Seat adjusting method and device, vehicle and storage medium

Abstract

The invention discloses a seat adjustment method and device, a vehicle and a storage medium, and relates to the technical field of seat adjustment. The method comprises the following steps: acquiring sitting and riding data of a passenger, and inputting the sitting and riding data into a preset comfort level evaluation model to output a comfort level vector; under the condition that a passenger feedback instruction is received within preset time, inputting the comfort level vector and the passenger feedback instruction into a preset seat adjusting parameter prediction model to output a target seat adjusting parameter; and determining a to-be-adjusted part of the seat according to the target seat adjustment parameter, and adjusting the to-be-adjusted part based on the target seat adjustment parameter. Thus, the defect that only a fixed mode or preset gear adjustment can be achieved during seat adjustment is overcome, intelligence, real-time performance and individuation of seat adjustment are achieved, the higher requirement for the comfort of the automobile seat under modern automobile development is met, and the experience and quality of the automobile seat are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of seat adjustment technology, and in particular to a seat adjustment method, a seat adjustment device, a computer-readable storage medium, and a vehicle. Background Technology

[0002] With the development of the automotive industry, seat comfort has become one of the core indicators for measuring vehicle quality. Modern seat design needs to integrate ergonomic support, dynamic damping systems, and intelligent adjustment functions.

[0003] The automotive seat adjustment systems in related technologies mainly rely on traditional motor-driven structures, achieving basic actions such as seat tilting, fore-aft adjustment, and height adjustment through manual adjustment or predefined gears. While this design meets basic usage needs to a certain extent, it has significant limitations in providing a personalized comfort experience. Typically based on fixed parameter settings, it lacks real-time collection and analysis of multi-dimensional information such as occupant body shape, posture, body position, and dynamic comfort, making it difficult to make fine-tuned adjustments for the actual needs of different occupants. Especially in complex environments such as long-distance driving, highway cruising, or traffic congestion, occupant body posture, local pressure distribution, and physiological parameters may change significantly, and traditional adjustment methods often cannot respond to these changes in a timely manner, resulting in delayed seat adjustment effects and an inability to provide continuous and precise comfort adjustments. Summary of the Invention

[0004] This application aims to at least partially address one of the technical problems in the related art. To this end, the first objective of this application is to propose a seat adjustment method, the method comprising: collecting occupant seating data and inputting the seating data into a preset comfort evaluation model to output a comfort vector; upon receiving occupant feedback instructions within a preset time, inputting the comfort vector and the occupant feedback instructions into a preset seat adjustment parameter prediction model to output target seat adjustment parameters; determining the adjustable components of the seat based on the target seat adjustment parameters, and adjusting the adjustable components based on the target seat adjustment parameters. This application continuously collects passenger seating data and inputs it into a preset comfort assessment model to generate a multi-dimensional comfort vector, thereby achieving real-time quantitative perception of passenger comfort status. When a passenger's active feedback instruction is received, the feedback instruction and the real-time comfort vector are input into a prediction model, outputting adjustment parameters that integrate objective state and subjective intention. Finally, based on these adjustment parameters, the components to be adjusted are precisely driven to operate collaboratively, overcoming the shortcomings of seat adjustment which can only achieve fixed modes or preset gear adjustments. This realizes intelligent, real-time, and personalized seat adjustment, meeting the higher requirements for car seat comfort under the development of modern automobiles, and significantly improving the experience and quality of car seats.

[0005] The second objective of this application is to provide a seat adjustment device.

[0006] The third objective of this application is to provide a computer-readable storage medium.

[0007] The fourth objective of this application is to propose a vehicle.

[0008] To achieve the above objectives, the first aspect of this application proposes a seat adjustment method, which includes: collecting occupant seating data and inputting the seating data into a preset comfort evaluation model to output a comfort vector; upon receiving an occupant feedback instruction within a preset time, inputting the comfort vector and the occupant feedback instruction into a preset seat adjustment parameter prediction model to output target seat adjustment parameters; determining the adjustable components of the seat based on the target seat adjustment parameters, and adjusting the adjustable components based on the target seat adjustment parameters.

[0009] According to one embodiment of this application, before inputting the comfort vector and occupant feedback instructions into the preset seat adjustment parameter prediction model, the method further includes: structuring the occupant feedback instructions to generate an occupant preference feature vector; fusing the comfort vector and the occupant preference feature vector to generate a fused feature vector; and inputting the fused feature vector into the preset seat adjustment parameter prediction model to output the target seat adjustment parameters.

[0010] According to one embodiment of this application, the method further includes: acquiring historical seating data streams and historical feedback instructions and historical seat adjustment parameters corresponding to each historical seating data slice of the historical seating data streams; inputting each historical seating data slice of the historical seating data streams into a preset comfort evaluation model to obtain a corresponding historical comfort vector; performing structured processing on the corresponding historical feedback instructions to obtain an occupant preference feature vector; performing feature fusion between the occupant preference feature vector and the corresponding historical comfort vector to generate a first input feature; constructing a model training set for a preset seat adjustment parameter prediction model based on multiple first input features and corresponding historical seat adjustment parameters; inputting any first input feature from the model training set of the preset seat adjustment parameter prediction model into an initial seat adjustment parameter prediction model to obtain predicted seat adjustment parameters; calculating a discrimination loss based on the predicted seat adjustment parameters and the corresponding seat adjustment parameters in the model training set of the preset seat adjustment parameter prediction model to obtain a first calculation result; updating the model parameters of the initial seat adjustment parameter prediction model using the first calculation result until the updated initial seat adjustment parameter prediction model satisfies a first preset convergence condition to obtain the preset seat adjustment parameter prediction model.

[0011] According to one embodiment of this application, the seating data includes body posture image data, vital sign data, and vehicle driving data. Before inputting the seating data into a preset comfort assessment model, the method further includes: extracting posture features from the body posture image data, obtaining posture parameters of various parts of the occupant's body through a human key point detection algorithm, and generating a posture feature vector; extracting physiological features from the vital sign data, obtaining rhythmic feature parameters representing physiological states through signal filtering and feature analysis algorithms, and generating a physiological feature vector; extracting dynamic features from the vehicle driving data, obtaining environmental feature parameters representing the vehicle's motion state through a time-frequency analysis algorithm, and generating an environmental feature vector; aligning the posture feature vector, physiological feature vector, and environmental feature vector in time synchronization and concatenating their features to form a multimodal feature matrix; and inputting the multimodal feature matrix into the preset comfort assessment model for comfort assessment.

[0012] According to one embodiment of this application, the method further includes: acquiring historical seating data streams, each historical seating data slice including synchronously acquired posture image data, vital sign data, and vehicle driving data; labeling each historical seating data slice to generate a corresponding comfort label vector, the comfort label vector including a comfort score for at least one body region; preprocessing the posture image data, vital sign data, and vehicle driving data of each historical seating data slice to determine corresponding posture feature vectors, physiological feature vectors, and environmental feature vectors; and performing time synchronization alignment on the posture feature vectors, physiological feature vectors, and environmental feature vectors. Features are concatenated to form a second input feature. A model training set for a preset comfort assessment model is constructed based on multiple second input features and corresponding comfort label vectors. Any second input feature from the model training set of the preset comfort assessment model is input into the initial comfort assessment model to obtain a predicted comfort vector. The discriminant loss is calculated based on the predicted comfort vector and the corresponding comfort label vector in the model training set of the preset comfort assessment model to obtain a second calculation result. The model parameters of the initial comfort assessment model are updated using the second calculation result until the updated initial comfort assessment model satisfies the second preset convergence condition, thus obtaining the preset comfort assessment model.

[0013] According to one embodiment of this application, occupant feedback instructions include voice adjustment instructions and seat adjustment instructions input via an in-vehicle terminal and a mobile terminal.

[0014] According to one embodiment of this application, the method further includes: if no occupant feedback instruction is received within a preset time, determining a comfort score for at least one body region based on a comfort vector; and determining the adjustable component of the seat and its adjustment parameters based on the comfort score for at least one body region.

[0015] To achieve the above objectives, a second aspect of this application provides a seat adjustment device, comprising: a data acquisition module for acquiring occupant seating data; a first output module for inputting the seating data into a preset comfort evaluation model to output a comfort vector; a second output module for inputting the comfort vector and the occupant feedback instruction into a preset seat adjustment parameter prediction model upon receiving an occupant feedback instruction within a preset time to output target seat adjustment parameters; and an adjustment module for determining the adjustable components of the seat based on the target seat adjustment parameters and adjusting the adjustable components based on the target seat adjustment parameters.

[0016] To achieve the above objectives, a third aspect of this application provides a computer-readable storage medium storing a seat adjustment program that, when executed by a processor, implements the aforementioned seat adjustment method.

[0017] To achieve the above objectives, a fourth aspect of this application provides a vehicle including a memory, a processor, and a seat adjustment program stored in the memory and executable on the processor. When the processor executes the seat adjustment program, it implements the aforementioned seat adjustment method.

[0018] According to the seat adjustment method, device, vehicle, and storage medium of this application embodiment, occupant seating data is collected and input into a preset comfort evaluation model to output a comfort vector. If an occupant feedback command is received within a preset time, the comfort vector and the occupant feedback command are input into a preset seat adjustment parameter prediction model to output target seat adjustment parameters. The components to be adjusted in the seat are determined based on the target seat adjustment parameters, and adjustments are made to these components based on the target seat adjustment parameters. This application continuously collects occupant seating data and inputs it into a preset comfort evaluation model to generate a multi-dimensional comfort vector, thereby achieving real-time quantitative perception of passenger comfort. When an occupant's active feedback command is received, the feedback command and the real-time comfort vector are input into the prediction model together, outputting adjustment parameters that integrate objective state and subjective intention. Finally, the adjustment parameters precisely drive the components to be adjusted to work collaboratively, overcoming the shortcomings of seat adjustment which can only achieve fixed modes or preset gear adjustments. This achieves intelligent, real-time, and personalized seat adjustment, meeting the higher requirements for car seat comfort in modern automotive development and significantly improving the experience and quality of car seats. Attached Figure Description

[0019] Figure 1 Here is a flowchart of a seat adjustment method according to some embodiments of this application; Figure 2 This is a block diagram of a seat adjustment device according to some embodiments of this application; Figure 3This is a block diagram of a vehicle according to some embodiments of this application. Detailed Implementation

[0020] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0021] The seat adjustment method, device, vehicle, and storage medium of this application are described in detail below with reference to the accompanying drawings.

[0022] Figure 1 This is a flowchart of a seat adjustment method according to some embodiments of this application. (Refer to...) Figure 1 The seat adjustment method in this application embodiment may include the following steps: S110 collects the occupant's seating data and inputs the seating data into a preset comfort assessment model to output a comfort vector.

[0023] Specifically, the occupant's seating data includes, but is not limited to, body image data, vital sign data, and vehicle driving data. Among these, vehicle driving data includes, but is not limited to, vehicle speed, road conditions, and vehicle vibration information during vehicle operation.

[0024] Occupant seating data can be acquired through multimodal sensors, including high-resolution cameras, infrared sensors, temperature and humidity sensors, and inertial measurement units (IMUs). These sensors collect occupant posture images, vital signs (such as heart rate and body temperature), and vehicle driving data. Vehicle driving data can include parameters such as vehicle speed, road conditions, vehicle vibration, and acceleration, as well as ambient temperature. The acquisition of multimodal data provides a comprehensive and three-dimensional reflection of the occupant's current physical state and environment, providing a data foundation for subsequent intelligent adjustments.

[0025] Furthermore, after collecting occupant seating data, this data can be input into a preset comfort assessment model to output a comfort vector. This preset comfort assessment model has a multi-layer convolutional neural network (CNN) structure and a recurrent neural network (RNN) module. It can extract and fuse features from occupant posture, vital signs, and vehicle driving data, accurately assess the occupant's current perceived comfort, and form a comfort vector. The comfort vector includes a comfort score for at least one body region and an environmental comfort score. The comfort score integrates visual, behavioral, and physiological features, achieving a quantitative description of occupant comfort for subsequent decision-making.

[0026] S120: Upon receiving occupant feedback instructions within a preset time, the comfort vector and occupant feedback instructions are input into a preset seat adjustment parameter prediction model to output target seat adjustment parameters. The preset time can be calibrated according to actual conditions; for example, it could be 20 seconds, but no specific limitation is imposed here.

[0027] Specifically, after collecting occupant seating data and inputting it into a preset comfort assessment model to obtain a comfort vector, if occupant feedback commands are received within a preset time, the comfort vector and the occupant feedback commands are input into a preset seat adjustment parameter prediction model to output target seat adjustment parameters. Occupant feedback commands include voice adjustment commands, seat adjustment commands input via in-vehicle terminals and mobile terminals.

[0028] In this way, when occupant feedback is received within a reasonable time window, both are used as a combined input. The preset seat adjustment parameter prediction model is used to output more accurate target seat adjustment parameters. This not only enhances the ability to respond to occupant's real-time intentions, but also makes the seat adjustment more in line with the occupant's subjective comfort needs through a closed-loop feedback mechanism of human-computer interaction, thereby significantly improving the practicality of the seat and user satisfaction.

[0029] S130: Determine the components of the seat to be adjusted based on the target seat adjustment parameters, and adjust the components to be adjusted based on the target seat adjustment parameters.

[0030] Specifically, after determining the target seat adjustment parameters, the vehicle drives the corresponding actuators (such as motors, air pumps, damping adjusters, etc.) in the seat to coordinately adjust various dimensions of the seat according to the target seat adjustment parameters. The adjustment process includes, but is not limited to, adjusting the seat height, fore-aft position, backrest tilt angle, seat cushion angle, lumbar support strength, side wing wrapping, local massage function parameters, heating function parameters, ventilation function parameters, etc.

[0031] This application continuously collects passenger seating data and inputs it into a preset comfort assessment model to generate a multi-dimensional comfort vector, thereby achieving real-time quantitative perception of passenger comfort status. When a passenger's active feedback instruction is received, the feedback instruction and the real-time comfort vector are input into a prediction model, outputting adjustment parameters that integrate objective state and subjective intention. Finally, based on these adjustment parameters, the components to be adjusted are precisely driven to operate collaboratively, overcoming the shortcomings of seat adjustment which can only achieve fixed modes or preset gear adjustments. This realizes intelligent, real-time, and personalized seat adjustment, meeting the higher requirements for car seat comfort under the development of modern automobiles, and significantly improving the experience and quality of car seats.

[0032] In some embodiments, before inputting the comfort vector and occupant feedback instructions into the preset seat adjustment parameter prediction model, the method further includes: structuring the occupant feedback instructions to generate an occupant preference feature vector; fusing the comfort vector and the occupant preference feature vector to generate a fused feature vector; and inputting the fused feature vector into the preset seat adjustment parameter prediction model to output the target seat adjustment parameters.

[0033] Specifically, before inputting the comfort vector and occupant feedback commands into the preset seat adjustment parameter prediction model, the comfort vector and occupant feedback commands need to be preprocessed. For example, firstly, through structured processing, discrete occupant feedback commands (such as voice adjustment of lumbar support or operation of the forward tilt button) are parsed, classified, and encoded, and associated with the context to extract the user's steady-state preference pattern (such as habitual high lumbar support intensity), which is quantified into a feature vector; then, the comfort vector and occupant preference feature vector are time-aligned and weighted to generate a multi-dimensional, complementary fused feature vector; finally, the fused feature vector is input into the preset seat adjustment parameter prediction model to output the target seat adjustment parameters.

[0034] In this way, the preset seat adjustment parameter prediction model can not only understand where the body is uncomfortable at this moment, but also combine it with how the occupant usually likes to relieve this discomfort. This results in outputting precise adjustment parameters that meet both real-time comfort needs and long-term usage habits, significantly improving the accuracy of adjustment, user satisfaction, and the intelligent experience of the system.

[0035] In some embodiments, the method further includes: acquiring a historical seating data stream and historical feedback instructions and historical seat adjustment parameters corresponding to each historical seating data slice of the historical seating data stream; inputting each historical seating data slice of the historical seating data stream into a preset comfort evaluation model to obtain a corresponding historical comfort vector; performing structured processing on the corresponding historical feedback instructions to obtain an occupant preference feature vector; performing feature fusion between the occupant preference feature vector and the corresponding historical comfort vector to generate a first input feature; constructing a model training set for a preset seat adjustment parameter prediction model based on multiple first input features and corresponding historical seat adjustment parameters; inputting any first input feature from the model training set of the preset seat adjustment parameter prediction model into an initial seat adjustment parameter prediction model to obtain predicted seat adjustment parameters; calculating a discriminant loss based on the predicted seat adjustment parameters and the corresponding seat adjustment parameters in the model training set of the preset seat adjustment parameter prediction model to obtain a first calculation result; updating the model parameters of the initial seat adjustment parameter prediction model using the first calculation result until the updated initial seat adjustment parameter prediction model satisfies a first preset convergence condition to obtain the preset seat adjustment parameter prediction model.

[0036] Specifically, the process begins by collecting continuous historical seating data streams and corresponding time slices containing user feedback commands and actual seat adjustment parameters. For each historical seating data slice, a pre-defined comfort assessment model is used to transform it into a historical comfort vector representing the objective comfort state. Simultaneously, the corresponding historical feedback commands are structured (including command parsing, classification encoding, and preference pattern extraction) to generate an occupant preference feature vector reflecting the user's subjective preferences. Subsequently, the occupant preference feature vector is fused with the corresponding historical comfort vector (e.g., through weighted concatenation or attention mechanism fusion) to generate a first input feature that simultaneously contains both objective state and subjective inclination. Finally, by summarizing the first input features of all historical seating data slices and their corresponding historical seat adjustment parameters, a complete model training set can be constructed.

[0037] During the model training phase, any first input feature is input into the initial seat adjustment parameter prediction model, which outputs the predicted seat adjustment parameters. The predicted seat adjustment parameters are then compared with the corresponding historical seat adjustment parameters recorded in the training set, and a first calculation result is calculated using a loss function (such as root mean square error). Based on the first calculation result, the model parameters are iteratively updated using optimization methods such as backpropagation until the model prediction accuracy reaches the first preset convergence condition, such as reaching a preset number of iterations. Finally, a pre-trained seat adjustment parameter prediction model that can accurately correlate the user's multi-dimensional state and personalized adjustment needs is obtained.

[0038] For example, the initial seat adjustment parameter prediction model can be a fully connected feedforward neural network architecture. This network consists of an input layer, two hidden layers, and an output layer connected sequentially. The layers are fully connected, meaning that each neuron in each layer is connected to all neurons in the next layer. The input layer receives the first input feature vector. The first hidden layer contains 64 neurons and uses the ReLU (Rectified Linear Unit) activation function to introduce nonlinearity and alleviate gradient vanishing. The second hidden layer contains 32 neurons and also uses the ReLU activation function. The output layer contains 8 neurons, corresponding to 8-dimensional seat adjustment parameters, and uses a linear activation function to achieve regression prediction of continuous values.

[0039] To ensure the model's performance and reproducibility, the selection criteria, adjustment range, and optimization strategies for key hyperparameters have been fully documented and validated through systematic experimental design: the learning rate was optimized using grid search within the range of 0.0001 to 0.01, and was finally selected as 0.001 based on the convergence stability of the validation set loss. An exponential decay strategy (decay coefficient of 0.9 every 50 rounds) was adopted to improve the accuracy of later training. The batch size was tested in 16, 32, 64, and 128, and 32 was selected based on the balance between training efficiency and gradient stability. Among the network structure hyperparameters, the number of hidden layer neurons was selected from the combination space {[32,16], [64,32], [128,64]} based on the trade-off between model capacity and validation set performance, with the [64,32] structure selected. The regularization hyperparameter Dropout ratio was selected as 0.2 within the range of 0.1 to 0.5 based on generalization performance evaluation. All hyperparameter optimizations were validated using five-fold cross-validation, supplemented by learning curve and validation set loss monitoring to ensure sufficient selection criteria and traceable adjustment processes.

[0040] On the same test set, the model's prediction results showed a root mean square error (RMSE) of 7.3 and a mean absolute error (MAE) of 4.2, demonstrating significant advantages over traditional rule-matching methods (RMSE 18.7) and support vector machine regression models (RMSE 14.2). Further ablation experiments showed that removing the comfort vector input increased the RMSE to 11.2, while removing user preference features increased it to 13.8, confirming the necessity of multi-feature fusion. During training, the validation set loss steadily converged to 7.9 without significant overfitting. All experiments were based on the same data partitioning and evaluation criteria, ensuring the comparability of the results and the independent verification of the actual effectiveness of the proposed solution.

[0041] In this way, the ability to automatically learn complex nonlinear mapping relationships from massive historical interaction data is realized, enabling the model not only to react based on real-time data, but also to predict users' potential preferences by mining historical patterns, thereby significantly improving the predictive accuracy, personalization and overall intelligence level of the system.

[0042] In some embodiments, the seating data includes body posture image data, vital sign data, and vehicle driving data. Before inputting the seating data into a preset comfort assessment model, the method further includes: extracting posture features from the body posture image data, obtaining posture parameters of various parts of the occupant's body through a human key point detection algorithm, and generating a posture feature vector; extracting physiological features from the vital sign data, obtaining rhythmic feature parameters representing physiological states through signal filtering and feature analysis algorithms, and generating a physiological feature vector; extracting dynamic features from the vehicle driving data, obtaining environmental feature parameters representing the vehicle's motion state through a time-frequency analysis algorithm, and generating an environmental feature vector; aligning the posture feature vector, physiological feature vector, and environmental feature vector in time synchronization and concatenating their features to form a multimodal feature matrix; and inputting the multimodal feature matrix into the preset comfort assessment model for comfort assessment.

[0043] Specifically, the seating data includes posture image data, vital sign data, and vehicle driving data. Before inputting the seating data into the preset comfort assessment model, it needs to be preprocessed. For example, feature extraction is first performed on the posture image data, vital sign data, and vehicle driving data respectively. For instance, posture feature extraction is performed on the posture image data by obtaining posture parameters of various parts of the occupant's body through a human key point detection algorithm, generating a posture feature vector. Physiological feature extraction is performed on the vital sign data by obtaining rhythmic feature parameters representing physiological states through signal filtering and feature analysis algorithms, generating a physiological feature vector. Dynamic feature extraction is performed on the vehicle driving data by obtaining environmental feature parameters representing the vehicle's motion state through a time-frequency analysis algorithm, generating an environmental feature vector. Then, the posture feature vector, physiological feature vector, and environmental feature vector, which are synchronized in time, are concatenated into a multimodal feature matrix. Finally, the multimodal feature matrix is ​​input into the preset comfort assessment model for comfort assessment.

[0044] In this way, the limitations of evaluation based on a single data source are overcome, and the objective and refined quantification of subjective comfort is achieved. This provides a more reliable and comprehensive basis for subsequent intelligent adjustment, enabling the seat system to truly understand the comfort needs of occupants in complex dynamic environments, thereby making more accurate and timely adjustment responses.

[0045] In some embodiments, the method further includes: acquiring historical seating data streams, each historical seating data slice including synchronously acquired posture image data, vital sign data, and vehicle driving data; labeling each historical seating data slice to generate a corresponding comfort label vector, the comfort label vector including a comfort score for at least one body region; preprocessing the posture image data, vital sign data, and vehicle driving data of each historical seating data slice to determine corresponding posture feature vectors, physiological feature vectors, and environmental feature vectors; and performing time synchronization alignment and feature matching on the posture feature vectors, physiological feature vectors, and environmental feature vectors. A second input feature is formed by concatenating multiple second input features and corresponding comfort label vectors. A model training set for a preset comfort assessment model is constructed based on these second input features and corresponding comfort label vectors. Any second input feature from the model training set of the preset comfort assessment model is input into the initial comfort assessment model to obtain a predicted comfort vector. The discriminant loss is calculated based on the predicted comfort vector and the corresponding comfort label vector in the model training set of the preset comfort assessment model to obtain a second calculation result. The model parameters of the initial comfort assessment model are updated using the second calculation result until the updated initial comfort assessment model satisfies the second preset convergence condition, thus obtaining the preset comfort assessment model.

[0046] Specifically, the process begins by collecting a synchronous historical data stream containing body posture images, physiological signs, and vehicle driving data. This stream is then segmented into time slices, and each slice is manually or semi-automatically labeled to generate a comfort label vector containing a specific comfort score for at least one body region (e.g., waist, back, legs). Simultaneously, feature extraction is performed on the raw data within each slice. For example, key point coordinates and angle information are extracted from body posture images using a human posture estimation algorithm to form a posture feature vector; time-frequency analysis and rhythm feature extraction are performed on physiological signals (e.g., heart rate variability, skin conductance response) to generate a physiological feature vector; and vibration spectrum and motion features are extracted from dynamic data such as vehicle acceleration and angular velocity through signal processing to obtain an environmental feature vector.

[0047] After aligning the three feature vectors within the same time slice using timestamps, a unified second input feature is formed through feature concatenation or fusion layer operations. The second input features of all slices are then paired with their corresponding comfort label vectors to construct a training set for the pre-defined comfort assessment model.

[0048] During the model training phase, the second input feature is fed into the initial preset comfort assessment model, and the model outputs a predicted comfort vector. The difference between the predicted vector and the true label vector is calculated (e.g., using the mean squared error loss function) to obtain the second calculation result (i.e., the loss value). The gradient of the loss with respect to the parameters of each layer of the model is calculated using the backpropagation algorithm, and the model weights and bias parameters are updated using stochastic gradient descent or its variant optimizer (e.g., Adam). This process is repeated iteratively on the training set. In each iteration, the model parameters are fine-tuned based on the loss calculated from a batch of data; simultaneously, the model performance is monitored using a validation set to prevent overfitting. Training terminates when the model's loss on the validation set no longer decreases significantly (satisfying the second preset convergence condition, such as the loss change being less than a threshold for 10 consecutive iterations), or when the preset maximum number of iterations is reached. The model obtained at this point is the preset comfort assessment model capable of accurately evaluating multi-dimensional comfort.

[0049] For example, the initial preset comfort assessment model can adopt a multi-layer fully connected neural network architecture, specifically consisting of an input layer, three hidden layers, and an output layer connected sequentially. The input layer receives a second input feature vector formed by time alignment and feature concatenation. This vector is composed of postural features (such as joint angles), physiological features (such as heart rate variability), and environmental features (such as vibration spectrum). The first hidden layer contains 256 neurons, uses the ReLU activation function, and introduces Dropout regularization (ratio 0.3); the second hidden layer contains 128 neurons, also using the ReLU activation function; the third hidden layer contains 64 neurons, using the Leaky ReLU activation function to alleviate neuronal inhibition. The number of neurons in the output layer is consistent with the dimension of the comfort label vector (e.g., corresponding to 5 body regions), and the Sigmoid activation function is used to map the output to a comfort score in the range of 0-1. All inter-layer connections are fully connected, and the output of each neuron serves as the weighted input of all neurons in the next layer. The model training uses an adaptive moment estimation optimizer with an initial learning rate of 0.001 and a cosine annealing scheduling strategy. The loss function uses mean squared error with weighted coefficients to balance the differences in importance of different body regions. During training, the validation set loss is monitored through an early stopping mechanism, and training is terminated early when the loss decreases by no more than 1e-4 for 15 consecutive rounds.

[0050] The number of hidden layer neurons in the network structure hyperparameters was searched using a grid search within the combination space of {[128,64,32], [256,128,64], [512,256,128]}. Based on the balance between the validation set mean squared error and model complexity, the [256,128,64] structure was ultimately selected. For activation functions, ReLU, Leaky ReLU (negative slope 0.01), and ELU were compared and tested in the hidden layers. Based on training stability, ReLU was chosen for the first two layers and Leaky ReLU for the third layer. The dropout ratio was fine-tuned in the range of 0.2-0.5 with a step size of 0.05, and 0.3 was determined as the optimal value based on the overfitting degree of the validation set. For optimizer selection, Adam, Nadam, and RMSprop were compared, and Adam was selected based on convergence speed. The learning rate was set within the range of [0.0001, ...]. [0.01] Logarithmic space sampling was used for testing, and the initial value was set to 0.001 using a cosine annealing strategy. The batch size was selected from {16, 32, 64, 128} based on training efficiency and memory limitations, with 32 selected. The weighting coefficients of the loss function were initialized using the inverse of the variance of each body region and dynamically adjusted during training. All hyperparameter optimizations used five-fold cross-validation, and the complete learning curve and validation metrics were recorded for each experiment to ensure the selection process was traceable and reproducible.

[0051] On the independent test set, the mean root mean square error (RMSE) between the predicted comfort scores and the manually labeled true values ​​was 0.124, and the mean absolute error (RMSE) was 0.087, significantly outperforming traditional methods based on support vector machines (RMSE 0.215), random forest models (RMSE 0.192), and baseline neural networks without multi-feature fusion (RMSE 0.158). In the detailed evaluation of different body regions, this model performed best in predicting lumbar comfort (RMSE 0.101), maintaining stable performance compared to the worst-performing leg region (RMSE 0.142). Ablation experiments further validated the effectiveness of multimodal features: the RMSE increased to 0.183 when using only posture features, 0.201 when using only physiological features, and 0.217 when using only environmental features, fully demonstrating the necessity of multi-feature fusion. Model robustness tests show that the prediction error standard deviations are only 0.015 and 0.012 in two typical scenarios: bumpy roads and long journeys. Training process monitoring indicates that the model converges after 135 epochs, with a final validation set loss of 0.131 and no overfitting (training / validation loss ratio of 1.02). All experiments were conducted on the same labeled dataset using the same partitioning method to ensure comparability and reproducibility of results.

[0052] In this way, by synchronously collecting and slicing body posture images, physiological signs, and vehicle dynamic data, and performing multi-dimensional feature extraction and time alignment, and further combining them with comfort label vectors containing specific body area scores, a high-quality supervised training sample set is formed. This enables the model to learn the accurate mapping relationship from complex multi-source data to subjective comfort. During the training process, through iterative prediction and loss calculation, and by continuously updating the model parameters using gradient optimization algorithms, the final preset comfort evaluation model has strong generalization ability and robustness. It can accurately and in real time output comfort vectors that reflect the actual feelings of occupants, providing a reliable and interpretable quantitative basis for subsequent intelligent adjustment decisions, thereby fundamentally improving the perception accuracy and intelligence level of the seat comfort system.

[0053] In some embodiments, the method further includes: if no occupant feedback instruction is received within a preset time, determining a comfort score for at least one body region based on a comfort vector; and determining the adjustable component of the seat and its adjustment parameters based on the comfort score for at least one body region.

[0054] Specifically, if no occupant feedback instruction is received within a preset time, the comfort score of at least one body area (e.g., lower back, legs) will be determined directly based on the comfort vector. Then, the adjustment parameters of the seat components to be adjusted and the adjustment parameters of these components will be determined based on the comfort scores of at least one body area. For example, if the lower back comfort score is low based on the comfort vector, the lumbar support will be identified as the component to be adjusted. The difference between the lower back comfort score and the standard comfort score will be calculated, and the adjustment parameters of the lumbar support will be determined based on this difference. For example, the adjustment parameters of the lumbar support can be determined by querying a preset mapping table between the comfort score difference and the adjustment parameters of the lumbar support. For example, the adjustment parameters of the lumbar support may include increasing the pressure of the central airbag by 0.15 bar and reducing the side wing support angle by 5 degrees. This preset mapping table includes multiple query comfort score differences and the corresponding adjustment parameters of the lumbar support for each comfort score difference.

[0055] In addition, an environmental comfort score, such as a temperature score, can be determined based on a comfort vector. This temperature score then determines the components of the seat to be adjusted and their adjustment parameters. For example, if the temperature score deviates from a preset comfort range (e.g., a low score indicates a feeling of coldness, and a high score indicates a feeling of heat), components related to environmental thermal comfort will be identified for adjustment, such as the seat heating pad or ventilation fan. Subsequently, the difference between the temperature score and the target value is mapped to specific adjustment parameters using a proportional-integral control algorithm or fuzzy logic rules. For instance, when it is too cold, the heating pad is automatically activated in stages, and the heating power is set to the level corresponding to the target temperature difference (e.g., set to level 3 to maintain a surface temperature of 38°C). When it is too hot, the ventilation system is activated and the fan speed is controlled according to similar logic.

[0056] In this way, in scenarios where there is no active feedback from passengers, the adjustment parameters of the seat components to be adjusted are determined directly based on the comfort vector. This not only ensures the timeliness and continuity of comfort maintenance, but also achieves multi-dimensional adaptive optimization from local support to overall thermal comfort through the coordinated consideration of the environment and body condition. This significantly reduces the frequency of manual intervention by passengers and improves the automation and intelligence level of the riding experience.

[0057] It should be noted that both the preset comfort assessment model and the preset seat adjustment parameter prediction model in this application embodiment include a self-learning mechanism, enabling the model to continuously adapt to occupant needs and environmental changes, and gradually improve the accuracy of the algorithm's prediction of adjustment parameters. Through the accumulation and relearning of historical data, a database for personalized adjustment needs of different occupants is gradually established, thereby providing more accurate and comfortable adjustment parameters directly to new occupants in the future, achieving continuous intelligent evolution.

[0058] Corresponding to the above embodiments, this application also proposes a seat adjustment device.

[0059] Reference Figure 2 The seat adjustment device 200 includes: a data acquisition module 210, a first output module 220, a second output module 230, and an adjustment module 240.

[0060] The acquisition module 210 is used to collect occupant seating data. The first output module 220 is used to input the seating data into a preset comfort assessment model to output a comfort vector. The second output module 230, upon receiving occupant feedback instructions within a preset time, inputs the comfort vector and the occupant feedback instructions into a preset seat adjustment parameter prediction model to output target seat adjustment parameters. The adjustment module 240 is used to determine the adjustable components of the seat based on the target seat adjustment parameters and adjust the adjustable components based on the target seat adjustment parameters.

[0061] According to one embodiment of this application, before inputting the comfort vector and occupant feedback instructions into the preset seat adjustment parameter prediction model, the occupant feedback instructions are structured to generate an occupant preference feature vector; the comfort vector and the occupant preference feature vector are fused to generate a fused feature vector; the fused feature vector is input into the preset seat adjustment parameter prediction model to output the target seat adjustment parameters.

[0062] According to one embodiment of this application, historical seating data streams and historical feedback instructions and historical seat adjustment parameters corresponding to each historical seating data slice of the historical seating data streams are obtained; each historical seating data slice of the historical seating data stream is input into a preset comfort evaluation model to obtain a corresponding historical comfort vector; the corresponding historical feedback instructions are structured to obtain an occupant preference feature vector; the occupant preference feature vector and the corresponding historical comfort vector are fused to generate a first input feature; a model training set for a preset seat adjustment parameter prediction model is constructed based on multiple first input features and corresponding historical seat adjustment parameters; any first input feature in the model training set of the preset seat adjustment parameter prediction model is input into an initial seat adjustment parameter prediction model to obtain predicted seat adjustment parameters; a discrimination loss is calculated based on the predicted seat adjustment parameters and the corresponding seat adjustment parameters in the model training set of the preset seat adjustment parameter prediction model to obtain a first calculation result; the model parameters of the initial seat adjustment parameter prediction model are updated using the first calculation result until the updated initial seat adjustment parameter prediction model satisfies a first preset convergence condition, thus obtaining the preset seat adjustment parameter prediction model.

[0063] According to one embodiment of this application, the seating data includes body posture image data, vital sign data, and vehicle driving data. Before inputting the seating data into a preset comfort assessment model, posture features are extracted from the body posture image data. Posture parameters of various parts of the occupant's body are obtained through a human key point detection algorithm to generate a posture feature vector. Physiological features are extracted from the vital sign data. Rhythmic feature parameters representing physiological states are obtained through signal filtering and feature analysis algorithms to generate a physiological feature vector. Dynamic features are extracted from the vehicle driving data. Environmental feature parameters representing the vehicle's motion state are obtained through a time-frequency analysis algorithm to generate an environmental feature vector. The posture feature vector, physiological feature vector, and environmental feature vector are time-synchronized and aligned and feature-stitched to form a multimodal feature matrix. The multimodal feature matrix is ​​then input into the preset comfort assessment model for comfort assessment.

[0064] According to one embodiment of this application, historical seating data streams are collected. Each historical seating data slice in the historical seating data stream includes synchronously collected posture image data, vital sign data, and vehicle driving data. Each historical seating data slice is labeled to generate a corresponding comfort label vector, which includes a comfort score for at least one body region. The posture image data, vital sign data, and vehicle driving data of each historical seating data slice are preprocessed to determine the corresponding posture feature vector, physiological feature vector, and environmental feature vector. The posture feature vector, physiological feature vector, and environmental feature vector are then time-synchronized aligned and feature-stitched. A second input feature is formed, and a model training set for a preset comfort assessment model is constructed based on multiple second input features and corresponding comfort label vectors. Any second input feature in the model training set of the preset comfort assessment model is input into the initial comfort assessment model to obtain a predicted comfort vector. The discriminant loss is calculated based on the predicted comfort vector and the corresponding comfort label vector in the model training set of the preset comfort assessment model to obtain a second calculation result. The model parameters of the initial comfort assessment model are updated using the second calculation result until the updated initial comfort assessment model satisfies the second preset convergence condition, thus obtaining the preset comfort assessment model.

[0065] According to one embodiment of this application, occupant feedback instructions include voice adjustment instructions and seat adjustment instructions input via an in-vehicle terminal and a mobile terminal.

[0066] According to one embodiment of this application, if no occupant feedback instruction is received within a preset time, a comfort score for at least one body region is determined based on a comfort vector; and the adjustment parameters of the seat components to be adjusted and the adjustment parameters of the components to be adjusted are determined based on the comfort scores of the at least one body region. Corresponding to the above embodiment, this application also proposes a computer-readable storage medium.

[0067] The computer-readable storage medium of this application stores a seat adjustment program thereon, which, when executed by a processor, implements the aforementioned seat adjustment method.

[0068] It should be noted that the above explanation of the embodiments and beneficial effects of the seat adjustment method also applies to the computer-readable storage medium of the embodiments of this application. To avoid redundancy, it will not be elaborated in detail here.

[0069] Corresponding to the above embodiments, this application also proposes a vehicle.

[0070] See Figure 3 As shown, the vehicle 300 of this application includes a memory 310, a processor 320, and a seat adjustment program stored in the memory 310 and executable on the processor 320. When the processor executes the seat adjustment program, it implements the aforementioned seat adjustment method.

[0071] It should be noted that the above explanation of the embodiments and beneficial effects of the seat adjustment method also applies to the vehicles in the embodiments of this application, and will not be elaborated in detail here to avoid redundancy.

[0072] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0073] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0074] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0075] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0076] In this application, unless otherwise expressly specified and limited, the terms "installation," "connection," "joining," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise expressly limited. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0077] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for adjusting a seat, characterized in that, The method includes: Passenger seating data is collected and input into a preset comfort assessment model to output a comfort vector; If an occupant feedback command is received within a preset time, the comfort vector and the occupant feedback command are input into a preset seat adjustment parameter prediction model to output the target seat adjustment parameters. The components of the seat to be adjusted are determined based on the target seat adjustment parameters, and the components to be adjusted are adjusted based on the target seat adjustment parameters.

2. The seat adjustment method according to claim 1, characterized in that, Before inputting the comfort vector and the occupant feedback command into the preset seat adjustment parameter prediction model, the method further includes: The occupant feedback instructions are structured to generate occupant preference feature vectors; The comfort vector and the occupant preference feature vector are fused to generate a fused feature vector. The fused feature vector is input into the preset seat adjustment parameter prediction model to output the target seat adjustment parameters.

3. The seat adjustment method according to claim 2, characterized in that, The method further includes: Acquire historical seating data streams and historical feedback commands and historical seat adjustment parameters corresponding to each historical seating data slice of the historical seating data stream; Each historical seating data slice of the historical seating data stream is input into the preset comfort evaluation model to obtain the corresponding historical comfort vector. The corresponding historical feedback instructions are structured to obtain the occupant preference feature vector. The occupant preference feature vector and the corresponding historical comfort vector are feature fused to generate the first input feature. The model training set of the preset seat adjustment parameter prediction model is constructed based on multiple first input features and the corresponding historical seat adjustment parameters. Input any first input feature from the model training set of the preset seat adjustment parameter prediction model into the initial seat adjustment parameter prediction model to obtain the predicted seat adjustment parameters; The discrimination loss is calculated based on the predicted seat adjustment parameters and the seat adjustment parameters corresponding to the model training set of the preset seat adjustment parameter prediction model to obtain the first calculation result; The model parameters of the initial seat adjustment parameter prediction model are updated using the first calculation result until the updated initial seat adjustment parameter prediction model meets the first preset convergence condition, thus obtaining the preset seat adjustment parameter prediction model.

4. The seat adjustment method according to claim 1, characterized in that, The seating data includes body posture image data, vital sign data, and vehicle driving data. Before inputting the seating data into a preset comfort assessment model, the method further includes: The posture features of the body image data are extracted, and the posture parameters of each part of the occupant's body are obtained by the human key point detection algorithm to generate a posture feature vector. Physiological features are extracted from the vital signs data, and rhythmic feature parameters representing physiological states are obtained through signal filtering and feature analysis algorithms to generate physiological feature vectors. Dynamic features are extracted from the vehicle driving data, and environmental feature parameters characterizing the vehicle's motion state are obtained through a time-frequency analysis algorithm to generate an environmental feature vector. The posture feature vector, the physiological feature vector, and the environmental feature vector are time-synchronized, aligned, and concatenated to form a multimodal feature matrix; The multimodal feature matrix is ​​input into the preset comfort assessment model for comfort assessment.

5. The seat adjustment method according to claim 4, characterized in that, The method further includes: Collect historical sitting data streams. Each historical sitting data slice in the historical sitting data stream includes synchronously collected body posture image data, vital sign data, and vehicle driving data. Each historical sitting data slice is labeled to generate a corresponding comfort label vector. The comfort label vector includes a comfort score for at least one body region. The body image data, vital sign data and vehicle driving data of each historical sitting data slice are preprocessed to determine the corresponding posture feature vector, physiological feature vector and environmental feature vector. The posture feature vector, the physiological feature vector and the environmental feature vector are time-synchronized and aligned and feature spliced ​​to form a second input feature. The model training set of the preset comfort evaluation model is constructed based on multiple second input features and corresponding comfort label vectors. Input any second input feature from the model training set of the preset comfort assessment model into the initial comfort assessment model to obtain the predicted comfort vector; The discrimination loss is calculated based on the predicted comfort vector and the corresponding comfort label vector in the model training set of the preset comfort evaluation model to obtain the second calculation result; The model parameters of the initial comfort assessment model are updated using the second calculation result until the updated initial comfort assessment model meets the second preset convergence condition, thus obtaining the preset comfort assessment model.

6. The seat adjustment method according to claim 1, characterized in that, The occupant feedback commands include voice adjustment commands and seat adjustment commands input via in-vehicle terminals and mobile terminals.

7. The seat adjustment method according to claim 1, characterized in that, The method further includes: If no occupant feedback instruction is received within a preset time, a comfort score for at least one body region is determined based on the comfort vector. The adjustable components of the seat and the adjustment parameters of the adjustable components are determined based on the comfort scores of the at least one body area.

8. A seat adjustment device, characterized in that, The device includes: The data acquisition module is used to collect passenger seating data. The first output module is used to input the seating data into a preset comfort evaluation model to output a comfort vector. The second output module is used to input the comfort vector and the occupant feedback instruction into a preset seat adjustment parameter prediction model when an occupant feedback instruction is received within a preset time, so as to output the target seat adjustment parameters. An adjustment module is used to determine the adjustable components of the seat based on the target seat adjustment parameters, and to adjust the adjustable components based on the target seat adjustment parameters.

9. A computer-readable storage medium, characterized in that, It stores a seat adjustment program, which, when executed by a processor, implements the seat adjustment method according to any one of claims 1-7.

10. A vehicle, characterized in that, The system includes a memory, a processor, and a seat adjustment program stored in the memory and executable on the processor. When the processor executes the seat adjustment program, it implements the seat adjustment method according to any one of claims 1-7.

Images