Subjective evaluation system and method for vehicle road test

Abstract

The invention discloses a subjective evaluation system and method for a vehicle road test, and the system comprises a data collection module which is used for synchronously collecting a vehicle operation objective parameter set and driver multi-mode subjective feedback data in a vehicle road test process; the data processing module is used for at least performing time alignment, denoising and feature extraction processing on the data; the model building module is used for training a mapping model by utilizing the extracted features so as to build a mapping relation between objective parameters and subjective experience; and the index quantification module is used for the mapping model to map the input objective parameters and subjective parameters into quantified index values. According to the technical scheme, subjective scores and objective data are fused, driver experience combination is established, the effects of comprehensively and objectively evaluating driver experience and improving evaluation accuracy and reliability are achieved, meanwhile, the method can be applied to optimization of vehicle design and performance improvement, and data support is provided for optimization of vehicle design and performance improvement.

Description

Technical Field

[0001] This invention belongs to the technical field of vehicle performance testing, specifically relating to a subjective evaluation system and method for vehicle road testing. Background Technology

[0002] In the automotive industry, driver experience evaluation is a crucial basis for vehicle design and performance improvement. With technological advancements, higher demands are placed on driver experience evaluation methods, requiring not only accurate reflection of driver subjective feelings but also scientific analysis incorporating objective data. However, traditional driver experience evaluation methods often have limitations, failing to comprehensively and objectively reflect the driver's true experience.

[0003] In related technologies, driver experience evaluation primarily relies on subjective assessments, such as questionnaires and interviews. While these methods can directly capture driver feedback, they are highly subjective and difficult to quantify and objectively compare. Furthermore, some objective data collection methods, such as vehicle performance testing and sensor monitoring, while providing some objective evidence, are often disconnected from driver subjective feelings and cannot directly reflect the driver's experience. This leads to vehicle evaluations either focusing solely on vehicle data or solely on driver perception, hindering scientific and effective vehicle improvement and optimization. Summary of the Invention

[0004] The purpose of this invention is to propose a subjective evaluation system and method for vehicle road testing to solve the problems in the prior art.

[0005] Therefore, the present invention provides a subjective evaluation system for vehicle road testing, comprising: The data acquisition module is used to simultaneously collect the set of objective parameters of vehicle operation and multimodal subjective feedback data of the driver during the road test of the vehicle. The data processing module is communicatively connected to the data acquisition module and is used to perform at least time alignment, noise reduction, and feature extraction processing on the acquired objective parameter set and subjective feedback data. The model building module is communicatively connected to the data processing module and is used to train a mapping model based on machine learning algorithms and extracted features to establish a mapping relationship between objective parameters and subjective experience. The index quantification module is used to map the input objective parameters and subjective parameters into quantified index values ​​using a trained mapping model.

[0006] In some embodiments, the data acquisition module includes at least an objective data acquisition module and a subjective data acquisition module; The objective data acquisition module includes at least a vibration acquisition unit, a noise acquisition unit, a temperature acquisition unit, and a CAN bus for acquiring control signals of the ACC system and actual vehicle response signals; The subjective data acquisition module includes at least a sound acquisition unit, an image acquisition unit, an eye-tracking acquisition unit, a bioelectric signal acquisition unit, and a subjective scoring acquisition unit.

[0007] In some embodiments, the data processing module includes a data preprocessing module and a feature extraction module. The data preprocessing module is used to perform at least cleaning, noise reduction and normalization processing on the objective parameter set and the subjective feedback data. The feature extraction module is used to extract key features from the objective parameter set and the subjective feedback data.

[0008] In some embodiments, the model building module includes a model training module, which trains the preprocessed data based on a machine learning algorithm to establish a mapping between the objective parameter set and the subjective feedback data; It also includes a model validation module, which inputs a validation set into the constructed model to evaluate the accuracy and generalization ability of the data and outputs the validation results; It also includes a model optimization module, which adjusts the model based on the verification results.

[0009] In some embodiments, the indicator quantification module includes an indicator definition module, which is used to define the numerical value of the experience indicator; The indicator calculation module, based on the module, quantifies the objective parameter set and subjective feedback into indicator values.

[0010] On the other hand, a vehicle road testing method is also provided, including: Collect objective parameters of vehicle operation and multimodal subjective feedback data from drivers; The objective parameter set and the subjective feedback data are processed, wherein the data processing includes at least time alignment, noise reduction and feature extraction. A mapping model is established based on the machine learning algorithm and the objective parameter set after data processing, and the subjective feedback data. The mapping model is input into an objective parameter set and subjective feedback data, and the mapping model maps them into indicator values.

[0011] In some embodiments, the data processing of the objective parameter set and the subjective feedback data includes: The collected set of objective parameters and the subjective feedback data are preprocessed, wherein the preprocessing includes at least cleaning, denoising and normalizing the data; Key features are extracted from the set of objective parameters and the subjective feedback data.

[0012] In some embodiments, establishing a mapping model based on machine learning algorithms and the processed set of objective parameters and the subjective feedback data includes: Extract a first feature vector from the preprocessed set of objective parameters, and extract a second feature vector from the preprocessed multimodal subjective feedback data; The training sample set is input into a machine learning algorithm for training, and the second feature vector or the subjective evaluation label derived therefrom is used as the training target to establish a mapping model from the first feature vector to the subjective experience.

[0013] In some embodiments, the first feature vector extracted from the set of objective parameters contains at least two or more of the following features: Vehicle NVH parameters include the effective value, peak frequency, and amplitude of vibration acceleration in a predetermined frequency band; Steady-state error, temperature fluctuation amplitude, and rate of change of air conditioner outlet temperature; The response delay time and its statistical distribution characteristics of the Adaptive Cruise Control (ACC) system; Extracting a second feature vector from the multimodal subjective feedback data includes extracting one or more features from the following data sources: Extract keyword frequency and contextual sentiment features from driver voice data based on speech recognition text; Extract the probability or duration of a preset emotion category from the driver's facial video data based on an expression recognition algorithm; Extract the percentage of fixation time in a preset area of ​​interest, the length of the saccade path, and the characteristics of pupil diameter changes from the driver's eye-tracking data; Heart rate variability indicators and skin conductance response characteristics were extracted from the driver's bioelectrical signals.

[0014] In some embodiments, inputting the mapping model with an objective parameter set and subjective feedback data, and mapping the mapping model to index values ​​includes: The processed set of objective parameters is input into the mapping model; The mapping model performs forward propagation calculations based on the weights and bias parameters stored internally and determined through training, mapping the input objective parameters and outputting one or more quantified user experience metric values.

[0015] Beneficial effects: This invention solves the problems of incompleteness in existing evaluation methods, difficulty in quantification due to the disconnect between subjective and objective data, and insufficient scientific rigor in testing methods by integrating subjective ratings and objective data to establish a joint driver experience technical solution. It achieves the effect of comprehensively and objectively evaluating driver experience and improving the accuracy and reliability of the evaluation. At the same time, it can be applied to optimize vehicle design and performance improvement, providing data support for optimizing vehicle design and performance improvement. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This invention provides a timing diagram for the synchronous acquisition of multiple vehicle parameters in the subjective evaluation system for vehicle road testing.

[0018] Figure 2 This is a timing diagram of subjective feedback data acquisition in the subjective evaluation system for vehicle road testing provided by the present invention.

[0019] Figure 3 Flowchart of the vehicle road testing method provided by the present invention Detailed Implementation The invention will be more readily understood by referring to the following detailed description of preferred embodiments and included examples. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. In case of conflict, the definitions in this specification shall prevail.

[0020] like Figure 1-2 As shown, a subjective evaluation system for vehicle road testing includes: The data acquisition module is used to simultaneously collect objective parameters of vehicle operation and multimodal subjective feedback data of the driver during the road test of the vehicle. This ensures that, at the same time reference, it can acquire physical signals reflecting the vehicle's own state and record various feedback information generated by the driver during driving. This provides a simultaneous and consistent data foundation for subsequent correlation analysis between vehicle performance and human feelings.

[0021] The data processing module, which is communicatively connected to the data acquisition module, is used to perform at least time alignment, noise reduction, and feature extraction processing on the acquired objective parameter set and subjective feedback data. This solves the problem of slight time misalignment of data acquired by different devices, removes noise introduced by environmental interference, and extracts key information that represents the essential characteristics of the data, thus preparing for model training.

[0022] The model building module, communicatively connected to the data processing module, is used to train a mapping model based on machine learning algorithms and extracted features to establish a mapping relationship between objective parameters and subjective experience. By analyzing a large number of paired samples of vehicle features and corresponding driver feedback features, it automatically identifies the hidden and complex mathematical patterns between the two, thereby constructing a computational model that can infer driver feelings based on vehicle data. The mapping model refers to a mathematical function containing many adjustable parameters, such as a neural network. The training process involves continuously adjusting these parameters so that when a set of vehicle features is input, the function's output value is as close as possible to the actually observed driver feedback rating or status label.

[0023] The index quantification module utilizes a trained mapping model to map input objective and subjective parameters into quantified index values. By deploying the trained model in real-world applications, it automatically outputs one or more specific scores for new test data, transforming the originally vague, qualitative driving experience into clear, comparable quantitative results. Specifically, the processed new vehicle feature data is input into a pre-trained, fixed model. The model calculates based on its internally defined parameters, ultimately generating a numerical index, such as a comprehensive comfort score of 7.5, or multiple numerical indices, such as comfort score of 8.0 and convenience score of 6.5.

[0024] At the same time, comparative analysis can be conducted based on user experience indicators of different vehicles and different drivers to identify differences and patterns, and these user experience indicators can be applied to vehicle design and performance improvement, thereby optimizing product functions and user experience.

[0025] In one embodiment, the data acquisition module includes at least an objective data acquisition module and a subjective data acquisition module; The objective data acquisition module 1 includes at least a vibration acquisition unit, a noise acquisition unit, a temperature acquisition unit, and a CAN bus for acquiring control signals of the ACC system and actual vehicle response signals; The subjective data acquisition module includes at least a sound acquisition unit, an image acquisition unit, an eye-tracking acquisition unit, a bioelectric signal acquisition unit, and a subjective scoring acquisition unit.

[0026] Specifically, the vibration acquisition unit is installed in key locations such as the vehicle's engine compartment, driver's cabin, and chassis, and preferably employs high-precision vibration sensors and noise sensors to collect the vehicle's NVH data. For example: The vibration sensors are triaxial accelerometers, and the noise sensors are microphones. At least six microphones are used, specifically two in the engine compartment, two in the cockpit, and two in the chassis. The sampling frequency is at least 1000 Hz, and the data acquisition time for each test is at least 30 minutes.

[0027] For example: The temperature acquisition unit consists of temperature sensors positioned at different locations on the air conditioner vent to monitor temperature changes in real time. At least four thermocouples are used as temperature sensors, positioned at the upper left, upper right, lower left, and lower right of the vent. The sampling frequency is 10Hz, and the acquisition time for each test is at least 30 minutes.

[0028] For example, the control signals of the ACC system and the actual response signals of the vehicle are acquired via the vehicle's CAN bus, and the response delay time is calculated. The acquisition frequency is at least 100Hz, and the response delay is calculated as the time difference between the issuance of the control signal and the actual response of the vehicle. The acquisition time for each test is at least 30 minutes.

[0029] Additionally, the voice acquisition unit in the subjective data acquisition module 1 is, exemplarily, an in-vehicle microphone. This microphone records the driver's voice feedback during the test. One microphone is located within the cockpit. Its sampling frequency is 16kHz, and the acquisition time for each test is at least 30 minutes.

[0030] The image acquisition unit is preferably a vehicle-mounted camera, which records the driver's facial expressions for subsequent analysis of emotional state. One camera is installed in the cockpit, facing the driver's face. The camera has a resolution of at least 1080p and a frame rate of at least 30fps. Each test session lasts at least 30 minutes.

[0031] The eye-tracking acquisition unit preferably uses an eye tracker to record the driver's gaze focus and fixation time. One eye tracker is installed in the cockpit, directly facing the driver's eyes, and its sampling frequency is 100Hz. The acquisition time for each test is at least 30 minutes.

[0032] The bioelectric signal acquisition unit is preferably a bioelectric signal sensor, specifically including a heart rate monitor and a skin conductance sensor, used to monitor the driver's heart rate and physiological indicators such as skin conductance. The heart rate monitor has a sampling frequency of at least 1 Hz, and the skin conductance sensor has a sampling frequency of at least 10 Hz. The acquisition time for each test is at least 30 minutes.

[0033] The subjective rating collection unit is preferably a questionnaire scoring system. After the test, drivers fill out a questionnaire to rate the vehicle's performance and driving experience. Scoring items can include multiple dimensions such as comfort, safety, handling, and satisfaction. A 5-point or 7-point scale is used for scoring. At least 10 points must be surveyed in each test.

[0034] In one embodiment, the data processing module includes a data preprocessing module and a feature extraction module. The data preprocessing module performs at least cleaning, denoising, and normalization processing on the objective parameter set and the subjective feedback data; transforming the collected raw, messy, and dimensionlessly variable data into clean, well-organized standard data that can be directly used for subsequent analysis. This eliminates the influence of invalid data segments and interference signals, and converts all feature values ​​to a uniform scale, laying a reliable data foundation for accurate correlation analysis and model training.

[0035] For example, the data acquisition module synchronously records data, specifically including the raw waveform of vertical vibration collected by the accelerometer located on the seat rail, at 1000 sampling points per second; and the vehicle speed signal read from the CAN bus, at 10 sampling points per second. Subjective feedback data includes the driver's voice recorded by the in-vehicle microphone and the video stream captured by the facial camera.

[0036] System checks revealed an abnormal zero value in the vibration signal at timestamp 125, lasting 0.1 seconds due to momentary interference. The module automatically marked this 0.1-second data segment as invalid and discarded it; this gap will be skipped in subsequent analysis. A low-pass digital filter with a cutoff frequency of 100Hz was applied to the original vibration waveform to filter out components above 100Hz, typically caused by minor high-frequency unevenness in the road surface and having minimal impact on human comfort, retaining the main low-frequency vibration information. The filtered vibration acceleration values ​​(in m / s²) and vehicle speed values ​​(in km / h) were linearly normalized to the [0,1] interval based on the maximum and minimum values ​​of the entire test data, respectively. For example, a vibration value of 0.5 m / s² and a vehicle speed value of 60 km / h were converted to 0.35 and 0.6, respectively.

[0037] The feature extraction module is used to extract key features from the objective parameter set and the subjective feedback data. This achieves data dimensionality reduction and information condensation. Instead of directly using massive amounts of raw data points, it calculates or identifies representative indicators that can represent the core characteristics and are highly correlated with the driver's experience, greatly improving the efficiency and effectiveness of subsequent model training.

[0038] For example, for the vibration data within the aforementioned 30 seconds, divided into 3-second sub-windows, the feature extraction module calculates the following features for each sub-window: Calculates the root mean square value of the vibration acceleration within each 3-second sub-window. Performs a Fast Fourier Transform on the vibration data of each 3-second sub-window to identify the frequency component with the highest energy as the dominant frequency. Uses the average normalized vehicle speed of each 3-second sub-window as a feature. Simultaneously, the audio within the 30 seconds is converted into text using speech recognition, and the frequency of negative words related to smoothness, such as "shaking" or "bumping," is counted. An facial expression recognition algorithm is used to analyze the driver's facial video within the 30 seconds, calculating the proportion of frames showing expressions that may indicate discomfort, such as "frowning" or "pursing lips," out of the total frames. Finally, for this 30-second data set, the feature extraction module outputs a feature vector. This vector represents a digital summary of the vehicle's state and the driver's feedback during that time period.

[0039] Through the above modules, various raw signals that are not easy to use directly are transformed into standardized feature datasets with clear structure and explicit meaning, which can be used to train the mapping relationship between objective vibration characteristics and subjective discomfort feedback.

[0040] In one embodiment, the model building module includes a model training module. The model training module trains the preprocessed data based on machine learning algorithms to establish a mapping between the set of objective parameters and the subjective feedback data. By utilizing historical data, namely the objective dataset and subjective feedback data mentioned above, it automatically learns and summarizes the complex and non-linear correspondence between the vehicle's objective performance and the driver's subjective feelings, thereby forming a computational tool that can predict the latter based on the former.

[0041] For example, the system has a total of 800 valid data samples. Each sample contains objective features extracted from a 3-second driving clip, such as the root mean square of vibration and the dominant frequency, and corresponding subjective labels, such as the driver's smoothness rating for the clip, from 1 to 5. These data are randomly divided into three parts: training set (640 sets, accounting for 80%), validation set (120 sets, accounting for 15%), and test set (40 sets, accounting for 5%).

[0042] Random forest regression was chosen as the machine learning algorithm. 640 sets of objective features and subjective ratings from the training set were input into the algorithm. By constructing a large number of decision trees, each tree learned the relationship between objective features and ratings. The training objective was to make the average of the predictions from all decision trees as close as possible to the actual driver ratings. After a predetermined number of iterations, a preliminary ride comfort prediction model A was obtained.

[0043] It also includes a model validation module, which inputs a validation set into the constructed model to evaluate the accuracy and generalization ability of the data and outputs the validation results. This module is used to objectively verify whether the trained model is reliable. It not only checks the model's fit to known data (accuracy), but more importantly, it evaluates whether it can still make correct predictions when faced with new data, such as generalization ability, to prevent the creation of a model that can only memorize training samples and has no practical value.

[0044] For example, 120 sets of objective features from the validation set are input into the newly trained model A. The model outputs 120 predicted scores for these 120 sets of data. The validation module calculates the root mean square error (RMSE) between these predicted scores and the true scores, which is 0.82. Simultaneously, it calculates the Pearson correlation coefficient between the predictions and the true values, which is 0.75. The module outputs the validation results: RMSE = 0.82, correlation coefficient = 0.75.

[0045] It also includes a model optimization module, which adjusts the model based on the validation results. Based on the problems exposed during the validation phase, targeted improvements are made to the model to enhance its final performance, making its predictions more accurate and stable. For example, adjusting the model's hyperparameters, such as changing the number of hidden layers or neurons in the neural network, adjusts the model's learning ability and complexity. Secondly, adjusting the training strategy, such as adding a regularization penalty to the loss function, constrains the size of the model parameters, thereby reducing the risk of the model becoming too complex and overfitting.

[0046] For example, analyzing the validation results, it was determined that the root mean square error (RMSE) was too large, indicating room for improvement in the model's prediction accuracy. An attempt was made to change the two main hyperparameters of the random forest: the number of decision trees and the maximum number of features allowed per tree. Based on the validation set, various hyperparameter combinations were automatically tried, followed by a brief retraining and validation. It was found that when the number of trees was 150 and the maximum number of features was the square root, the lowest RMSE (0.71) was achieved on the validation set, with the correlation coefficient increasing to 0.80. Based on this, the optimization module generated a set of optimized hyperparameters and instructed the model training module to use these new parameters and the complete training set data to retrain a new smoothness prediction model. The new smoothness prediction model was finally tested using a test set (40 sets) that had not participated in the training and optimization process. The results showed an RMSE of 0.73 and a correlation coefficient of 0.79, demonstrating performance close to and good compared to the validation set results.

[0047] In one embodiment, the indicator quantification module includes an indicator definition module, which is used to define numerical values ​​for experience indicators. This transforms abstract user experience goals into one or more specific, calculable mathematical goals, providing a clear measurement standard and interpretive basis for the output of the entire system. Defining an experience indicator means explicitly specifying a numerical variable representing a certain dimension of experience and its calculation rules. For example, the smoothness index can be defined as a value between 0 and 100, with higher values ​​indicating a smoother and less bumpy driving experience perceived by the driver. Its initial calculation rules (before model establishment) can be tentatively set as linear transformations of relevant subjective questionnaire scores.

[0048] The indicator calculation module, based on the module, quantifies the objective parameter set and subjective feedback into indicator values. By utilizing the established mapping model and mapping relationships, it automatically transforms newly collected, complex, multi-source test data directly into simple scores with clear business significance, thereby making the evaluation results more intuitive and operable.

[0049] like Figure 3 As shown, the present invention also provides a vehicle road testing method, which includes: The system collects a set of objective parameters for vehicle operation and multimodal subjective feedback data from the driver. For example, sensors and data acquisition devices are arranged on a selected test vehicle according to a predetermined plan. Triaxial accelerometers are installed at the engine mounts and driver's seat rails to collect vibration data, thermocouple temperature sensors are placed at the air conditioning vents, and control and response signals from the adaptive cruise control (ACC) system are read in real time via the vehicle's CAN bus interface, thus constituting the collection of the objective parameters for vehicle operation. Simultaneously, a high-definition camera facing the driver is installed inside the vehicle to record facial video, a near-infrared eye tracker tracks the driver's gaze, an onboard microphone records the entire audio input, and the driver wears a wristband that monitors heart rate and skin conductance, thus constituting the collection of multimodal subjective feedback data from the driver. All devices are controlled by a central acquisition unit to ensure millisecond-level synchronization of the data stream.

[0050] The objective parameter set and the subjective feedback data are processed, including at least time alignment, denoising, and feature extraction. For example, the time alignment step uses a unified timescale to precisely align vibration peaks, air conditioning temperature fluctuations, ACC delay pulses, and the driver's exclamation, frowning expression frame, and gaze shift signal at a specific moment to the same time point. The denoising step uses a digital filter to remove electrical noise from the vibration signal and eliminates environmental wind noise from the voice recording. The feature extraction step calculates key features from the aligned and cleaned data, extracting the weighted root mean square value and dominant frequency within each 3-second time window from the vibration signal; extracting the frequency of keywords such as "bumpy" and "shaky" from the voice text; calculating the duration percentage of the "frowning" expression from the video using an image recognition algorithm; and statistically analyzing the proportion of fixation time on bumpy road areas from eye-tracking data.

[0051] Based on machine learning algorithms and the processed set of objective parameters and subjective feedback data, a mapping model is established. For example, hundreds of objective features (vibration, temperature, ACC delay features) and corresponding subjective features (keyword frequency, frowning ratio, gaze ratio) within the aforementioned 3-second time window are used as paired training samples. A random forest regression algorithm is used for training, aiming at the comprehensive score of subjective features, allowing the algorithm to learn the complex relationship between objective features and subjective feelings. After training, a reserved validation set is used for testing, the root mean square error between the predicted value and the actual score is calculated, and the model is optimized by adjusting the number and depth of the tree models, ultimately obtaining a stable smoothness experience prediction model.

[0052] The mapping model is input into an objective parameter set and subjective feedback data, which are then mapped into index values. For example, during the model application phase, objective parameter features (such as a new root mean square vibration sequence) collected from another SUV of the same type on the same road segment and processed in the same way are input into the deployed ride comfort prediction model. The model automatically runs its internal algorithm, outputting a series of corresponding predicted values. These predicted values ​​are linearly converted into a ride comfort index ranging from 0 to 100. The system-generated report shows that the vehicle's average index during the test was 72 points, but the index plummeted to 48 points when traversing a specific damaged asphalt section.

[0053] Based on this report, and combined with the corresponding high-vibration raw data for that period, the engineers clearly defined the optimization objective as improving the attenuation performance of the vehicle suspension under such road impacts, thus achieving a quantitative transformation from multi-source heterogeneous data to clear, executable engineering instructions.

[0054] In one embodiment, a first feature vector is extracted from the set of objective parameters that have been synchronously acquired and preprocessed by time alignment, denoising, and normalization.

[0055] In this embodiment, from a 10-minute urban road test data set, the following objective features were extracted in 30-second analysis windows: the effective value of the 1-80Hz frequency band of the seat rail vertical vibration, the standard deviation of the air conditioning vent temperature fluctuation range, and the average response delay time of the ACC system to changes in the speed of the vehicle in front. These values, after standardization, together constitute a first feature vector representing the vehicle's state during that time period, for example, [0.85, 0.42, 1.15] (corresponding to the normalized vibration, temperature fluctuation, and delay values, respectively).

[0056] Extract the corresponding second feature vector from the multimodal subjective feedback data after synchronous preprocessing.

[0057] In this embodiment, for the same 30-second time window, the system extracts the frequency of complaining words (such as "bumpy" and "noisy") in the driver's speech through speech recognition and sentiment analysis; and calculates the percentage of frames in which the driver shows "discomfort" or "focus" expressions through a facial expression recognition model. From eye-tracking data, the system extracts the percentage of total time the driver's gaze is taken away from the road ahead (e.g., looking at the central control screen or out the window). These subjective metrics are combined into a second feature vector, for example [0.08, 0.15, 0.03] (corresponding to the frequency of complaints, the percentage of uncomfortable expressions, and the percentage of gaze deviation, respectively).

[0058] Supervised learning model training is performed. A training sample set is composed of hundreds of such time windows (each window contains a first feature vector and a second feature vector).

[0059] In this embodiment, the training objective is set using a combination of two methods: first, the second feature vector is directly used as a multi-dimensional objective for regression fitting; second, experts assign a comprehensive "comfort level" label (e.g., A / B / C / D) to each window based on the overall pattern of the second feature vector. Support Vector Regression (SVR) is selected as the machine learning algorithm, using the aforementioned first feature vector (e.g., [0.85, 0.42, 1.15]) as input and the corresponding second feature vector or comprehensive label as the output objective for model training. The algorithm learns the complex nonlinear mapping relationship from objective states such as "vibration, temperature, and delay" to subjective feedback or comprehensive levels such as "complaint frequency, facial expression, and gaze" by finding the optimal hyperplane.

[0060] The performance of the initial model was evaluated using a reserved validation set (approximately 20% of the data). For example, given the objective features of the validation set, the model predicted subjective features or grades, which were then compared to the real data. Calculations showed an accuracy of 78% for predicting the overall grade. The model optimization module was then initiated, adjusting key hyperparameters of the SVR (such as the penalty coefficient C and the kernel parameter gamma) through grid search and using cross-validation to find the optimal combination. After optimization, the new model achieved an accuracy of 85% on the validation set, and the correlation coefficients between the predicted and actual values ​​of each subjective feature component exceeded 0.8.

[0061] Finally, a deployable mapping model is generated.

[0062] In one embodiment, the first feature vector extracted from the set of objective parameters contains at least two or more of the following features: Vehicle NVH parameters include the effective value, peak frequency, and amplitude of vibration acceleration in a predetermined frequency band; Steady-state error, temperature fluctuation amplitude, and rate of change of air conditioner outlet temperature; The response delay time and its statistical distribution characteristics of the Adaptive Cruise Control (ACC) system; Extracting a second feature vector from the multimodal subjective feedback data includes extracting one or more features from the following data sources: Extract keyword frequency and contextual sentiment features from driver voice data based on speech recognition text; Extract the probability or duration of a preset emotion category from the driver's facial video data based on an expression recognition algorithm; Extract the percentage of fixation time in a preset area of ​​interest, the length of the saccade path, and the characteristics of pupil diameter changes from the driver's eye-tracking data; Heart rate variability indicators and skin conductance response characteristics were extracted from the driver's bioelectrical signals.

[0063] In one embodiment, inputting the mapping model with an objective parameter set and subjective feedback data, and mapping the mapping model to index values ​​includes: The processed set of objective parameters is input into the mapping model; The mapping model performs forward propagation calculations based on the weights and bias parameters stored internally and determined through training, mapping the input objective parameters and outputting one or more quantified user experience metric values.

[0064] For example, for an evaluation window (such as a continuous 60 seconds starting from the 10th minute), the current objective feature vector containing the following information is extracted: the effective value of low-frequency vibration of the seat rail (normalized to 0.72), the temperature fluctuation amplitude of the air conditioning vents (0.35), the vibration amplitude of the steering wheel center area (0.18), and the fluctuation coefficient of ACC following distance control (0.41). This feature vector X_current=[0.72,0.35,0.18,0.41,...] is the input to the mapping model.

[0065] Subsequently, the index quantization module 4 calls a pre-loaded, fully trained and validated neural network mapping model (e.g., a multilayer perceptron with two hidden layers). This model internally stores fixed, trained parameters, including weight matrices W1, W2, W3 and bias vectors b1, b2, b3 connecting the neurons in each layer.

[0066] The forward propagation computation process of the model is entirely determined by these fixed parameters. The input layer receives the vector X_current. The activation values ​​of the first hidden layer are calculated: H1 = ReLU(W1 * X_current + b1). The weights W1 and bias b1 determine how the input features are combined and mapped to the intermediate features represented by the neurons in the first layer. The activation values ​​of the second hidden layer are calculated: H2 = ReLU(W2 * H1 + b2). The output layer is then computed: Y_raw = W3 * H2 + b3. Here, the weights W3 and bias b3 act as scorers, synthesizing high-level features into a final score.

[0067] In this calculation, the model output layer produces two raw values, for example, Y_raw=[6.8,7.5]. According to the definitions at the time of model deployment, these two values ​​correspond to the Highway Comfort Index (HCI) and the Driving Convenience Index (DCI), respectively. They are then mapped to a standard scale of 0-100 points using a fixed linear scaling.

[0068] Ultimately, the system output quantified user experience metrics: HCI = 68 points, DCI = 75 points. Simultaneously, the system automatically generated a time-series curve, showing that during the entire test, the HCI briefly dropped to 52 points when the vehicle entered a strong crosswind zone. Engineers could immediately pinpoint the problem based on this quantified result: combining the raw data, they discovered that the decrease in HCI was highly synchronized with the sudden increase in steering wheel vibration characteristic values ​​at that time, thus clearly identifying optimizing steering wheel vibration suppression under high-speed crosswinds as a specific improvement task based on quantified user experience metrics.

[0069] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A subjective evaluation system for vehicle road testing, characterized in that, include: The data acquisition module is used to simultaneously collect the set of objective parameters of vehicle operation and multimodal subjective feedback data of the driver during the road test of the vehicle. The data processing module is communicatively connected to the data acquisition module and is used to perform at least time alignment, noise reduction, and feature extraction processing on the acquired objective parameter set and subjective feedback data. The model building module is communicatively connected to the data processing module and is used to train a mapping model based on machine learning algorithms and extracted features to establish a mapping relationship between objective parameters and subjective experience. The index quantification module is used to map the input objective parameters and subjective parameters into quantified index values ​​using a trained mapping model.

2. The vehicle road test subjective evaluation system according to claim 1, characterized in that, The data acquisition module includes at least an objective data acquisition module and a subjective data acquisition module; The objective data acquisition module includes at least a vibration acquisition unit, a noise acquisition unit, a temperature acquisition unit, and a CAN bus for acquiring control signals of the ACC system and actual vehicle response signals; The subjective data acquisition module includes at least a sound acquisition unit, an image acquisition unit, an eye-tracking acquisition unit, a bioelectric signal acquisition unit, and a subjective scoring acquisition unit.

3. The vehicle road test subjective evaluation system according to claim 1, characterized in that, The data processing module includes a data preprocessing module and a feature extraction module. The data preprocessing module is used to perform at least cleaning, noise reduction and normalization processing on the objective parameter set and the subjective feedback data. The feature extraction module is used to extract key features from the objective parameter set and the subjective feedback data.

4. The vehicle road test subjective evaluation system according to claim 1, characterized in that, The model building module includes a model training module, which trains the preprocessed data based on machine learning algorithms to establish a mapping between the objective parameter set and the subjective feedback data. It also includes a model validation module, which inputs a validation set into the constructed model to evaluate the accuracy and generalization ability of the data and outputs the validation results; It also includes a model optimization module, which adjusts the model based on the verification results.

5. The vehicle road test subjective evaluation system according to claim 1, characterized in that, The indicator quantification module includes an indicator definition module, which is used to define the numerical values ​​of experience indicators. The indicator calculation module, based on the module, quantifies the objective parameter set and subjective feedback into indicator values.

6. A method for road testing of vehicles, characterized in that, include: Collect objective parameters of vehicle operation and multimodal subjective feedback data from drivers; The objective parameter set and the subjective feedback data are processed, wherein the data processing includes at least time alignment, noise reduction and feature extraction. A mapping model is established based on the machine learning algorithm and the objective parameter set after data processing, and the subjective feedback data. The mapping model is input into an objective parameter set and subjective feedback data, and the mapping model maps them into indicator values.

7. The vehicle road testing method according to claim 6, characterized in that, The data processing of the objective parameter set and the subjective feedback data includes: The collected set of objective parameters and the subjective feedback data are preprocessed, wherein the preprocessing includes at least cleaning, denoising and normalizing the data; Key features are extracted from the set of objective parameters and the subjective feedback data.

8. The vehicle road testing method according to claim 6, characterized in that, The establishment of the mapping model based on the machine learning algorithm and the processed objective parameter set and the subjective feedback data includes: Extract a first feature vector from the preprocessed set of objective parameters, and extract a second feature vector from the preprocessed multimodal subjective feedback data; The training sample set is input into a machine learning algorithm for training, and the second feature vector or the subjective evaluation label derived therefrom is used as the training target to establish a mapping model from the first feature vector to the subjective experience.

9. The vehicle road testing method according to claim 8, characterized in that, The first feature vector extracted from the set of objective parameters contains at least two or more of the following features: Vehicle NVH parameters include the effective value, peak frequency, and amplitude of vibration acceleration in a predetermined frequency band; Steady-state error, temperature fluctuation amplitude, and rate of change of air conditioner outlet temperature; The response delay time and its statistical distribution characteristics of the Adaptive Cruise Control (ACC) system; Extracting a second feature vector from the multimodal subjective feedback data includes extracting one or more features from the following data sources: Extract keyword frequency and contextual sentiment features from driver voice data based on speech recognition text; Extract the probability or duration of a preset emotion category from the driver's facial video data based on an expression recognition algorithm; Extract the percentage of fixation time in a preset area of ​​interest, the length of the saccade path, and the characteristics of pupil diameter changes from the driver's eye-tracking data; Heart rate variability indicators and skin conductance response characteristics were extracted from the driver's bioelectrical signals.

10. The vehicle road testing method according to claim 6, characterized in that, The step of inputting the mapping model into an objective parameter set and subjective feedback data, and mapping the mapping model into indicator values, includes: The processed set of objective parameters is input into the mapping model; The mapping model performs forward propagation calculations based on the weights and bias parameters stored internally and determined through training, mapping the input objective parameters and outputting one or more quantified user experience metric values.

Images