sets from unloaded / light-load conditions). After correction, the matching degree between the samples and actual driving behavior improved from 92.1% to 98.6%, and reached 99.2% for heavy-load conditions, forming a high-quality labeled dataset as the basis for model training and testing.

[0072] 2. Construct a driver driving style determination model based on the physical constraints of vehicle load ratio: 2.1 Dataset Partitioning and Model Building: The corrected labeled dataset was randomly divided into a training set (3697 groups, used for parameter training) and a test set (1584 groups, used for accuracy validation) at a 7:3 ratio. This division ratio conforms to standard machine learning practices, ensuring sufficient training and reliable validation. During the division, the proportion of samples from the three load conditions was strictly controlled to be consistent with the original dataset to avoid sample bias leading to insufficient accuracy for any particular condition. To prevent overfitting, 10% of the training set was extracted as an internal validation set, and an early stopping mechanism was used to monitor the training process, promptly terminating invalid iterations to ensure optimal model performance.

[0073] A decision-making model is built using an LSTM neural network, adapted to the temporal characteristics of driving behavior, and effectively captures time-series features. The main body of the model is a 3-layer LSTM structure with 128 neurons per layer. After optimization and verification, this structure balances model performance and computational cost. The Tanh activation function is used to handle non-linear features. A Dropout layer (with a dropout rate of 0.2) is added after each LSTM layer to randomly discard some neurons to suppress overfitting. The output layer uses the Softmax activation function to output the probability distributions and corresponding labels of three driving styles, ensuring accurate and interpretable decision results. The model input is the standardized 13-dimensional original features, and the output is the three driving style labels, which completely correspond to the clustering labels, ensuring consistency between training and data.

[0074] 2.2 Model Loss Function Optimization and Training: To address the limitation of traditional LSTM cross-entropy loss functions in failing to reflect load differences, this invention employs a load-weighted cross-entropy loss function. This function introduces an adaptive load weighting factor to dynamically enhance the loss penalty weights for medium and heavy-load samples, prioritizing accuracy under heavy-load conditions. This aligns with the actual needs of commercial vehicles facing high energy consumption and high risk under heavy loads, and overcomes the shortcomings of traditional models that treat different loads equally. The loss function formula is as follows: ; In the formula, The global total loss value is the core indicator for measuring the training effect of the model. The smaller the value, the smaller the deviation of the model's prediction results and the higher the judgment accuracy. The total number of training samples; The load-adaptive weighting factor is calculated as follows: ( This is the weighting adjustment coefficient; Basic weights; For the first The core innovation of this loss function is the ratio of vehicle load to weight for each sample (ranging from 0 to 1). This ratio is positively correlated with vehicle load; that is, the higher the vehicle load, the greater the increase in load. The higher the value, the stronger the corresponding penalty; This is the single-sample cross-entropy loss, used to quantify the deviation between the model's predictions and the true labels. For the first Real driving style labels for each sample The loss value is higher for the driving style label predicted by the model and the greater the deviation between the two.

[0075] The model training uses the Adam optimizer (fast convergence, adaptive parameter tuning), with an initial learning rate of 0.001 (balancing convergence speed and stability), a batch size of 64, a maximum of 200 iterations, and an early stopping mechanism (termination occurs if the validation set loss does not decrease for 10 consecutive iterations) to avoid overfitting. After training, the accuracy is verified using a test set; if it meets the target, the model is considered complete; otherwise, the parameters are adjusted and retraining is performed.

[0076] 2.3 Model Performance Verification and Application Results: This model achieved an overall accuracy of 96.2% on the test set, a 6.2 percentage point improvement over the traditional LSTM model without load ratio and a 1.7 percentage point improvement over the LSTM-SVM fusion model, demonstrating significant performance advantages. Specifically, the model achieved 97.1% precision and 95.8% recall for the aggressive driving style, 96.5% precision and 97.2% recall for the robust driving style, and 94.8% precision and 95.1% recall for the conservative driving style, demonstrating excellent performance in distinguishing between different driving types.

[0077] To address the pain points in determining variable load conditions, a comparative accuracy test was conducted for three types of load conditions. The accuracy matrix is ​​shown in the table below: Table 1

[0078]

[0079] Test results show that this model maintains stable high accuracy under all load conditions, from no-load to full-load. The average accuracy of the three types of conditions is 6 percentage points higher than that of the traditional model, completely solving the industry pain point that the accuracy of the traditional model drops significantly with increasing load. Among them, the accuracy of the judgment of aggressive driving style under heavy load conditions reaches 98.2% and the recall rate reaches 97.5%, which are 16.4 and 15.7 percentage points higher than the traditional model, respectively, fully demonstrating the core advantage of introducing physical constraints on load ratio.

[0080] Most importantly, this system relies entirely on the vehicle's existing onboard CAN bus terminal to collect data, eliminating the need for additional high-cost sensors. This enables highly accurate and low-cost determination of driving styles for new energy commercial vehicles under all operating conditions. Based on this core capability, the model can directly provide accurate and reliable input for the "determination-adaptation-guidance" closed-loop energy-saving system, supporting the formulation of differentiated energy-saving strategies for different driving styles and load conditions. Through joint simulation verification using Cruise and Simulink, the application of the personalized energy-saving technology corresponding to this invention achieves an overall average energy-saving rate of 2.8%, with the highest energy-saving rate reaching 6.2% for heavy-load aggressive driving, representing a 1.2 percentage point improvement over traditional general energy-saving strategies.

[0081] Real-world vehicle generalization verification results show that the model maintains an overall accuracy rate of 94.7% in different geographical scenarios, including plains and mountainous areas, demonstrating excellent generalization ability. This invention provides a cost-effective solution for determining driving style in commercial vehicle variable load scenarios, directly contributing to the large-scale application of energy-saving technologies in new energy commercial vehicles, effectively reducing operational energy consumption and battery degradation costs for logistics companies, and promoting overall cost reduction and efficiency improvement in the logistics industry.

[0082] The advantages of this embodiment include: This embodiment provides a driver driving style determination model construction system based on physical constraints, solving the core problems of existing driving style determination models that do not consider the dynamic changes in the load ratio of commercial vehicles, have insufficient determination accuracy due to different feature parameters for the same driving style under different load conditions, and are difficult to adapt to different load operation scenarios of new energy commercial vehicles. First, real-world driving data with multiple driving styles and operating conditions is collected, and 13-dimensional core features including the vehicle load ratio are extracted. A driving aggression index fused with the load ratio is introduced to construct an original segment database that fits the operating characteristics of commercial vehicles. Then, a K-means clustering algorithm with optimized weighted Euclidean distance is used to complete the initial classification of driving styles. A clustering result correction mechanism based on driving aggression is established to effectively improve the accuracy of labeled data and avoid extreme misclassification problems. Subsequently, a load-adaptive weighting factor is introduced to construct a load-weighted cross-entropy loss function, optimizing the LSTM determination model, strengthening the feature learning ability under different load conditions, and ensuring the determination accuracy of different load ranges. Finally, it achieves high-precision determination of driving style under all working conditions from empty to fully loaded for new energy commercial vehicles, greatly improving the generalization ability of the model in variable load scenarios. It provides a high-precision and low-cost method for determining the driving style of commercial vehicles in variable load scenarios, which can directly support the formulation of personalized energy-saving strategies and help the large-scale application of energy-saving technologies for new energy commercial vehicles and reduce costs and increase efficiency in the logistics industry.

[0083] For ease of explanation, the above description has been provided in conjunction with specific embodiments. However, the discussion in some embodiments is not intended to be exhaustive or to limit the embodiments to the specific forms disclosed above. Various modifications and variations can be obtained based on the above teachings. The selection and description of the above embodiments are for the purpose of better explaining the contents of this disclosure, thereby enabling those skilled in the art to better utilize the embodiments.

Claims

1. A physical constraint-based driver driving style determination model construction system characterized by, The system includes: The data acquisition module is configured to acquire actual road driving data; the actual road driving data is the driving data of the vehicle in actual driving based on several drivers with different driving styles, taking into account the vehicle load ratio. The data processing module is configured to perform K-means clustering calculation based on driving aggression based on the actual road driving data to obtain the target clustering result; the K-means clustering calculation based on driving aggression has a clustering result correction mechanism. The model building module is configured to train a driver driving style determination model based on the actual road driving data and the target clustering results to obtain a target driver driving style determination model; the driver driving style determination model is built based on an improved loss function, which includes a load adaptive weighting factor.

2. The driver driving style determination model construction system based on physical constraints according to claim 1, characterized by, The actual road driving data includes time, vehicle speed, accelerator pedal travel, brake pedal status, drive motor speed, drive motor torque, vehicle load capacity, and vehicle rated full load capacity. 3.The physical constraint based driver driving style determination model construction system according to claim 1, wherein, The data processing module includes: A feature extraction unit is configured to extract features based on the actual road driving data to obtain actual road driving features. A clustering calculation unit is configured to perform K-means clustering calculation based on driving aggression according to the actual road driving characteristics to obtain clustering results; The result correction unit is configured to correct the clustering results based on the clustering results to obtain the target clustering result; the target clustering result is the driving style corresponding to the actual road driving data.

4. The driver driving style determination model construction system based on physical constraints according to claim 3, characterized by, The feature extraction unit includes: A data segmentation component, configured to segment the actual road driving data into several trip segments according to a fixed duration; A feature calculation component is configured to calculate feature parameters in each of the trip segments to obtain the actual driving characteristics of the vehicle and the driving aggression. A feature standardization component is configured to standardize the actual road driving features and the driving aggression using the Z-score method to obtain the actual road driving features.

5. The physical constraint-based driver driving style determination model construction system according to claim 4, characterized by, The actual driving characteristics of the vehicle include: The following parameters are included: accelerator pedal travel rate of change, brake pedal travel rate of change, longitudinal acceleration, motor speed response rate, average accelerator pedal opening, average brake pedal opening, standard deviation of longitudinal acceleration, standard deviation of vehicle speed, number of rapid accelerations, number of rapid brakings, average motor speed, standard deviation of motor speed, and vehicle load ratio; the vehicle load ratio is the ratio of the actual weight of the vehicle to its rated full load weight in each travel segment. The driving aggression is the product of the average vehicle speed, the absolute value of the average acceleration, and the vehicle load ratio for each travel segment.

6. The physical constraint-based driver driving style determination model construction system according to claim 4, characterized by, The clustering calculation unit includes: Principal component extraction component, configured to perform PCA principal component analysis based on the actual driving characteristics of the vehicle after standardization, to obtain multidimensional principal component features; A clustering calculation component is configured to perform K-means clustering calculation based on the multidimensional principal component features and the standardized driving aggression to obtain several cluster centers; A style segmentation component is configured to segment the driving style of the corresponding trip segments based on the parameter features of all the cluster centers to obtain the clustering result; the clustering result includes the driving style corresponding to each trip segment, and the driving style includes aggressive driving style, stable driving style and conservative driving style.

7. The driver driving style determination model construction system based on physical constraints according to claim 6, characterized in that, The result correction unit is further configured to: A first threshold for the driving aggression of the aggressive driving style and a second threshold for the driving aggression of the conservative driving style are set respectively; the first threshold is greater than the second threshold, the first threshold is taken as the 90th percentile of the cumulative distribution of driving aggression, and the second threshold is taken as the 10th percentile of the cumulative distribution of driving aggression. In the trip segments where the driving style is either the stable driving style or the conservative driving style, the trip segments where the driving aggression is greater than the first threshold are classified as the aggressive driving style; in the trip segments where the driving style is either the aggressive driving style or the stable driving style, the trip segments where the driving aggression is less than the second threshold are classified as the conservative driving style.

8. The driver driving style determination model construction system based on physical constraints according to claim 7, characterized in that, The model building module is also configured to: Construct the driver driving style determination model; The driver's driving style determination model is trained based on the actual road driving data, and the predicted driving style is output. The loss function is calculated based on the predicted driving style, the actual driving style in the target clustering results, and the load adaptive weighting factor until the loss value meets the preset requirements, and then the target driver driving style determination model is output.

9. The driver driving style determination model construction system based on physical constraints according to claim 8, characterized in that, The step of training the driver driving style determination model based on the actual road driving data includes: The actual road driving data is divided into a training dataset and a test dataset according to a preset ratio; The driver driving style determination model is trained using the training dataset.

10. The driver driving style determination model construction system based on physical constraints according to claim 9, characterized in that, The model building module is also configured to: When the loss value meets the preset requirements, the model accuracy of the target driver driving style determination model is verified using the test dataset; when the model accuracy reaches the target requirements, the model construction is completed; when the model accuracy does not reach the target requirements, the target driver driving style determination model is trained again using the training dataset.

Images