A precise evaluation method for ride comfort based on subjective and objective data fusion

Abstract

The application discloses a kind of based on subjective and objective data fusion's ride comfort precision evaluation method, comprising the following steps: step 1, real-time synchronous acquisition passenger comfort evaluation data and vehicle kinematics data;Step 2, the passenger comfort evaluation data and vehicle kinematics data obtained in step 1 are preprocessed, the passenger comfort evaluation data with subjective score 0 and the vehicle kinematics data corresponding time period of time stamp are deleted, the vehicle kinematics data is detected and replaced, low-pass filtering and signal smoothing processing are carried out;Step 3, establish the improved MSDV model as objective evaluation model, obtain MSDV model output value;Step 4, based on the subjective and objective fusion model of combination of unscented Kalman filter and neural network is obtained after the comfort score of subjective and objective fusion score.

Description

Technical Field

[0001] This invention relates to the field of vehicle ride comfort evaluation technology, and in particular to a precise ride comfort evaluation method based on the fusion of subjective and objective data. By jointly processing objective vehicle kinematic data and subjective passenger ratings, the accuracy and stability of comfort evaluation are improved, and this method can be applied to fields such as autonomous vehicles, intelligent transportation systems, and comfort optimization control. Background Technology

[0002] Existing methods for evaluating ride comfort mainly fall into two categories:

[0003] 1. Subjective rating method: Passengers rate their comfort level based on their own experience using a specific rating scale or handheld rating device. However, due to individual differences in perception and the discrete nature of ratings, this method has low accuracy and consistency.

[0004] 2. Objective Evaluation Method: This method collects kinematic data such as acceleration and angular velocity from the vehicle's inertial navigation system (INS) and uses the motion sickness dose value (MSDV) model from the ISO 2631-1:1997 standard to quantify comfort. Although this method has high repeatability, it ignores the subjective experience of passengers.

[0005] Using the methods described above alone cannot accurately measure actual ride comfort; therefore, it is necessary to integrate subjective and objective data to improve the accuracy of the evaluation. However, there is a strong nonlinear relationship between subjective and objective data, and the nonlinear mapping relationship between them cannot be directly described by commonly used functional relationships. Traditional linear filtering algorithms (such as standard Kalman filtering) cannot handle this effectively. Therefore, there is an urgent need for a fusion method that can combine subjective and objective data and effectively handle nonlinear relationships. Summary of the Invention

[0006] The purpose of this invention is to address the technical deficiencies in the existing technology by providing a method for accurately evaluating ride comfort based on the fusion of subjective and objective data.

[0007] The technical solution adopted to achieve the purpose of this invention is:

[0008] A method for accurately evaluating ride comfort based on the fusion of subjective and objective data includes the following steps:

[0009] Step 1: Real-time synchronous collection of passenger comfort evaluation data and vehicle kinematic data. The passenger comfort evaluation data includes subjective scores, and the vehicle kinematic data includes longitudinal acceleration, lateral acceleration, vertical acceleration, longitudinal angular velocity, lateral angular velocity, and vertical angular velocity.

[0010] Step 2: Preprocess the passenger comfort evaluation data and vehicle kinematics data obtained in Step 1. Delete the passenger comfort evaluation data with a subjective score of 0 and the vehicle kinematics data of the corresponding time period. Perform outlier detection and replacement, low-pass filtering and signal smoothing on the vehicle kinematics data.

[0011] Step 3: Establish an improved MSDV model as an objective evaluation model: Where MSDV is the output value of the improved MSDV model, a w,i (t) is the frequency-weighted value of acceleration and angular velocity, k i Here are the weighting coefficients for each dimension, and T is the total time length;

[0012] Step 4: Establish a subjective-objective fusion model based on the combination of unscented Kalman filtering and neural networks:

[0013] Step 4.0: Determine the initial uncertainty P0 of the system, the system process noise covariance Q(i) as a function of subjective rating levels, and the perceived error covariance R of subjective ratings using statistical methods. t ;

[0014] Step 4.1, Unscented Kalman Filter (UKF) Prediction Stage: Use the MSDV model output value from the previous time step as the comfort prior estimate to obtain the MSDV prediction value at the current time step;

[0015] Step 4.2, in the prediction phase, based on the error covariance P of the previous time step. t-1 Calculate the prior error covariance P at the current time. t - ;

[0016] Step 4.3: Fit the nonlinear observation equation using a multilayer neural network. The mapping relationship between the MSDV prediction value obtained in step 4.1 and the subjective rating obtained in step 1 is studied.

[0017] Step 4.4, Unscented Kalman Filter (UKF) Observation Update Stage: A trained multilayer neural network is used as the nonlinear observation equation. The calculated values ​​are first input from the output values ​​of the MSDV model after Sigma point transformation. The predicted observation sigma point was calculated. Then calculate the model observation prediction mean. Observation error covariance P zz and the covariance P between the state and the observation xz ;

[0018] Step 4.5, using the observation error covariance P obtained in step 4.4 zz and the covariance P between the state and the observationxz Calculate the Kalman gain K t ;

[0019] Step 4.6, based on the Kalman gain K obtained in step 4.5 t The model obtained in step 4.4 is used to predict the observed mean. The deviation between the subjective rating obtained in step 1 and the MSDV prediction obtained in step 4.1 is corrected to obtain the comfort rating after integrating subjective and objective ratings.

[0020] Step 4.7, based on the Kalman gain K obtained in step 4.5 t Update the prior error covariance P obtained in step 4.2 t - The updated error covariance P is obtained. t .

[0021] In the above technical solution, in step 4.0, the initial uncertainty P0 is calculated using the following formula:

[0022]

[0023] n i It is the number of data points with a subjective rating of i, MSDV ij It is the MSDV model output value of the j-th sample with subjective rating i, MSDV i It is the mean of the MSDV model output values ​​for subjective rating i, and M is the total number of all rating levels;

[0024] System process noise covariance

[0025] Perceived error covariance of subjective ratings in, It can be calculated from experimental statistical data.

[0026] In the above technical solution, in step 4.1,

[0027] This is the MSDV prediction value at the current moment. It is the posterior estimate of comfort obtained after the end of the previous moment, U. t It is the input at the current moment. A and B are the state transition matrix and the control matrix, respectively.

[0028] In the above technical solution, in step 4.2, P t - =AP t-1 A T +Q t ;

[0029] Where: P t - It is the prior error covariance at the current moment, P t-1 It is the posterior error covariance of the previous time step, Q. t Q is the process noise covariance at the current moment; t Based on adaptive adjustment of subjective ratings, Q t =Q(i).

[0030] In the above technical solution, in step 4.3, the nonlinear observation equation

[0031] in: The subjective rating given by passengers at time t; To obtain the MSDV prediction value in step 4.1; This represents the nonlinear mapping relationship learned by the neural network; v t To observe the noise, assume it follows a zero-mean normal distribution: v t ~N(0,R t ).

[0032] In the above technical solution, in step 4.3, the multilayer neural network includes an input layer, a hidden layer, and an output layer. The input of the input layer includes the output value of the MSDV model, the one-hot encoded values ​​of the passenger's physiological characteristics and physical condition. The hidden layer adopts a three-layer hidden layer structure, all using ReLU as the activation function. The output layer uses the Softmax activation function to calculate the probability of each rating category.

[0033]

[0034] Where, p i The predicted probability of the i-th type of rating calculated by the neural network. This represents the unnormalized score output by the neural network.

[0035] In the above technical solution, in step 4.3, the training of the multilayer neural network is optimized using the cross-entropy loss function: Among them, y i The one-hot encoding of the true rating represents that if the true rating is k, then y k =1, other categories are 0, p i Let C be the predicted probability of the i-th class rating calculated by the neural network, and C be the number of classes.

[0036] In the above technical solution, in step 4.4, the model observes and predicts the mean. and observation error covariance P zz The covariance P between state and observationxz Calculated using the following formula:

[0037]

[0038] in, W represents the model's observed predicted mean. i (m) Weighted by mean, To predict the observed sigma point;

[0039]

[0040] Among them, W i (c) Covariance weights;

[0041]

[0042] in, It is the calculated value after the MSDV model output value has undergone Sigma point transformation.

[0043] In the above technical solution, in step 4.5, K t This is the Kalman gain.

[0044] In the above technical solution, in step 4.6,

[0045] To achieve a comfort score that integrates subjective and objective ratings, K is the predicted value of MSDV. t The Kalman gain determines the degree to which subjective ratings correct for the state, z. t These are the actual observed values, i.e., the subjective scores obtained in step 1. The deviation term represents the difference between the actual observed value and the model-predicted mean observed value. The deviation between them.

[0046] In the above technical solution, in step 4.7,

[0047] Where: P t For the updated error covariance, P t - For the prior error covariance, K t For Kalman gain, P zz This represents the observation error covariance.

[0048] Compared with the prior art, the beneficial effects of the present invention are:

[0049] 1. This invention uses an improved MSDV model: based on the ISO 2631 standard, the MSDV calculation model is extended, and the frequency weighting of six dimensions of acceleration and angular velocity is comprehensively considered to improve the accuracy of objective comfort evaluation.

[0050] 2. This invention uses statistical concepts to model the perceived error of subjective ratings: it utilizes experimental statistical analysis of the rating distribution of different passenger groups in different scenarios, adopts a normal distribution to model the rating error, and calculates the rating variance as perceived noise.

[0051] 3. This invention uses a neural network to fit the nonlinear observation equation: It utilizes a neural network to fit the nonlinear mapping relationship between the MSDV model and subjective ratings, thereby solving the problem of nonlinear correspondence between subjective and objective data.

[0052] 4. This invention uses unscented Kalman filtering to fuse subjective and objective data: Unscented Kalman filtering (UKF) is used to fuse subjective scores with objective MSDV model output values, thereby improving the accuracy of the evaluation. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating the fusion of subjective and objective elements in an embodiment of the present invention.

[0054] Figure 2 This is a comparison chart of data before and after preprocessing in an embodiment of the present invention.

[0055] Figure 3 This is a comparison diagram of the full-band frequency weighting before and after in an embodiment of the present invention.

[0056] Figure 4 This is a comparison chart of the frequency weighting effect of longitudinal acceleration signals based on the ISO-2631 standard in an embodiment of the present invention.

[0057] Figure 5 This is a neural network mapping framework diagram of the MSDV model and the true value in an embodiment of the present invention.

[0058] Figure 6 This is a heatmap of the rating frequency of each group under two specific scenarios in this embodiment of the invention.

[0059] Figure 7 This is a subjective rating distribution diagram of different passenger groups in typical scenarios in an embodiment of the present invention. Detailed Implementation

[0060] The present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0061] Example 1

[0062] like Figure 1 As shown, a method for accurately evaluating ride comfort based on the fusion of subjective and objective data includes the following steps:

[0063] Step 1: Real-time synchronous collection of passenger comfort evaluation data (subjective data) and vehicle kinematics data (objective data);

[0064] Passenger comfort evaluation data collection:

[0065] The system utilizes a specially designed mobile webpage for evaluation. Upon initial login, passengers are required to fill in physiological information such as gender, age, and physical condition. These parameters will be used in subsequent models to personalize evaluation data for different groups. Once inside the vehicle, passengers automatically log into the evaluation platform via the in-vehicle local area network. During the ride, passengers receive system prompts at preset time intervals (e.g., every 15 or 20 seconds) to rate their current comfort level in real time, typically ranging from 1 to 5 points. Each passenger comfort evaluation (subjective rating) is accompanied by a precise timestamp to ensure accurate temporal alignment with vehicle kinematic data.

[0066] Vehicle kinematics data acquisition:

[0067] An inertial navigation system (INS), GPS positioning equipment, and related inertial sensors are installed inside the vehicle to collect the vehicle's dynamic parameters in real time. These dynamic parameters include longitudinal acceleration, lateral acceleration, vertical (upward) acceleration, longitudinal angular velocity, lateral angular velocity, and vertical (upward) angular velocity. The collected vehicle kinematic data is transmitted to a central server in real time via an onboard Wi-Fi wireless network, and then processed in conjunction with subjective comfort evaluation data for subsequent model training.

[0068] The above data acquisition system is based on the Ubuntu 20.4 operating system and uses a real-vehicle algorithm platform jointly built with Python 3 and C++. The platform realizes real-time data acquisition, storage, and preliminary analysis, and ensures that all data contains accurate time synchronization information, guaranteeing high-precision alignment during the fusion of subjective and objective data.

[0069] Step 2, as follows Figure 2 As shown, the passenger comfort evaluation data and vehicle kinematics data obtained in step 1 are preprocessed;

[0070] Data alignment and invalid score removal: The passenger comfort evaluation data obtained in step 1 is checked. For data that is recorded as "0 points" due to passengers' failure to respond in time or distraction, the timestamp is used to locate the data and the vehicle kinematic data of the corresponding time period is deleted simultaneously, thereby ensuring strict consistency between subjective and objective data in time.

[0071] Outlier Detection and Replacement: The Z-score method is used to detect outliers in vehicle kinematics data. Specifically, the standard deviation of each data point from the mean is calculated, and a threshold (usually 3) is set. Data points exceeding the threshold are considered outliers and replaced using linear interpolation of adjacent normal data to avoid the impact of outliers on subsequent analysis.

[0072] Low-pass filtering: Considering that the vehicle kinematic data is INS data, which contains high-frequency noise introduced by sensor electronic noise and environmental vibration, the system uses a first-order low-pass filter for data filtering. In the filter design, Fourier transform (FFT) is performed on the original data to analyze the signal spectrum, confirming that the effective signal is mainly concentrated in the 0-5Hz range. The cutoff frequency is set to 5Hz, thereby effectively filtering out high-frequency noise and retaining the low-frequency effective signal generated by vehicle motion.

[0073] Signal smoothing: To further reduce fluctuations in vehicle kinematic data, the system employs a moving average algorithm to smooth the low-pass filtered signal. This method calculates local averages for adjacent data points, resulting in smoother data that more accurately reflects the vehicle's dynamic state.

[0074] The data acquisition in this embodiment has the following characteristics:

[0075] Real-time performance: Passenger comfort evaluation data and vehicle kinematics data are collected in real time, and each data point is timestamped to ensure data timeliness and synchronization.

[0076] Multi-source data fusion: By collecting data from both mobile web and on-board sensors, complementary fusion of subjective and objective data is achieved, providing comprehensive data support for subsequent comfort evaluation.

[0077] High-precision data preprocessing: Through data alignment, outlier processing, low-pass filtering and smoothing, the data is made to ensure that it truly reflects the vehicle's motion and passenger experience, reduce noise interference, and improve the accuracy of subsequent model fusion.

[0078] Step 3: Establish an improved MSDV model as an objective evaluation model.

[0079] To more comprehensively assess occupant comfort, the influence of angular velocity needs to be considered in addition to acceleration. Based on this, this invention extends the MSDV model in ISO 2631 by adding angular velocity as an evaluation factor and considering six dimensions of acceleration and angular velocity. The calculation formula is as follows:

[0080]

[0081] Where MSDV is the improved MSDV model output value, i.e., the improved motion sickness dose value, a w,i (t) is the frequency-weighted value of acceleration and angular velocity, such as Figure 3 , Figure 4 As shown (first, the original acquired time-domain signal is decomposed into frequency-domain signals of different frequency bands by performing a Fourier transform, then weighted according to the different frequency bands, and finally synthesized and subjected to an inverse Fourier transform to obtain the total time-domain frequency-domain weighted value), k i These are the weighting coefficients for each dimension.

[0082] Step 4: Establish a subjective-objective fusion model based on the combination of unscented Kalman filtering and neural networks:

[0083] In comfort evaluation, subjective and objective evaluations each have their advantages and disadvantages. Subjective evaluations can directly reflect passengers' comfort feelings, but they are easily affected by individual differences and perceptual biases, and may contain certain subjective errors. Objective evaluations are based on physical measurements (such as acceleration and angular velocity) and mathematical models (such as the MSDV model), and have high stability and repeatability, but they are difficult to directly characterize an individual's true comfort experience. Therefore, a single data source is insufficient to accurately assess passengers' actual comfort level. To quantify comfort more accurately, this study proposes a subjective-objective fusion model, aiming to combine the intuitiveness of subjective evaluations with the stability of objective evaluations to obtain a more accurate comfort evaluation standard. The core idea is to use the weighted fusion mechanism of Kalman filtering to establish a dynamic balance between subjective and objective evaluation data, making the final fusion result closer to passengers' true comfort feelings.

[0084] In the process of integrating subjective and objective factors in this invention, it is assumed that passengers' subjective comfort rating has a high subjective perception ability, but there is a perception error; the objective comfort rating is calculated based on the improved MSDV model and is affected by modeling error; the goal is to obtain a comfort standard rating that is closer to the true value.

[0085] Step 4.0: Determine the initial uncertainty P0 of the system, the system process noise covariance Q(i) as a function of subjective rating levels, and the perceived error covariance R of subjective ratings using statistical methods. t ;

[0086] To determine the initial uncertainty P0 of the system, the following steps are taken based on statistical methods:

[0087] Obtain a large number of subjective ratings and their corresponding MSDV values ​​to construct a sample set. Calculate the variance of the MSDV values ​​for each rating level:

[0088]

[0089] n iIt is the number of data points with a subjective rating of i, MSDV ij It is the MSDV model output value of the j-th sample with subjective rating i, MSDV i It is the mean of the output values ​​of the MSDV model with subjective rating i;

[0090] Take the mean of the MSDV variances for all rating levels as P0:

[0091]

[0092] Where M is the total number of all rating levels.

[0093] Based on the subjective rating level settings, let: Q(i) is the MSDV variance corresponding to the current rating level i, used to set Q. t Q t It represents the dynamic uncertainty of the system state as it changes over time.

[0094] This allows different rating levels to correspond to different process noise levels, reflecting the uncertainty of comfort ratings under different conditions. Q t It is dynamic, not a fixed value, allowing the system to adaptively adjust based on subjective ratings. Changes in rating levels affect the uncertainty of the MSDV value, thereby optimizing state estimation.

[0095] Step 4.1, Unscented Kalman Filter (UKF) Prediction Stage: Using the current MSDV model output value as the comfort prior estimate, the predicted MSDV value for the next time step is obtained.

[0096] First, let's look at the formula for calculating MSDV. Discretization yields:

[0097]

[0098] Where MSDV(t) is the prior estimate of comfort at the current time, and MSDV(t-1) is the posterior estimate of comfort obtained after the end of the previous time step. It is the input at the current moment, a w (τ) is the comprehensive frequency weighting value, which is the general term for frequency weighted signals (that is, different weighting processes are applied to different frequency domains of a time domain signal, corresponding to different human body sensitivities to comfort in different frequency bands).

[0099] However, the comfort rating obtained by the model is a general model that does not take into account group differences, thus containing some error, but it can usually reflect the true value to a certain extent. Therefore, in the prediction stage, the objective MSDV model output value is used as the prior estimate of comfort:

[0100]

[0101] in: This is the MSDV prediction value at the current moment. It is the posterior estimate of comfort obtained after the end of the previous moment, U. t It is the input at the current moment, and the specific expression is: A and B are the state transition matrix and control matrix, respectively, both of which are one-dimensional unit vectors in this invention. 1×1 ;

[0102] Step 4.2, in the prediction phase, based on the error covariance P of the previous time step. t-1 Calculate the prior error covariance P at the current time. t - During the prediction phase, the prior error covariance at the current time is calculated, i.e.:

[0103] P t - =AP t-1 A T +Q t

[0104] Where: P t - The prior error covariance at the current moment, i.e., the uncertainty of the prediction. P t-1 The posterior error covariance of the previous time step, i.e., the error covariance updated after observation. Q t The process noise covariance at the current moment is used to describe the uncertainty of the system model.

[0105] By using the above methods, we can ensure that the calculation of prediction error covariance accurately reflects the dynamic uncertainty of MSDV calculation and improve the robustness of Kalman filtering in the fusion of subjective and objective data.

[0106] Step 4.3, as follows Figure 5 The diagram shows the fitting of nonlinear observation equations using a neural network. The mapping relationship between the MSDV predicted value obtained in step 4.1 and the subjective rating obtained in step 1 is as follows:

[0107] In this invention, due to the significant nonlinear relationship between the subjective ratings (discrete values ​​1-5) obtained in step 1 and the MSDV model output values ​​(continuous values) obtained in step 3, traditional linear models cannot directly establish a correspondence between the two. Therefore, a neural network is used for fitting to learn the mapping relationship between the MSDV model output values ​​obtained in step 3 and the subjective ratings obtained in step 1, ultimately establishing a nonlinear observation equation:

[0108]

[0109] in: The subjective rating given by the passenger at time t (discrete value 1 to 5); To obtain the MSDV predicted value (continuous value) in step 4.1; This represents the non-linear mapping learned by the neural network, used to describe the relationship between subjective ratings and the output value of the MSDV model; v t To observe the noise, assume it follows a zero-mean normal distribution: v t ~N(0,R t ); where R t The covariance of perceived error represents subjective ratings, and the magnitude of perceived error is measured by the variance of the ratings. in, It can be calculated from experimental statistical data.

[0110] Neural Network Structure Design: To ensure that the neural network can effectively learn the nonlinear mapping relationship between the MSDV model output value and subjective ratings, this study constructs a multi-layer neural network, the structure of which is as follows:

[0111] (1) Input layer:

[0112] The neural network inputs include: MSDV value (1 input neuron), passenger physiological characteristics (including gender, age group, and physical condition), gender (2D one-hot encoding), age group (3D one-hot encoding: youth / middle-aged / elderly), and physical condition (2D one-hot encoding: prone to motion sickness / not prone to motion sickness), totaling 8 input neurons. One-hot encoding is used to process gender, age group, and physical condition to ensure the reasonableness of the input feature values.

[0113] (2) Hidden layer:

[0114] A three-layer hidden layer structure is adopted, consisting of: Layer 1: 32 neurons, using ReLU (Rectified Linear Unit) as the activation function; Layer 2: 16 neurons, using ReLU as the activation function; Layer 3: 8 neurons, using ReLU as the activation function.

[0115] (3) Output layer:

[0116] The output layer contains 5 neurons, corresponding to the 5 categories (1-5) of the subjective rating. A softmax activation function is used to calculate the probability of each rating category:

[0117]

[0118] in, This represents the unnormalized score (logits) output by the neural network. A Softmax output layer is used to ensure that the output follows a probability distribution, thereby improving classification accuracy.

[0119] Training a neural network:

[0120] The training data comes from subjective ratings collected in the experiment and the corresponding MSDV model output values. It mainly includes input data and label data. The input data includes MSDV model output values ​​and passenger physiological characteristics (8 dimensions). The label data includes subjective ratings (1 to 5, using one-hot encoding).

[0121] Since subjective ratings are discrete classification labels, the neural network is trained using the cross-entropy loss function for optimization.

[0122]

[0123] The cross-entropy loss function is used to improve the model's convergence speed. Where: y i The one-hot encoding representation of the true rating (if the true rating is k, then y) k =1, other categories are 0). p i This represents the predicted probability of the i-th class rating calculated by the neural network. C is the number of classes, which in this embodiment is C = 5 (corresponding to the ratings).

[0124] Step 4.4, Unscented Kalman Filter (UKF) Observation Update Stage: Using the trained neural network Calculate the mean of observed predictions from the model And calculate the observation error covariance P. zz and the covariance P between the state and the observation xz ;

[0125] In the Unscented Kalman Filter (UKF) observation update phase, a trained multilayer neural network is used as the nonlinear observation equation. The calculated MSDV value after Sigma point transformation is first input. Then through neural network mapping Calculate and predict the observed sigma points Finally, calculate the mean of the model's observed predictions. and the covariance P between the state and the observation xz .

[0126] Model observation prediction mean Calculated using the following formula Among them, W i (m) It is the mean weight.

[0127] Observation error covariance P zzCalculated using the following formula: Among them, W i (c) This represents the covariance weight.

[0128] Covariance P between state and observation xz Calculated using the following formula:

[0129] Step 4.5, using the observation error covariance P obtained in step 4.4 zz and the covariance P between the state and the observation xz Calculate the Kalman gain K t ;

[0130] This embodiment solves the mapping problem between subjective and objective data by learning the non-linear relationship between MSDV values ​​and subjective ratings through a neural network. (Observation noise R) t The variance of subjective ratings is calculated to ensure the rationality of the noise model. The neural network mapping is used as the observation equation for the unscented Kalman filter, improving the accuracy of the filter estimation.

[0131] In the process of integrating subjective and objective comfort, the output value of the MSDV model is predicted through the state transition equation, while the subjective rating, as an observed value, has significant uncertainty due to discretization and individual perception errors. Therefore, in order to optimize the state estimation, it is necessary to calculate the Kalman gain to reasonably balance the uncertainty between the predicted value (MSDV calculation result) and the observed value (subjective rating).

[0132] Kalman gain K t This is used to dynamically adjust the predicted value, varying the degree of correction for subjective ratings under different circumstances. When the subjective rating error is large, the system relies more on the calculated MSDV value; conversely, when the MSDV calculation error is large, the system relies more on the correction of subjective ratings. The calculation formula is as follows:

[0133]

[0134] Step 4.6, based on the Kalman gain K obtained in step 4.5 t The MSDV prediction value obtained in step 4.1 is corrected by using the deviation between the predicted observation value obtained in step 4.4 and the subjective rating obtained in step 1 to obtain the comfort rating after integrating subjective and objective ratings.

[0135]

[0136] in, This is the updated state estimate, which is the comfort score after integrating subjective and objective scores. This is the MSDV prediction value obtained in step 4.1. Kt The Kalman gain determines the degree to which subjective ratings correct for the state. The deviation term represents the difference between the actual observed value (the subjective score in step 1) and the mean of the model's observed predictions.

[0137] The purpose of this step is to use subjective rating information to correct the comfort score calculated by MSDV, so that the final estimate is closer to the actual comfort level and the accuracy of the evaluation is improved.

[0138] Step 4.7, based on the Kalman gain K obtained in step 4.5 t Update the prior error covariance P obtained in step 4.2 t - The updated error covariance P is obtained. t .

[0139] After completing the state update, the error covariance needs to be recalculated to reflect the uncertainty of the current estimate. The update formula is as follows:

[0140]

[0141] Where: P t The updated error covariance describes the uncertainty of the corrected state. P t - Prior error covariance represents the estimation error during the prediction phase. t Kalman gain controls the effect of observation information on error covariance. zz Observation error covariance measures the uncertainty of subjective ratings.

[0142] The purpose of this step is to adjust the error range so that it gradually converges as the filtering process proceeds, thereby ensuring the stability and reliability of the comfort estimation.

[0143] Comparative Example 1

[0144] To demonstrate the existence of perceptual error in subjective comfort evaluation, this comparative study designed a univariate independent repeated experiment. The core idea of ​​the experiment was to control experimental conditions, using unmanned vehicles on a closed road, to ensure that the same group of subjects could repeatedly evaluate the subject's comfort within the same scenario and time period, thereby observing the distribution of scores. By analyzing the distribution characteristics of the scores, the study verified whether perceptual bias exists in subjective comfort evaluation.

[0145] Experimental Setting and Subjects: The experiment was conducted in a closed environment, consisting of two typical scenarios: crossing speed bumps and making a right turn. Subjects were divided into eight groups based on gender, physical condition, and age: gender (male, female), physical condition (prone to motion sickness, not prone to motion sickness), and age (youth, middle-aged and elderly). Each group contained four subjects, for a total of 32 subjects.

[0146] Experimental Grouping and Procedure: Two participants were selected from each of the eight groups for each experiment. Participants rode in driverless vehicles, following pre-set start and end points. Every 15 seconds, participants rated the comfort of the ride on a scale of 1 to 5. Each experiment was repeated 10 times, with no changes to the participants, their locations, or the setting.

[0147] Data collection: Only the first two minutes of data were collected for each scenario, for a total of 8 time periods (15 seconds each). The data for each group, each scenario, and each time period is equivalent to 4 subjects each scoring 10 times, for a total of 40 independent replicate experiments.

[0148] Total number of experiments: 8 (time period) × 2 (scenario) × 8 (group) × 4 (subjects) × 10 (repetitions) = 5120 scores.

[0149] To demonstrate the existence of perceptual errors in subjective comfort assessment, this study uses heatmaps such as... Figure 6 As shown, this is the main data visualization tool used to analyze the central tendency and dispersion of scores by comparing the score distribution of the same group in multiple experiments under the same scenario. The specific analysis method is as follows:

[0150] Heatmaps can clearly show the frequency distribution of different groups on different ratings. Through heatmaps, we can observe the central tendency of ratings (whether the ratings of the same group in the same scenario are concentrated on a specific rating value) and the degree of dispersion of ratings (whether the distribution range of ratings is wide and whether there is obvious dispersion).

[0151] Taking the scenarios of going over speed bumps and making a right turn as examples, heatmaps were collected and drawn to show the subjective comfort scores of eight groups under specific conditions in the two scenarios, after each group underwent multiple repeated experiments.

[0152] like Figure 7 As shown, the ratings of the same group in the same scenario are not completely consistent, but rather exhibit a certain concentrated distribution; the rating distribution is concentrated around a certain central value, and the frequency gradually decreases as the rating deviates from the central value.

[0153] The ratings of the same group in the same scenario are mainly concentrated in a specific rating value, indicating that the ratings of the subjects under the same conditions have a certain central tendency; the rating distribution shows a certain degree of dispersion, and the frequency gradually decreases as the rating deviates from the central value; the rating distribution characteristics are different in different scenarios.

[0154] The above description is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for accurately evaluating ride comfort based on the fusion of subjective and objective data, characterized in that, Includes the following steps: Step 1: Real-time synchronous collection of passenger comfort evaluation data and vehicle kinematic data. The passenger comfort evaluation data includes subjective scores, and the vehicle kinematic data includes longitudinal acceleration, lateral acceleration, vertical acceleration, longitudinal angular velocity, lateral angular velocity, and vertical angular velocity. Step 2: Preprocess the passenger comfort evaluation data and vehicle kinematics data obtained in Step 1. Delete the passenger comfort evaluation data with a subjective score of 0 and the vehicle kinematics data of the time period corresponding to the timestamp. Perform outlier detection and replacement, low-pass filtering and signal smoothing on the vehicle kinematics data. Step 3: Establish an improved MSDV model as an objective evaluation model: Where MSDV is the output value of the improved MSDV model, a w,i (t) is the frequency-weighted value of acceleration and angular velocity, k i Here are the weighting coefficients for each dimension, and T is the total time length; Step 4: Establish a subjective-objective fusion model based on the combination of unscented Kalman filtering and neural networks: Step 4.0: Determine the initial uncertainty P0 of the system, the system process noise covariance Q(i) as a function of subjective rating levels, and the perceived error covariance R of subjective ratings using statistical methods. t ; Step 4.1, Unscented Kalman Filter (UKF) Prediction Stage: Use the MSDV model output value from the previous time step as the comfort prior estimate to obtain the MSDV prediction value at the current time step; Step 4.2, in the prediction phase, based on the error covariance P of the previous time step. t-1 Calculate the prior error covariance P at the current time. t - ; Step 4.3: Fit the nonlinear observation equation using a multilayer neural network. The mapping relationship between the MSDV prediction value obtained in step 4.1 and the subjective rating obtained in step 1 is studied. Step 4.4, Unscented Kalman Filter (UKF) Observation Update Stage: A trained multilayer neural network is used as the nonlinear observation equation. The calculated values ​​are first input from the output values ​​of the MSDV model after Sigma point transformation. The predicted observation sigma point was calculated. Then calculate the model observation prediction mean. Observation error covariance P zz and the covariance P between the state and the observation xz ; Step 4.5, using the observation error covariance P obtained in step 4.4 zz and the covariance P between the state and the observation xz Calculate the Kalman gain K t ; Step 4.6, based on the Kalman gain K obtained in step 4.5 t The model obtained in step 4.4 is used to predict the observed mean. The deviation between the subjective rating obtained in step 1 and the MSDV prediction obtained in step 4.1 is corrected to obtain the comfort rating after integrating subjective and objective ratings. Step 4.7, based on the Kalman gain K obtained in step 4.5 t Update the prior error covariance P obtained in step 4.2 t - The updated error covariance P is obtained. t .

2. The method for accurately evaluating ride comfort based on the fusion of subjective and objective data as described in claim 1, characterized in that, In step 4.0, the initial uncertainty P0 is calculated using the following formula: n i It is the number of data points with a subjective rating of i, MSDV ij It is the MSDV model output value of the j-th sample with subjective rating i, MSDV i It is the mean of the MSDV model output values ​​for subjective rating i, and M is the total number of all rating levels; System process noise covariance Perceived error covariance of subjective ratings in, It can be calculated from experimental statistical data.

3. The method for accurately evaluating ride comfort based on the fusion of subjective and objective data as described in claim 1, characterized in that, In step 4.1, This is the MSDV prediction value at the current moment. It is the posterior estimate of comfort obtained after the end of the previous moment, U. t It is the input at the current moment. A and B are the state transition matrix and the control matrix, respectively.

4. The method for accurately evaluating ride comfort based on the fusion of subjective and objective data as described in claim 1, characterized in that, In step 4.2, P t - =AP t-1 A T +Q t ; Where: P t - It is the prior error covariance at the current moment, P t-1 It is the posterior error covariance of the previous time step, Q. t Q is the process noise covariance at the current moment; t Based on adaptive adjustment of subjective ratings, Q t =Q(i).

5. The method for accurately evaluating ride comfort based on the fusion of subjective and objective data as described in claim 1, characterized in that, In step 4.3, the nonlinear observation equation in: The subjective rating given by passengers at time t; To obtain the MSDV prediction value in step 4.1; This represents the nonlinear mapping relationship learned by the neural network; v t To observe the noise, assume it follows a zero-mean normal distribution: v t ~N(0,R t ).

6. The method for accurately evaluating ride comfort based on the fusion of subjective and objective data as described in claim 1, characterized in that, In step 4.3, the multilayer neural network includes an input layer, a hidden layer, and an output layer. The input layer includes the output value of the MSDV model, the one-hot encoded values ​​of the passenger's physiological characteristics and physical condition. The hidden layer adopts a three-layer hidden layer structure, all using ReLU as the activation function. The output layer uses the Softmax activation function to calculate the probability of each rating category. Where, p i The predicted probability of the i-th type of rating calculated by the neural network. The unnormalized score of the neural network output; The training of the multilayer neural network is optimized using the cross-entropy loss function: Among them, y i The one-hot encoding of the true rating represents that if the true rating is k, then y k =1, other categories are 0, p i Let C be the predicted probability of the i-th class rating calculated by the neural network, and C be the number of classes.

7. The method for accurately evaluating ride comfort based on the fusion of subjective and objective data as described in claim 1, characterized in that, In step 4.4, the model observes the predicted mean. and observation error covariance P zz The covariance P between state and observation xz Calculated using the following formula: in, W represents the model's observed predicted mean. i (m) Weighted by mean, To predict the observed sigma point; Among them, W i (c) Covariance weights; in, It is the calculated value after the MSDV model output value has undergone Sigma point transformation.

8. The method for accurately evaluating ride comfort based on the fusion of subjective and objective data as described in claim 1, characterized in that, In step 4.5, K t For Kalman gain.

9. The method for accurately evaluating ride comfort based on the fusion of subjective and objective data as described in claim 1, characterized in that, In step 4.6, To achieve a comfort score that integrates subjective and objective ratings, K is the predicted value of MSDV. t The Kalman gain determines the degree to which subjective ratings correct for the state, z. t These are the actual observed values, i.e., the subjective scores obtained in step 1. The deviation term represents the difference between the actual observed value and the model-predicted mean observed value. The deviation between them.

10. The method for accurately evaluating ride comfort based on the fusion of subjective and objective data as described in claim 1, characterized in that, In step 4.7, Where: P t For the updated error covariance, P t - For the prior error covariance, K t For Kalman gain, P zz This represents the observation error covariance.

Images