Model for quantifying the relationship between vehicle handling behavior and passenger ride experience

Abstract

The application provides a model for quantifying the relationship between vehicle longitudinal and lateral operation behavior and passenger driving experience, comprising: a psychophysics mechanism submodel: establishing a driving comfort quantification submodel, extracting a feature vector from input parameters based on quantification psychology principles; a data-driven submodel: quantitatively analyzing the relationship between vehicle dynamics sequences generated by different vehicle operation behaviors and passenger subjective feelings; an integrated learning model: taking the output of the psychophysics mechanism submodel and the output of the data-driven submodel as inputs, fusing the outputs of the two submodels through a full connection network to obtain the output result of the whole model, which is an index of driving comfort corresponding to the current input vehicle dynamics sequence. The application solves the problem of difficulty in feature coding and gathering of long vehicle dynamics parameter sequences in a unit traffic condition, and improves the modeling capability of the driving comfort model for the dependence relationship in long sequence input.

Description

Technical Field

[0001] This invention relates to the field of learning algorithm technology, and more specifically, to a model for quantifying the relationship between vehicle lateral and longitudinal handling behavior and passenger driving experience. Background Technology

[0002] In the field of psychology, psychologist Stevens proposed a theory called the "Stevens power law" to quantify the relationship between people's subjective feelings and external stimulus input. Ao applied Stevens' power law to explore the correlation between passengers' subjective feelings and objective assessments. Furthermore, the law has been used to predict discomfort as a complementary method to quantifying passenger comfort indices. As a mechanism-based methodology, psychophysical methods possess strong interpretability and stability. However, in complex traffic scenarios, psychophysical methods often exhibit limited expressive power.

[0003] With sufficiently high-quality datasets, data-driven approaches can develop powerful machine learning models that quantify driving comfort indices caused by vehicle handling behavior. Luntian Mou introduced an attention-based convolutional neural network model called the Long Short-Term Memory (CNN-LSTM) model to assess driver stress as a measure of comfort. However, learning-based models typically operate as black boxes, exhibiting low interpretability.

[0004] Patent document CN112109735A (CN202011141795.X) discloses an intelligent driving control system and method that considers vehicle comfort, including: a manual start unit for the driver to manually start the autonomous driving state and select the corresponding driving mode; a driving mode control module, including an anti-motion sickness mode control module; a normal mode control module; and a destination approaching mode control module; an autonomous driving control unit for autonomous driving control in various modes; a data processing center for processing and analyzing the onboard computer's actions based on the driver's driving behavior and vehicle information during manual driving to help select the appropriate driving mode in autonomous driving mode; and a vehicle information monitoring unit. However, this patent cannot solve the aforementioned technical problems. Summary of the Invention

[0005] In view of the deficiencies in the prior art, the purpose of this invention is to provide a model for quantifying the relationship between vehicle lateral and longitudinal handling behavior and passenger driving experience.

[0006] The model provided by the present invention for quantifying the relationship between vehicle lateral and longitudinal handling behavior and passenger driving experience includes:

[0007] Psychophysical Mechanism Sub-model: Based on Steven's power function law in psychophysics, a quantitative sub-model of driving comfort is established. Feature vectors are extracted from the input parameters to obtain the cumulative stimulus CS, the rate of change of stimulus within a unit traffic condition SR, and the instantaneous maximum stimulus IMS within a unit traffic condition.

[0008] Data-driven sub-model: Using the Encoder-Transformer module, a masked soft attention mechanism, and a fully connected layer, we quantitatively analyze the relationship between vehicle dynamics sequences generated by different vehicle operating behaviors and passengers' subjective feelings;

[0009] The ensemble learning model takes the outputs of the psychophysical mechanism sub-model and the data-driven sub-model as inputs, and fuses the outputs of the two sub-models through a fully connected network to obtain the output of the entire model. This output is an index of the driving comfort corresponding to the current input vehicle dynamics sequence.

[0010] Preferably, the vehicle dynamics characteristics related to vehicle behavior are defined as external stimuli applied to passengers, and the subjective evaluation score is a quantification of psychological changes, expressed as:

[0011] S = b * I α

[0012] Where S represents psychological quantity, I represents physical quantity, α represents power exponent, and b represents scaling factor.

[0013] The lateral and longitudinal dynamic sequences of a vehicle are defined as external stimuli, and are expressed as follows:

[0014]

[0015] Where ΔT represents the duration of the stimulus, T represents the starting point of the stimulus, a(t) represents the vehicle dynamics sequence, and f represents the feature projection relationship.

[0016] Preferably, a fixed-length time window is used to calculate CS, and the starting point of the selected time window in the vehicle dynamics sequence is determined by the Cross Time Window (CTW) method. For a single dimension VD of the vehicle dynamics parameter sequence k (t), the sequence is divided into equal lengths ΔT cs Δt is the step size of the CTW method. In the nth step, the starting point of the first segment moves backward by n*Δt, where ΔT is the step size. cs In the following formula, is an integer multiple of Δt. Therefore, the CTW of the current iteration will end when the starting point of the first segment in the nth step matches the ending point of the first segment in the first step.

[0017]

[0018] Where n represents the number of steps in CTW, ΔT cs The time window width of CTW is a hyperparameter, and Δt represents the step size of CTW.

[0019] In the i-th step of CTW, the cumulative stimulus value of the k-th dimension vehicle dynamics parameter is calculated. It is obtained by calculating each subsequence;

[0020]

[0021] Among them, VD k (t) is the k-th dimension parameter of the vehicle dynamics parameter sequence at time (t), μ k It's VD k The average value of (t), σ k It's VD k The variance of (t), where k is the dimension index and t is the time point. It is a power exponent, a learnable parameter, where i is the index of the traffic scene in the dataset. It is the starting point of the k-th dimension parameter of the selected i-th traffic scenario;

[0022] In the calculation of the CS term, the selected segment is the largest term in the following equation:

[0023]

[0024] Among them, CS k This represents the CS value of the k-th dimension of the vehicle dynamics sequence.

[0025] Preferably, ΔT si Defined as VD k The expression for the duration between the maximum and minimum values ​​of (t) is:

[0026]

[0027] The rate of change of stimulus (SR) is defined as the maximum average slope of a sequence of vehicle dynamic parameters under a traffic condition, and its expression is:

[0028]

[0029] Among them, SR k The SR value of the k-th dimension of the vehicle dynamic sequence, the power exponent. These are learnable parameters in PCQM.

[0030] Preferably, the instantaneous maximum stimulus (IMS) is defined as the maximum value of the vehicle dynamics parameter sequence in all dimensions throughout the entire traffic condition, expressed as:

[0031]

[0032] Among them, IMS k The power exponent represents the IMS value of the k-th dimension of the vehicle dynamic sequence. Learnable parameters.

[0033] Preferably, the three stimulus quantification terms are combined into a vector through weighted summation. Specifically as follows:

[0034]

[0035] in, These are the weights of the learnable items. It is the subjective psychological quantification result of the external stimulus in the k-th dimension of the vehicle dynamics sequence;

[0036] Under a single traffic condition, the stimulus quantification results generated by the 11-dimensional vehicle dynamics parameter sequence are as follows:

[0037]

[0038] in, G is the stimulus quantization value of a single-dimensional sequence of vehicle dynamics parameters, and G is a learnable parametric linear transformation used to enhance the expressive power of the model.

[0039] Preferably, the Encoder-Transformer module maps the vehicle dynamics sequence from the input matrix to a higher-dimensional feature matrix. Then, the masked soft attention layer, consisting of a masked softmax operation and a soft attention convergence layer, converges the high-dimensional feature matrix into a feature vector. This feature vector is finally mapped to a specific real value through a fully connected layer, which serves as the input to the subsequent ensemble learning layer.

[0040] The Encoder-Transformer module consists of N identical layers stacked together. Each layer contains two sub-layers: the first sub-layer is a multi-head self-attention mechanism layer, and the second sub-layer is a position-based feedforward neural network (FFN). Residual connections are added between each layer to enhance the model's feature extraction capabilities. Then, layer normalization is performed. The attention function maps the query and a set of key-value pairs to an output, where the query, key, value, and output are all vectors. The output is a weighted sum of values, where the weight of a given value is calculated by the attention scoring function between the query and the corresponding key.

[0041] Using scaled dot product attention, a set of queries is packaged into a matrix Q, and the keys and values ​​are packaged into matrices K and V. The output matrix is ​​calculated as follows:

[0042]

[0043] Where, d k It is the dimension of Q, K T It is the transpose of K;

[0044] In multi-head attention mechanisms, the attention calculation for a single head is as follows:

[0045] H j =Attention(W j Q ·Q,W j K ·K,W j v ·V)

[0046] Among them, H j It is the output of the i-th attention head, {W j Q W j K W j V} is a learnable weight matrix;

[0047] The outputs of different attention mechanism heads are concatenated along a specified dimension and then passed through another linear layer W. o By projecting, we obtain the final output, expressed as:

[0048] MH(Q,K,V)=Con(H1,H2,…,H j )·W o

[0049] Where MH is the output of the multi-head attention layer, j is the head index of the attention mechanism, Con is an abbreviation for connection, and W... o It is a learnable weight matrix.

[0050] Preferably, the output MH(Q,K,V) of the Encoder-Transformer module is a sequence {O(0),O(1),…,O(T)}, with the same length as the input sequence. The masked soft attention layer maps the sequence {O(0),O(1),…,O(T)} to a one-dimensional vector M, expressed as:

[0051]

[0052] Where M is the output of the masked soft attention mechanism module, α it The normalized weights are calculated in the following formula after occlusion Softmax preprocessing:

[0053] α it =MS(U T *Uω )

[0054] Where MS is the occluded Softmax operation, and U... ω U is a random initialization vector with learnable parameters. T It is the transpose of U, which is calculated by a nonlinear transformation in the following formula:

[0055] U = tanh(W) ω *{O(0),O(1),…,O(T)}+b ω )

[0056] Among them, W ω It is a learnable weight matrix.

[0057] Preferably, the vector M obtained after attention convergence is fed into the subsequent fully connected layer to produce the final output of the data-driven sub-model. As input parameters for subsequent ensemble learning modules;

[0058] The input to the data-driven sub-model is the vehicle dynamics parameter sequence VD, which is an eleven-dimensional two-dimensional matrix. For the passenger driving experience quantification model, these vectors VD in the sequence are regarded as queries, keys, and values.

[0059] VD≡Q≡K≡V.

[0060] Preferably, the meta-model used is a fully connected network, corresponding to the stacked fusion layer in the network structure diagram, and the calculated basic model outputs are as follows: and Then, the fusion is performed, expressed as follows:

[0061]

[0062] PCS is a real number obtained by combining two sub-models. The value of this real number is a number between 0 and 5, representing the output index of the quantitative model of driving experience under input traffic conditions. The closer the value is to 5, the better the driving experience; the closer it is to 1, the worse the driving experience.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] (1) By adopting psychophysical methods, a quantitative relationship between vehicle handling behavior and driving comfort was modeled from the perspective of psychological mechanisms, which solved the problem of poor model interpretability in the process of quantifying driving comfort and improved the interpretability of the model.

[0065] (2) By adopting the data-driven method based on Encoder-Transformer, the problem of feature encoding and convergence of long vehicle dynamic parameter sequences within a unit traffic condition is solved, and the ability of the ride comfort model to model the dependencies between long input sequences is improved.

[0066] (3) By adopting an integrated learning strategy, the interpretability of psychological models and the powerful expressive ability of data-driven models are combined, thereby realizing a quantitative model of driving comfort that has both powerful feature expression ability and interpretability. Attached Figure Description

[0067] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0068] Figure 1 This is a schematic diagram of the model structure used in this invention to quantify the relationship between vehicle lateral and longitudinal handling behavior and ride comfort.

[0069] Figure 2 A schematic diagram illustrating the determination of the starting point of the selected time window in the vehicle dynamics sequence using the cross-time window method. Detailed Implementation

[0070] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0071] Example

[0072] like Figure 1 This invention provides a model for quantifying the relationship between vehicle lateral and longitudinal handling behavior and passenger driving experience. The input parameters of the model are a sequence of 11 dynamic parameters measured by in-vehicle integrated navigation, including: roll angle, roll rate, yaw rate, yaw acceleration, lateral jerk, lateral acceleration, lateral velocity, longitudinal jerk, longitudinal acceleration, longitudinal velocity, and the reciprocal of the turning radius.

[0073] A sub-model for quantifying driving comfort is established based on Steven's power function law in psychophysics. The sub-model includes a feature extraction module, which extracts feature vectors from the input: the cumulative stimulus (CS) per unit traffic condition, the rate of change of stimulus (SR) per unit traffic condition, and the instantaneous maximum stimulus (IMS) per unit traffic condition. The output of the feature extraction module is a mechanistic feature vector per unit traffic condition. This vector is passed through a fully connected network to generate the output value of the current sub-model.

[0074] In psychophysics, Stevens' power law is a commonly used mathematical model. Stevens used quantitative assessment methods to study the relationship between stimuli and sensations, and proposed Stevens' power law, which became the foundation of modern psychophysics. This model shows that there is a non-linear functional relationship between the intensity of a stimulus and subjective perception; that is, the perceived experience of a stimulus is not simply proportional to its intensity. This model provides researchers with a method to quantify and predict the impact of stimuli on subjective perception, further deepening our understanding of psychophysics. In this invention, vehicle dynamics characteristics related to vehicle behavior are defined as external stimuli applied to passengers, while subjective evaluation scores are a quantification of psychological changes, expressed as:

[0075] S = b * I α

[0076] Where S represents the psychological quantity, I represents the physical quantity, α represents the power exponent, and b represents the scaling factor.

[0077] In this invention, the lateral and longitudinal dynamic sequences of a vehicle are defined as external stimuli. Stimuli are inputs that cause subjective psychological fluctuations in volunteers. For the passenger driving experience quantification model, the mathematical expression of external stimuli is as follows:

[0078]

[0079] Where ΔT represents the duration of the stimulus, T represents the starting point of the stimulus, a(t) represents the vehicle dynamics sequence, and f represents the feature projection relationship. The input to the driving experience quantification model consists of an 11-dimensional sequence of vehicle dynamics parameters. Therefore, CS, SR, and IMS are calculated for each dimension. CS represents the stimulus intensity within a given time interval. SR captures the rate of stimulus change throughout the traffic conditions. Normalization is used when calculating CS to eliminate the numerical stability problem caused by the different value ranges of different dimensions of vehicle dynamics features. The purpose of introducing IMS is to quantify the highest stimulus level in each dynamics input dimension.

[0080] Cumulative Stimulus (CS): The CS is calculated by accumulating the positive or negative trends of the vehicle dynamics sequence within a single dimension under a traffic condition. Considering the varying durations of different traffic scenarios, a fixed-length time window is used to calculate the CS during modeling. This invention proposes the Cross-Time Window (CTW) method to determine the starting point of the selected time window within the vehicle dynamics sequence. like Figure 2 As shown. For a single dimension VD of the vehicle dynamics parameter sequence. k(t), such as the yaw rate, the sequence is divided into equal lengths ΔT. cs Δt is the step size of the CTW method. In the nth step, the starting point of the first segment moves backward by n*Δt. ΔT cs In the following formula, is an integer multiple of Δt. Therefore, the CTW of the current iteration will end when the starting point of the first segment in the nth step matches the ending point of the first segment in the first step.

[0081]

[0082] Where n represents the number of steps in CTW, ΔT cs The time window width of CTW is a hyperparameter, while Δt represents the step size of CTW.

[0083] In the i-th step of CTW, the cumulative stimulus value of the k-th dimension vehicle dynamics parameter is calculated. It is obtained by calculating each subsequence.

[0084]

[0085] Among them, VD k (t) is the k-th dimension parameter of the vehicle dynamics parameter sequence at time (t), μ k It's VD k The average value of (t), σ k It's VD k The variance of (t), where k is the dimension index and t is the time point. It is a power exponent, a learnable parameter, where i is the index of the traffic scene in the dataset. It is the starting point of the k-th dimension parameter of the selected i-th traffic scenario.

[0086] In the calculation of the CS term, the selected segment is the largest term in the following equation:

[0087]

[0088] Among them, CS k This represents the CS value of the k-th dimension of the vehicle dynamics sequence.

[0089] Stimulus rate of change part: ΔT si Defined as VD k The expression for the duration between the maximum and minimum values ​​of (t) is:

[0090]

[0091] The rate of change of stimulus (SR) is defined as the maximum average slope of a sequence of vehicle dynamic parameters under a traffic condition.

[0092]

[0093] Among them, SR k The SR value of the k-th dimension of the vehicle dynamic sequence, the power exponent. These are learnable parameters in PCQM.

[0094] Instantaneous maximum stimulus (IMS): IMS is defined as the maximum value of the sequence of vehicle dynamic parameters in all dimensions throughout the entire traffic condition.

[0095]

[0096] Among them, IMS k The power exponent represents the IMS value of the k-th dimension of the vehicle dynamic sequence. Learnable parameters.

[0097] Mechanistic model feature combination: These three stimulus quantification terms are combined in a weighted manner to form a vector. Specifically as follows:

[0098]

[0099] in These are the weights of the learnable items. It is the subjective psychological quantification result of the external stimulus in the k-th dimension of the vehicle dynamics sequence.

[0100] Under a single traffic condition, the stimulus quantification results generated by the 11-dimensional vehicle dynamics parameter sequence are as follows:

[0101]

[0102] in, G is the stimulus quantization value of a single-dimensional sequence of vehicle dynamics parameters, and G is a learnable parametric linear transformation used to enhance the expressive power of the model.

[0103] The psychophysical-based quantification sub-model of passenger driving experience is built by referencing Stevens' power law and incorporating statistical information from time series. Learnable parameters include:

[0104]

[0105] Data-Driven Sub-Model: In the modeling of the passenger driving experience quantification model, the driving comfort quantification model aims to quantitatively analyze the relationship between the vehicle dynamics sequence generated by different vehicle operating behaviors and passenger subjective feelings. In addition to the CS, IMS, and SR indices included in the sub-model based on psychophysical mechanisms, passenger subjective feelings are also influenced by the changing trend of the vehicle dynamics sequence along the time axis. Considering the high-dimensional nonlinear relationship between the vehicle dynamics sequence and passenger subjective feelings, as well as the continuity of driving behavior, this invention proposes a data-driven sub-model based on an Encoder-Transformer as the backbone network to extract valuable features from the vehicle dynamics sequence. The architecture of the data-driven sub-model is as follows: Figure 1 As shown on the right, it includes the Encoder-Transformer module, the soft attention mechanism module with masking, and the fully connected layer.

[0106] The backbone network is based on Encoder-Transformer: The Encoder-Transformer module in the model maps the vehicle dynamics sequence from the input matrix to a higher-dimensional feature matrix. Then, a masked soft attention layer, consisting of a masked softmax operation and a soft attention pooling layer, pools the high-dimensional feature matrix into a single feature vector. This feature vector is then mapped to a specific real value through a fully connected layer, serving as the input to subsequent ensemble learning layers.

[0107] The Encoder-Transformer module consists of N identical layers stacked together, such as Figure 1 As shown, each layer contains two sub-layers. The first layer is a multi-head self-attention mechanism layer, and the second layer is a position-based feedforward neural network (FFN). Residual connections are added between each layer to enhance the model's feature extraction capabilities, followed by layer normalization. The basic idea of ​​the attention mechanism is that humans selectively focus on target information and attenuate other less important information through their field of vision. The attention function maps a query and a set of key-value pairs to an output, where the query, key, value, and output are all vectors. The output is a weighted sum of values, where the weight of a given value is calculated by the attention scoring function of the query and the corresponding key.

[0108] In the Encoder-Transformer-based data-driven sub-model, a self-attention mechanism is used as the modeling method for the attention mechanism. Queries, keys, and values ​​all come from the same sequence. The attention function used in this invention is scaled dot-product attention, which packages a set of queries into a matrix Q. Keys and values ​​are packaged into matrices K and V. The output matrix is ​​calculated as follows:

[0109]

[0110] Where, d k It is the dimension of Q, K T It is the transpose of K.

[0111] In self-attention mechanisms, multi-head attention, compared to single-head attention, transforms queries, keys, and values ​​into different representation subspaces through independent linear projections and learnable weight matrices, rather than performing only a single attention function operation. Multi-head attention allows the model to simultaneously focus on information from different representation subspaces at different locations. Through multi-head attention, different heads capture useful information from different parts of the input sequence, thereby improving the model's capabilities in prediction, generation, and other tasks.

[0112] In multi-head attention mechanisms, the attention calculation for a single head is as follows:

[0113] H j =Attention(W j Q ·Q,W j K ·K,W j v ·V)

[0114] Among them, H j It is the output of the i-th attention head, {W j Q W j K W j V} is a learnable weight matrix.

[0115] The outputs of different attention mechanism heads are concatenated along a specified dimension and then passed through another linear layer W. o Project the image to obtain the final output.

[0116] MH(Q,K,V)=Con(H1,H2,…,H j )·W o

[0117] Where MH is the output of the multi-head attention layer, j is the head index of the attention mechanism, Con is an abbreviation for connection, and W... o It is a learnable weight matrix.

[0118] Masked Softmax and Soft Attention Mechanism: The output MH(Q,K,V) of the Encoder-Transformer module is still a sequence {O(0),O(1),…,O(T)}, with the same length as the input sequence. The masked soft attention layer maps the sequence {O(0),O(1),…,O(T)} to a one-dimensional vector M.

[0119]

[0120] Where M is the output of the masked soft attention mechanism module, α it These are the normalized weights calculated in the following formula after occlusion Softmax preprocessing.

[0121] α it =MS(U T *U ω )

[0122] Where MS is the occluded Softmax operation, and U... ω U is a random initialization vector with learnable parameters. T It is the transpose of U.

[0123] U is calculated using a nonlinear transformation in the following formula.

[0124] U = tanh(W) ω *{O(0),O(1),…,O(T)}+b ω )

[0125] Among them, W ω It is a learnable weight matrix.

[0126] The vector M obtained after attention convergence is fed into the subsequent fully connected layer to produce the final output of the data-driven sub-model. As input parameters for subsequent ensemble learning modules.

[0127] In this invention, the input to the data-driven sub-model is a sequence of vehicle dynamics parameters, VD, which is an eleven-dimensional two-dimensional matrix. For the passenger driving experience quantification model, these vectors VD in the sequence are treated as queries, keys, and values.

[0128] VD≡Q≡K≡V

[0129] The following two formulas provide the mathematical implementation of self-attention and multi-head attention after incorporating the vehicle dynamics parameter sequence into the data-driven sub-model.

[0130]

[0131] H j=Attention(W j Q ·VD,W j K ·VD,W j V ·VD)

[0132] MH(VD,VD,VD)=Con(H1,H2,…,H j )·W o

[0133] Ensemble learning model: An ensemble learning model that integrates two sub-models: The ensemble learning model takes the output of the psychological mechanism model and the output of the data-driven sub-model as input, and integrates the outputs of the two sub-models through a fully connected network to produce the output of the entire model. This output is an index of driving comfort corresponding to the current input vehicle dynamics sequence.

[0134] Stacked learning, a type of ensemble learning, combines multiple independent machine learning models for prediction. Introducing ensemble learning into a passenger driving experience quantification model offers several advantages: First, it enables the model to simultaneously extract psychophysical mechanistic features and temporal features, effectively enhancing its expressive power. Second, by fusing multiple sub-models, stacked learning helps reduce model bias and improves prediction accuracy. Unlike other ensemble techniques, stacked learning uses the predictions from multiple models as input features to train a meta-model before making the final prediction. Stacked learning provides a powerful approach to improving prediction accuracy, leveraging model diversity, mitigating overfitting, and offering flexibility and interpretability during model development.

[0135] Stacked learning consists of two training phases: Phase 0 and Phase 1, as shown in Algorithm ~\ref{alg:alg1}. In Phase 0, the base model is trained on a driving experience quantification dataset to generate the input for the Layer 1 meta-model. The Layer 0 model includes a psychophysical mechanism-based sub-model and an Encoder-Transformer data-driven sub-model. In Phase 1, a linear regression model serves as the meta-model, used to combine the outputs of the Layer 0 model. These two phases begin training sequentially, and data is automatically passed from the Layer 0 phase to the Layer 1 phase.

[0136] This invention establishes a psychophysical mechanism sub-model based on Stevens' power function law and a data-driven sub-model based on Encoder-Transformer. These two sub-models are the basic models that generate the meta-model input. The calculated outputs of the basic models are as follows: and The meta-model used in this invention is a fully connected network, corresponding to the stacked fusion layer in the network structure diagram.

[0137]

[0138] The final PCS is a real number obtained by combining these two sub-models. This real number ranges from 0 to 5, representing the output index of the quantitative model of driving experience under input traffic conditions. A value closer to 5 indicates a better driving experience, while a value closer to 1 indicates a worse driving experience.

[0139] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.

[0140] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for quantifying the relationship between vehicle lateral and longitudinal handling behavior and passenger driving experience, characterized in that, include: Psychophysical Mechanism Sub-model: Based on Steven's power function law in psychophysics, a quantitative sub-model of driving comfort is established. Feature vectors are extracted from the input parameters to obtain the cumulative stimulus CS, the rate of change of stimulus within a unit traffic condition SR, and the instantaneous maximum stimulus IMS within a unit traffic condition. Used to indicate the amount of stimulus within a given time interval; This captures the rate of change of stimuli throughout the entire traffic condition; IMS is defined as the maximum value of the sequence of vehicle dynamic parameters in all dimensions throughout the entire traffic condition; Data-driven sub-model: Using the Encoder-Transformer module, a masked soft attention mechanism, and a fully connected layer, we quantitatively analyze the relationship between vehicle dynamics sequences generated by different vehicle operating behaviors and passengers' subjective feelings; Integrated learning model: The output of the psychophysical mechanism sub-model and the output of the data-driven sub-model are used as inputs. The outputs of the two sub-models are fused through a fully connected network to obtain the output of the entire model. This output is an index of the driving comfort corresponding to the current input vehicle dynamics sequence. The Encoder-Transformer module maps the vehicle dynamics sequence from the input matrix to a higher-dimensional feature matrix. Then, a masked soft attention layer, consisting of a masked softmax operation and a soft attention pooling layer, pools the high-dimensional feature matrix into a feature vector. This feature vector is then mapped to a specific real value through a fully connected layer, which serves as the input to the subsequent ensemble learning layer. The Encoder-Transformer module consists of N identical layers stacked together. Each layer contains two sub-layers: the first sub-layer is a multi-head self-attention mechanism layer, and the second sub-layer is a position-based feedforward neural network (FFN). Residual connections are added between each layer to enhance the model's feature extraction capabilities. Then, layer normalization is performed. The attention function maps the query and a set of key-value pairs to an output, where the query, key, value, and output are all vectors. The output is a weighted sum of values, where the weight of a given value is calculated by the attention scoring function between the query and the corresponding key. Scaled dot product attention is used to package a set of queries into a matrix. Pack the keys and values ​​into a matrix and The output matrix is ​​calculated as shown in the following formula: in, yes Dimensions yes transpose; In multi-head attention mechanisms, the attention calculation for a single head is as follows: in, It is the first The output of each attention head, It is a learnable weight matrix; The outputs of different attention mechanism heads are concatenated along a specified dimension and then passed through another linear layer. By projecting, we obtain the final output, expressed as: in, It is the output of the multi-head attention layer. It is the head index of the attention mechanism. It is an abbreviation for connection. It is a learnable weight matrix.

2. The method for quantifying the relationship between vehicle lateral and longitudinal handling behavior and passenger driving experience according to claim 1, characterized in that, Vehicle dynamics characteristics related to vehicle behavior are defined as external stimuli applied to passengers, and subjective evaluation scores are the quantification of psychological changes, expressed as: in, Represents psychological quantity. Represents a physical quantity. Represents the power index. Represents the scaling factor; The lateral and longitudinal dynamic sequences of a vehicle are defined as external stimuli, and are expressed as follows: in, Indicates the duration of the stimulus. Indicates the starting point of the stimulus. Represents the vehicle dynamics sequence. This represents the feature projection relationship.

3. The method for quantifying the relationship between vehicle lateral and longitudinal handling behavior and passenger driving experience according to claim 2, characterized in that, Calculation using a fixed-length time window The starting point of the selected time window in the vehicle dynamics sequence is determined by the Cross Time Window (CTW) method. For a single dimension of the vehicle dynamics parameter sequence The sequence is divided into equal lengths. , It is the step size of the CTW method, at the th... During the step, the starting point of the first segment moves backward. , In the following formula is Multiples of , therefore, when the The CTW of the current iteration will end when the starting point of the first segment in the first step matches the ending point of the first segment in the first step. in, Indicates the number of steps in CTW. The time window width of CTW is a hyperparameter. Indicates the step size of CTW; In CTW's In the first step, the calculation of the first step... Cumulative stimulus values ​​of vehicle dynamic parameters It is obtained by calculating each subsequence; in, It is a moment The first of the vehicle dynamics parameter sequence Dimensional parameters, yes The average value, yes variance It is a dimensional index. It's a point in time. It is the power exponent, a learnable parameter. It is an index of traffic scenes in the dataset. The first choice The first traffic scenario The starting point of the dimensional parameter; exist In the calculation of the terms, the selected segment is the largest term in the following equation: in, Represents the vehicle dynamics sequence number 1 Vie value.

4. The method for quantifying the relationship between vehicle lateral and longitudinal handling behavior and passenger driving experience according to claim 3, characterized in that, Defined as The duration between the maximum and minimum values ​​is expressed as: Stimulus rate of change Defined as the maximum average slope of a sequence of vehicle dynamic parameters under a traffic condition, the expression is: in, The first vehicle in the dynamic sequence Dimensional Value, power exponent These are learnable parameters in PCQM.

5. The method for quantifying the relationship between vehicle lateral and longitudinal handling behavior and passenger driving experience according to claim 4, characterized in that, Instantaneous maximum stimulus Defined as the maximum value of the vehicle dynamics parameter sequence across all dimensions within the entire traffic condition, expressed as: in, Indicates the vehicle dynamic sequence number Dimensional Value, power exponent Learnable parameters.

6. The method for quantifying the relationship between vehicle lateral and longitudinal handling behavior and passenger driving experience according to claim 5, characterized in that, The three stimulus quantification terms are combined into a vector through a weighted sum. The details are as follows: in, These are the weights of the learnable items. It is the vehicle dynamics sequence number The subjective psychological quantification results of external stimuli in each dimension; Under a single traffic condition, the stimulus quantification results generated by the 11-dimensional vehicle dynamics parameter sequence are as follows: in, It is a stimulus quantization value of a single-dimensional sequence of vehicle dynamics parameters. It is a learnable parameterized linear transformation used to enhance the expressive power of the model.

7. The method for quantifying the relationship between vehicle lateral and longitudinal handling behavior and passenger driving experience according to claim 1, characterized in that, Output of the Encoder-Transformer module It is a sequence The output length is equal to the input sequence, and the masked soft attention layer will... Mapped to a one-dimensional vector The expression is: in, It is the output of the module with masking soft attention mechanism. The normalized weights are calculated in the following formula after occlusion Softmax preprocessing: in, It is a Softmax operation with occlusion. It is a random initialization vector with learnable parameters. yes transpose, The following formula is obtained through nonlinear transformation: in, It is a learnable weight matrix.

8. The method for quantifying the relationship between vehicle lateral and longitudinal handling behavior and passenger driving experience as described in claim 7, characterized in that, The vector obtained after attention convergence It is fed into subsequent fully connected layers to produce the final output of the data-driven sub-model. , which serve as input parameters for subsequent integrated learning modules; The input to the data-driven sub-model is a sequence of vehicle dynamics parameters. The sequence is an eleven-dimensional two-dimensional matrix. For a passenger driving experience quantification model, these vectors in the sequence... Considered as queries, keys, and values; 。 9. The method for quantifying the relationship between vehicle lateral and longitudinal handling behavior and passenger driving experience according to claim 8, characterized in that, The meta-model used is a fully connected network, corresponding to the stacked fusion layer in the network structure diagram. The basic model outputs calculated are as follows: and Then, the fusion is performed, and the expression is: PCS is a real number obtained by combining two sub-models. The value of this real number is a number between 0 and 5, representing the output index of the quantitative model of driving experience under input traffic conditions. The closer the value is to 5, the better the driving experience; the closer it is to 1, the worse the driving experience.

Images