Adaptive vibration suppression method for vehicle seat based on projection iterative disturbance rejection

Abstract

The application discloses an adaptive vibration suppression method for a vehicle seat based on projection iteration disturbance rejection, after real-time sensing data and a current working condition characteristic vector are acquired, three types of disturbance components, i.e., road excitation, seat damper nonlinearity and system unmodeled dynamics, are decoupled and estimated by a physical information neural network extended state observer; based on the three types of disturbance components and the real-time sensing data, real-time control current is generated by a linear active disturbance rejection controller to suppress vibration; according to a current seat vibration comfort performance index and a current working condition characteristic vector trigger event, an initialization hyperparameter of a projection iteration optimization algorithm is output by calling a meta-learning optimizer; based on the hyperparameter, manifold projection iteration optimization is performed in a subspace defined by a vehicle seat-human coupled dynamics manifold model, and a controller parameter set and an observer network weight fine-tuning amount are updated and deployed. Through event triggering and meta-learning guided online parameter optimization, adaptive adjustment of the control system to changing working conditions is realized.

Description

Technical Field

[0001] This invention relates to the field of vehicle vibration control technology, and in particular to an adaptive vibration suppression method for vehicle seats based on projection iterative disturbance rejection. Background Technology

[0002] Vehicle seats are widely used in engineering vehicles, special vehicles, and high-end commercial vehicles to improve ride comfort and operational stability due to their advantages such as continuous damping force, wide adjustable range, and rapid response. They primarily counteract vibrations transmitted to the seat from road unevenness by adjusting the current input to the seat damper in real time, thereby changing the damping characteristics. However, effective vibration suppression faces multiple challenges: First, the seat damper itself has complex nonlinear and hysteretic dynamic characteristics, making it difficult to establish an accurate mathematical model, which constitutes a significant internal disturbance source; second, the road excitation encountered during vehicle operation is random and time-varying, representing a major external disturbance; furthermore, the suspension system also exhibits unmodeled dynamic characteristics. These factors combined make it difficult for traditional control methods based on fixed parameters or simple adaptive rules (such as ceiling damping control and fuzzy PID control) to maintain optimal vibration suppression performance under all operating conditions. Linear active disturbance rejection control (ADR) technology, by using an extended state observer to uniformly estimate and compensate for internal and external disturbances in the system, reduces the dependence on accurate models to some extent and enhances robustness. However, the core control parameters of this method (such as observer bandwidth and controller gain) are usually determined through empirical trial and error or offline optimization. Once determined, they remain fixed or can only be adjusted to a limited extent based on a finite set of prior rules. When vehicle operating conditions (such as load, speed, and road surface type) change significantly, these fixed parameter sets may not be able to keep the controller in an optimal operating state, leading to a decline in vibration suppression performance and limited adaptive capability. Therefore, how to enable the vehicle seat control system not only to resist known and unknown disturbances, but also to enable its core control parameters and observer model to optimize online, autonomously, and efficiently, thereby maintaining optimal vibration suppression under changing operating conditions, is a problem that urgently needs to be solved to achieve truly comprehensive adaptive control. Summary of the Invention

[0003] In view of this, the purpose of this invention is to propose an adaptive vibration suppression method for vehicle seats based on projection iterative disturbance rejection. This method uses an event-triggered meta-learning optimizer to adjust the parameters of the controller and observer online, thereby solving the problem that fixed-parameter active disturbance rejection control is unable to maintain optimal vibration suppression performance under varying operating conditions.

[0004] To achieve the above-mentioned technical objectives, the technical solution adopted by this invention is as follows:

[0005] An adaptive vibration suppression method for vehicle seats based on projection iterative disturbance rejection includes:

[0006] Acquire real-time sensor data and current operating condition feature vectors of the vehicle seat. The real-time sensor data includes seat vertical acceleration, seat telescopic displacement, and seat damper drive current.

[0007] Real-time sensor data is input into the physical information neural network extended state observer to decouple and estimate the vibration disturbance components transmitted by the seat, the nonlinear disturbance components of the seat damper, and the unmodeled dynamic disturbance components of the seat-human coupling system.

[0008] Based on the vibration disturbance components transmitted by the seat, the nonlinear disturbance components of the seat damper, the unmodeled dynamic disturbance components of the seat-human coupling system, and real-time sensing data, a real-time control current for the seat damper is generated through a linear active disturbance rejection controller to perform vibration suppression.

[0009] The current seat vibration comfort performance index is calculated based on real-time sensor data, and combined with the current operating condition feature vector, the parameter optimization process is determined by event triggering rules.

[0010] When the event triggering rule determines that the optimization conditions are met, the meta-learning optimizer is called. The meta-learning optimizer outputs the initial hyperparameters of the projection iterative optimization algorithm based on the feature vector of the current working condition.

[0011] Based on the initial hyperparameters, within the parameter subspace defined by the vehicle seat-human coupled dynamics manifold model, a manifold projection iterative optimization algorithm is executed to search and update the control parameter set of the linear active disturbance rejection controller and the network weight fine-tuning of the physical information neural network extended state observer.

[0012] The control parameter set and network weight fine-tuning are deployed to the linear active disturbance rejection controller and the physical information neural network extended state observer, respectively, to complete the adaptive adjustment.

[0013] In some embodiments, real-time sensing data is input to a physical information neural network extended state observer to decouple and estimate the seat-transmitted vibration disturbance component, the seat damper nonlinear disturbance component, and the unmodeled dynamic disturbance component of the seat-human coupling system, including:

[0014] The time window data, consisting of real-time sensing data and its historical sequence, is input into the encoder module of the physical information neural network extended state observer to extract high-dimensional mixed perturbation features.

[0015] The high-dimensional hybrid perturbation features are input into the parallel first feature decoupling branch, the second feature decoupling branch and the third feature decoupling branch. The first feature decoupling branch extracts the first feature related to the frequency domain characteristics of the seat transmitted vibration through the first set of filter weights. The second feature decoupling branch extracts the second feature related to the hysteresis nonlinearity of the seat damper through the second set of filter weights. The third feature decoupling branch extracts the remaining dynamic third features of the seat-human coupling system that are not modeled through the third set of filter weights.

[0016] The first feature, the second feature, and the third feature are respectively input into the corresponding fully connected regression layer to obtain the estimated values ​​of the seat-transmitted vibration disturbance component, the estimated value of the seat damper nonlinear disturbance component, and the estimated value of the unmodeled dynamic disturbance component of the seat-human coupling system.

[0017] The estimated values ​​of the vibration disturbance components transmitted by the seat, the nonlinear disturbance components of the seat damper, and the unmodeled dynamic disturbance components of the seat-human coupling system are used as the final decoupled output of the physical information neural network extended state observer.

[0018] In some embodiments, the first set of filter weights, the second set of filter weights, and the third set of filter weights are implemented through a physical information-guided training method, including:

[0019] Construct a training dataset for a physical information neural network extended state observer. The training dataset contains real-time sensor data time series under different operating conditions and corresponding true total disturbance labels.

[0020] Initialize the encoder module, the first feature decoupling branch, the second feature decoupling branch, the third feature decoupling branch, and the network parameters of each fully connected regression layer;

[0021] The training data is input into the physical information neural network extended state observer for forward propagation to obtain the estimated values ​​of the vibration disturbance components transmitted by the sample seat, the estimated values ​​of the nonlinear disturbance components of the sample seat damper, and the estimated values ​​of the unmodeled dynamic disturbance components of the sample seat-human coupling system.

[0022] The composite loss function is calculated. The composite loss function is composed of a weighted sum of a data fitting loss term, a physical equation constraint loss term, and a decoupling regularization loss term. The data fitting loss term constrains the sum of the estimated values ​​of the sample seat transmitted vibration disturbance component, the sample seat damper nonlinear disturbance component, and the sample seat-human coupling system unmodeled dynamic disturbance component to be consistent with the true total disturbance label. The physical equation constraint loss term forces the calculation of the equation residuals obtained by substituting the estimated values ​​of the sample seat transmitted vibration disturbance component, the sample seat damper nonlinear disturbance component, and the sample seat-human coupling system unmodeled dynamic disturbance component into the vehicle seat-human coupling dynamic equation. The decoupling regularization loss term constrains the correlation between the estimated values ​​of the sample seat transmitted vibration disturbance component, the sample seat damper nonlinear disturbance component, and the sample seat-human coupling system unmodeled dynamic disturbance component.

[0023] Based on the composite loss function value, the network parameters of the encoder module, the first feature decoupling branch, the second feature decoupling branch, the third feature decoupling branch, and each fully connected regression layer are updated through the error backpropagation algorithm until the network converges.

[0024] After training convergence, the network parameters corresponding to the feature extraction layer in the first feature decoupling branch, the second feature decoupling branch, and the third feature decoupling branch are fixed as the first set of filter weights, the second set of filter weights, and the third set of filter weights, respectively.

[0025] In some embodiments, based on the seat-transmitted vibration disturbance component, the seat damper nonlinear disturbance component, the unmodeled dynamic disturbance component of the seat-human coupling system, and real-time sensing data, a real-time control current for the seat damper is generated by a linear active disturbance rejection controller to perform vibration suppression, including:

[0026] The total disturbance of the system is estimated by summing the vibration disturbance component transmitted by the seat, the nonlinear disturbance component of the seat damper, and the unmodeled dynamic disturbance component of the seat-human coupling system.

[0027] The total disturbance estimate of the system, the seat vertical acceleration and seat telescopic displacement in the real-time sensor data, and the preset seat vertical acceleration reference value are all input into the linear state error feedback law of the linear active disturbance rejection controller.

[0028] In the linear state error feedback law, the preliminary control quantity is calculated based on the difference between the seat vertical acceleration and the reference value of the seat vertical acceleration, the seat extension displacement and its derivative, and in combination with the preset controller gain parameters.

[0029] The total control quantity is obtained by superimposing the initial control quantity with the estimated total system disturbance after feedforward compensation.

[0030] Based on the current-force mapping relationship between the total control quantity and the seat damper, the real-time control current of the seat damper is calculated and output.

[0031] In some embodiments, in the linear state error feedback law, based on the difference between the seat vertical acceleration and the reference value of the seat vertical acceleration, the seat extension displacement and its derivative, and in conjunction with preset controller gain parameters, a preliminary control quantity is calculated, including:

[0032] The difference between the seat vertical acceleration and the reference value of the seat vertical acceleration is taken as the first tracking error;

[0033] The derivative of the seat's telescopic displacement is used as the second tracking error;

[0034] The seat extension and retraction displacement is used as the third tracking error;

[0035] The first tracking error, the second tracking error, and the third tracking error are multiplied by the proportional gain, the derivative gain, and the displacement feedback gain in the preset controller gain parameters, respectively.

[0036] Sum the three results after multiplication to obtain the preliminary control quantity;

[0037] Based on the current-force mapping relationship between the total control quantity and the seat damper, the real-time control current of the seat damper is calculated and output, including:

[0038] The total control quantity is used as the desired output damping force of the seat damper;

[0039] Based on the current-force mapping relationship of the seat damper, the desired output damping force is solved by inverse mapping to obtain the corresponding desired control current value.

[0040] The desired control current value is limited to keep it within the safe current range of the seat damper actuator.

[0041] The desired control current value after limiting is used as the real-time control current output to the driver of the seat damper.

[0042] In some embodiments, the current seat vibration comfort performance index is calculated based on real-time sensing data, and combined with the current operating condition feature vector, an event triggering rule is used to determine whether to initiate the parameter optimization process, including:

[0043] Based on the seat vertical acceleration in the real-time sensing data within a preset time window, the current seat vibration comfort performance index is calculated. The current seat vibration comfort performance index is the root mean square value of the seat vertical acceleration.

[0044] The current seat vibration comfort performance index is compared with the dynamic performance threshold, which is adaptively determined based on the historical best performance index and the current operating condition feature vector.

[0045] The similarity between the current operating condition feature vector and the operating condition feature vector recorded during the last start of the parameter optimization process is calculated to obtain the operating condition drift degree.

[0046] Determine whether the current seat vibration comfort performance index is worse than the dynamic performance threshold, or whether the operating condition drift exceeds the preset operating condition drift threshold;

[0047] If the current seat vibration comfort performance index is worse than the dynamic performance threshold, or the operating condition drift exceeds the operating condition drift threshold, then the optimization conditions are met, and the parameter optimization process is triggered.

[0048] In some embodiments, when the event triggering rule determines that the optimization conditions are met, the meta-learning optimizer is invoked. The meta-learning optimizer outputs the initial hyperparameters of the projection iterative optimization algorithm based on the current operating condition feature vector, including:

[0049] Input the current working condition feature vector into the pre-trained meta-learning optimizer;

[0050] The forward propagation of the meta-learning optimizer outputs the initial hyperparameters of the projection iterative optimization algorithm. The initial hyperparameters include the center and covariance of the population initialization distribution, the initial step size factor of each projection operator, and the initial scale parameter of the Lévy fly operator.

[0051] The initialization hyperparameters are loaded into the corresponding configuration variables of the projection iterative optimization algorithm to complete the initialization settings of the projection iterative optimization algorithm.

[0052] In some embodiments, the training process of the pre-trained meta-learning optimizer is as follows:

[0053] Collect a historical training task set containing feature vectors of various vehicle seat operating conditions and corresponding optimal controller parameter sets;

[0054] Initialize the network parameters of the meta-learning optimizer;

[0055] A batch of training tasks are sampled from the historical training task set. For each training task, the working condition feature vector is input into the meta-learning optimizer to obtain the corresponding sample initialization hyperparameters.

[0056] Based on the sample initialization hyperparameters, the projection iterative optimization algorithm is initialized and run for a fixed number of iterations to obtain the optimized sample controller parameter set.

[0057] Calculate the performance loss of the optimized sample controller parameter set on the vehicle seat-human coupled real system or high-fidelity simulation model corresponding to the training task.

[0058] Based on the performance loss of each training task, the network parameters of the meta-learning optimizer are updated using the meta-learning backpropagation algorithm.

[0059] Repeat the above sampling, optimization, and update process until the meta-learning optimizer converges, resulting in a trained meta-learning optimizer.

[0060] In some embodiments, initialization hyperparameters are loaded into the corresponding configuration variables of the projection iterative optimization algorithm to complete the initialization settings of the projection iterative optimization algorithm, including:

[0061] The center of the population initialization distribution in the initialization hyperparameters is assigned to the population center vector of the projection iterative optimization algorithm;

[0062] The covariance of the population initialization distribution in the initialization hyperparameters is assigned to the population covariance matrix of the projection iterative optimization algorithm.

[0063] The initial step size factors of each projection operator in the initialization hyperparameters are assigned to the step size control variables of the corresponding gradient projection operator, random projection operator, and Lévy flight projection operator in the projection iterative optimization algorithm, respectively.

[0064] The initial scale parameter of the Lévy flight operator in the initialization hyperparameters is assigned to the scale control variable of the Lévy flight projection operator in the projection iterative optimization algorithm.

[0065] In some embodiments, based on the initialized hyperparameters, a manifold projection iterative optimization algorithm is executed within the parameter subspace defined by the vehicle seat-human coupled dynamics manifold model to search for and update the control parameter set of the linear active disturbance rejection controller and the network weight fine-tuning of the physical information neural network extended state observer, including:

[0066] Based on the center and covariance of the population initialization distribution in the initialization hyperparameters, an initial candidate solution population is generated in the joint parameter space composed of the control parameter set of the linear active disturbance rejection controller and the network weight fine-tuning of the physical information neural network extended state observer.

[0067] Each candidate solution in the initial candidate solution population is input into the encoder of the vehicle seat-human body coupled dynamics manifold model and mapped to a low-dimensional manifold space to obtain the corresponding manifold coordinates.

[0068] In the low-dimensional manifold space, based on the initial step size factor of each projection operator in the initialization hyperparameters, an iterative projection update operation is performed on the manifold coordinates. The iterative projection update operation includes manifold gradient-based projection, random direction-based projection, and long-range exploration projection based on the Lévy fly operator.

[0069] The manifold coordinates obtained after each iteration are reconstructed back into the joint parameter space through the decoder of the vehicle seat-human body coupled dynamics manifold model to obtain the updated candidate solution;

[0070] Using the performance predictor built into the vehicle seat-human body coupled dynamics manifold model, the updated candidate solutions are quickly evaluated to obtain the predicted performance index.

[0071] Based on the predicted performance index, the candidate solution with the best performance is selected in the joint parameter space. The subset of control parameters and the set of network weight fine-tuning quantum contained in the candidate solution with the best performance are used as the control parameter set of the linear active disturbance rejection controller to be updated and the network weight fine-tuning amount of the physical information neural network extended state observer, respectively.

[0072] Compared with existing technologies, the present invention, employing the above technical solution, has the following advantages: The present invention provides an adaptive vibration suppression method for vehicle seats based on projection iterative disturbance rejection. After acquiring real-time sensor data and the current operating condition feature vector, it decouples and estimates three types of disturbance components—road excitation, seat damper nonlinearity, and unmodeled system dynamics—through a physical information neural network-extended state observer. Based on these three types of disturbance components and real-time sensor data, a linear active disturbance rejection controller generates a real-time control current to suppress vibration. According to the current seat vibration comfort performance index and the current operating condition feature vector triggering events, a meta-learning optimizer is invoked to output the initialization hyperparameters of the projection iterative optimization algorithm. Based on these hyperparameters, manifold projection iterative optimization is performed within the subspace defined by the vehicle seat-human body coupled dynamics manifold model, updating the controller parameter set and the observer network weight fine-tuning and deployment. Through event-triggered and meta-learning-guided online parameter optimization, the present invention achieves adaptive adjustment of the control system to changing operating conditions, improving the continuous optimality of vibration suppression. Attached Figure Description

[0073] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0074] Figure 1This is a schematic diagram of steps S101 to S107 of the method described in the specific implementation embodiment;

[0075] Figure 2 This is a schematic diagram of steps S201 to S204 of the method described in the specific implementation. Detailed Implementation

[0076] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be particularly noted that the following embodiments are for illustrative purposes only and do not limit the scope of the invention. Similarly, the following embodiments are only some, not all, embodiments of the present invention, and all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0077] Please see Figure 1 This embodiment provides an adaptive vibration suppression method for vehicle seats based on projection iteration disturbance rejection, including:

[0078] S101. Acquire real-time sensing data and current operating condition feature vector of the vehicle seat. The real-time sensing data includes seat vertical acceleration, seat telescopic displacement and seat damper drive current.

[0079] S102. Input the real-time sensing data into the physical information neural network extended state observer to decouple and estimate the vibration disturbance component transmitted by the seat, the nonlinear disturbance component of the seat damper, and the unmodeled dynamic disturbance component of the seat-human coupling system.

[0080] S103. Based on the seat-transmitted vibration disturbance component, the seat damper nonlinear disturbance component, the unmodeled dynamic disturbance component of the seat-human coupling system, and real-time sensing data, a real-time control current for the seat damper is generated through a linear active disturbance rejection controller to perform vibration suppression.

[0081] S104. Calculate the current seat vibration comfort performance index based on real-time sensor data, and combine it with the current working condition feature vector to determine whether to start the parameter optimization process through event triggering rules.

[0082] S105. When the event triggering rule determines that the optimization conditions are met, the meta-learning optimizer is called. The meta-learning optimizer outputs the initial hyperparameters of the projection iterative optimization algorithm based on the current working condition feature vector.

[0083] S106. Based on the initial hyperparameters, within the parameter subspace defined by the vehicle seat-human body coupled dynamics manifold model, execute the manifold projection iterative optimization algorithm to search and update the control parameter set of the linear active disturbance rejection controller and the network weight fine-tuning of the physical information neural network extended state observer.

[0084] S107. Deploy the control parameter set and network weight fine-tuning amount to the linear active disturbance rejection controller and the physical information neural network extended state observer respectively to complete the adaptive adjustment.

[0085] In step S101, real-time sensing data is obtained directly through sensor measurements. The seat vertical acceleration reflects the vertical vibration intensity of the seat body, the seat telescopic displacement characterizes the real-time deformation of the vehicle seat suspension system, and the seat damper drive current indicates the magnitude of the control command currently applied to the seat damper. The current operating condition feature vector describes the operating environment state of the system. Its components can be obtained through the onboard information system or statistical analysis of the sensing data, and may include, for example, seat load, occupant posture parameters, vehicle ride comfort level, and seat adjustment position. These data are collected synchronously and form a time series, providing real-time input for subsequent control decisions. This step achieves the synchronous acquisition of key states and external environmental information of the vehicle seat vibration control system.

[0086] In step S102, the physical information neural network extended state observer is a hybrid model that embeds prior knowledge of physical dynamics into a neural network structure. Essentially, it takes real-time sensor data as input, processes it internally, and outputs independent estimates of three different types of disturbances. The seat-transmitted vibration disturbance component originates from the vibration energy transmitted from external vibrations to the seat through the vehicle body; the nonlinear disturbance component of the seat damper originates from the complex rheological properties and hysteresis effect of the magnetorheological fluid inside the seat damper; and the unmodeled dynamic disturbance component of the seat-human coupling system originates from higher-order dynamic characteristics that were ignored or simplified during the modeling of the seat-human coupling system. This observer, through decoupling estimation, separates the originally coupled total disturbance into components with different physical sources, thereby providing the controller with more refined disturbance information. This step, through a combination of data-driven and physical model approaches, achieves effective separation and observation of complex disturbances in the seat-human coupling system.

[0087] In step S103, the linear active disturbance rejection controller generates a control signal based on disturbance estimation and state feedback. It receives the estimated values ​​of the three disturbance components and real-time sensor data. The controller's internal processing involves synthesizing the disturbance information to form a total disturbance estimate for compensation, and simultaneously calculating the control error based on the deviation between the sensor data and the seat's vertical acceleration reference value. Using its internal algorithm, the controller combines the control error and the total disturbance estimate to calculate the ideal control force required to suppress vibration. Subsequently, based on the inherent characteristics of the seat damper, it maps this ideal control force into a corresponding real-time current command and outputs it. By applying this current, the damping force of the seat damper is actively adjusted to counteract the vibration. This step utilizes a strategy combining feedforward compensation and feedback control to effectively and actively suppress seat vibration, ensuring occupant comfort.

[0088] In step S104, the current seat vibration comfort performance index is a quantitative value used to evaluate the effectiveness of the control. Its calculation is based on real-time sensor data, such as the statistical characteristics of the seat's vertical acceleration. The event triggering rule is a set of preset logical judgment criteria that considers both the current performance index and the current operating condition feature vector. This rule only determines that the subsequent parameter optimization process needs to be initiated when the control performance shows a downward trend or the operating conditions change significantly, thus avoiding unnecessary optimization calculations when the system is running smoothly. This step, by introducing a condition- and state-based triggering mechanism, achieves adaptive start and stop of the optimization process, balancing computational load and control performance.

[0089] In step S105, the meta-learning optimizer is a model capable of learning from experience. This optimizer is activated when an event-triggered rule determines that optimization is needed. It takes the current operating condition feature vector as input and, through its internally learned mapping relationships, outputs a set of initialization configuration parameters—i.e., initialization hyperparameters—tailored to the subsequent projection iterative optimization algorithm. These hyperparameters aim to provide the optimization algorithm with a starting point close to the optimal region, thereby accelerating its search process. This step improves the efficiency and effectiveness of online parameter adjustment by utilizing historical optimization experience to guide the current optimization initialization.

[0090] In step S106, the vehicle seat-human coupling dynamics manifold model is a mathematical model that reveals the inherent low-dimensional structure of the high-dimensional parameter space, and its core adapts to the dynamic characteristics of the seat-human coupling system. Initializing hyperparameters provides initial guidance information such as direction and step size for the search within this low-dimensional manifold space. The manifold projection iterative optimization algorithm performs the search within this low-dimensional parameter subspace. This process can be understood as projecting the high-dimensional parameters to be optimized onto the low-dimensional manifold, iteratively updating in the low-dimensional space to find points with better seat vibration suppression performance and better fit for occupant comfort, and then mapping these points back to the original high-dimensional space to obtain new parameter candidates. Through repeated iterative evaluation and updates, the parameter modification amount for performance improvement is finally obtained, including adjustments to controller parameters and observer network weights. This step significantly reduces the computational complexity of online parameter optimization through a dimensionality reduction optimization strategy.

[0091] In step S107, the deployment process refers to applying the latest optimized parameters to the actual operating controller and observer. Specifically, the searched set of control parameters is updated to the corresponding parameter positions of the linear active disturbance rejection controller, while the network weight fine-tuning is superimposed on the existing weights of the physical information neural network extended state observer. After this operation, the vehicle seat vibration control system operates with the updated parameters, thereby adapting to new performance requirements or environmental conditions. This step realizes the closed-loop application of optimization results to actual control, completing the self-adjustment of the vehicle seat vibration control system.

[0092] This embodiment constructs a complete adaptive control loop that includes fine-grained disturbance observation, real-time vibration suppression, intelligent optimization triggering, efficient parameter search, and online deployment. This method not only improves the accuracy of disturbance estimation in the seat-human coupling system by decoupling observations and ensures real-time suppression using active disturbance rejection control, but more importantly, it introduces an intelligent parameter adjustment mechanism consisting of event triggering, meta-learning guidance, and manifold projection optimization. This mechanism enables the control system to autonomously perceive performance changes and operating condition shifts, and proactively and efficiently adjust its core parameters. This overcomes the inherent shortcomings of traditional methods, such as fixed parameters and difficulty in adapting to changing operating conditions, achieving continuous optimization and stable maintenance of the vehicle seat's vibration suppression performance across the entire operating range.

[0093] Please see Figure 2 In some embodiments, real-time sensing data is input to a physical information neural network extended state observer to decouple and estimate the seat-transmitted vibration disturbance component, the seat damper nonlinear disturbance component, and the unmodeled dynamic disturbance component of the seat-human coupling system, including:

[0094] S201. Input the time window data composed of real-time sensing data and its historical sequence into the encoder module of the physical information neural network extended state observer to extract high-dimensional mixed perturbation features.

[0095] S202. Input the high-dimensional hybrid disturbance features into the parallel first feature decoupling branch, second feature decoupling branch and third feature decoupling branch. The first feature decoupling branch extracts the first feature related to the frequency domain characteristics of seat transmitted vibration through the first set of filter weights. The second feature decoupling branch extracts the second feature related to the hysteresis nonlinearity of seat damper through the second set of filter weights. The third feature decoupling branch extracts the remaining dynamic third feature of the seat-human coupling system that has not been modeled through the third set of filter weights.

[0096] S203. Input the first feature, the second feature and the third feature into the corresponding fully connected regression layer respectively to obtain the estimated values ​​of the seat-transmitted vibration disturbance components, the estimated values ​​of the nonlinear disturbance components of the seat damper and the estimated values ​​of the unmodeled dynamic disturbance components of the seat-human coupling system.

[0097] S204. The estimated values ​​of the vibration disturbance components transmitted by the seat, the nonlinear disturbance components of the seat damper, and the unmodeled dynamic disturbance components of the seat-human coupling system are used as the final decoupling output of the physical information neural network extended state observer.

[0098] In step S201, the time window data refers to a data sequence composed of real-time sensing data from the current moment and several consecutive sampling moments prior to it, arranged chronologically. The real-time sensing data includes the seat's vertical acceleration, seat telescopic displacement, and seat damper drive current. The length of this window can be preset according to the dynamic characteristics of the seat-human coupling system, for example, including data from the past 0.5 seconds. The encoder module is a sub-network in the physical information neural network extended state observer, typically composed of several fully connected or convolutional layers. Its function is to perform nonlinear transformation and feature abstraction on the input time window data, thereby extracting high-dimensional hybrid perturbation features contained in the time-series data. This feature is a fusion vector containing all perturbation information, including seat-transmitted vibration, seat damper nonlinearity, and unmodeled dynamics of the seat-human coupling system. This step, by introducing time window data and using the encoder for feature extraction, provides a rich feature representation containing temporal dependencies for subsequent decoupling operations.

[0099] In step S202, the first feature decoupling branch, the second feature decoupling branch, and the third feature decoupling branch are three parallel neural network sub-modules with identical structures but independent parameters. The first set of filter weights, the second set of filter weights, and the third set of filter weights are the core parameter matrices used for feature transformation within these three branches. These weights are fixed after the observer training is completed and are each optimized into a filter sensitive to specific types of disturbances. Specifically, the first set of filter weights is trained to have bandpass or high-pass filtering characteristics to separate components related to the seat's transmitted vibration frequency band (usually corresponding to lower frequencies) from high-dimensional mixed disturbance features, forming the first feature; the second set of filter weights is trained to capture nonlinear patterns related to velocity and displacement history to extract the second feature reflecting the hysteresis loop characteristics of the seat damper; the third set of filter weights is used to extract the remaining dynamic information that cannot be covered by the first two types of features, i.e., the third feature related to the unmodeled dynamics of the seat-human coupling system. This step, through three parallel branches with dedicated filtering characteristics, achieves the initial separation of seat-related feature components from different physical sources from mixed features.

[0100] In step S203, the fully connected regression layer is a simple linear layer or shallow neural network. Each decoupled branch is connected to an independent regression layer, which maps the abstract features (first feature, second feature, and third feature) extracted from each branch back to the physical quantity space, i.e., outputs scalar or vector estimates of the corresponding disturbance components. For example, the fully connected regression layer connected to the decoupled branch of the first feature outputs the estimated value of the seat-transmitted vibration disturbance component, which represents the equivalent disturbance magnitude or force caused by external vibration transmission at the current moment. The other two regression layers are similar, outputting the estimated values ​​of the nonlinear disturbance component of the seat damper and the estimated values ​​of the unmodeled dynamic disturbance component of the seat-human coupling system, respectively. These estimates have clear physical meaning and dimensions. This step, through regression mapping, transforms the feature representation inside the neural network into disturbance observations that can be directly used for control calculations.

[0101] In step S204, the final decoupling output refers to the three explicit disturbance estimates obtained after the physical information neural network extended state observer completes one forward calculation: the estimated value of the seat-transmitted vibration disturbance component, the estimated value of the seat damper nonlinear disturbance component, and the estimated value of the unmodeled dynamic disturbance component of the seat-human coupling system. These three output values ​​are directly passed to the subsequent linear active disturbance rejection controller as precise input for disturbance compensation, helping the controller to accurately adjust the seat damper's damping force and suppress seat vibration. This step provides the decoupled disturbance state through the observer, providing a reliable input for subsequent precise control of seat vibration suppression.

[0102] This embodiment extracts temporal mixed features through an encoder, then uses three parallel and functionally specific decoupled branches with specific filter weights for feature separation. Finally, a regression layer outputs a quantized estimate, enabling the observer to adaptively learn and separate different physical disturbances coupled in the sensing signal in a data-driven manner. Its decoupling capability stems from the optimization of network weights during training. Compared to directly inputting the total disturbance into the controller, this refined decoupling provides the controller with clearer and more targeted seat-related disturbance information, making feedforward compensation more accurate. This lays a crucial foundation for improving the control accuracy and adaptability of the entire vehicle seat vibration suppression system, focusing entirely on the core objective of occupant comfort.

[0103] In some embodiments, the first set of filter weights, the second set of filter weights, and the third set of filter weights are implemented through a physical information-guided training method, including:

[0104] Construct a training dataset for a physical information neural network extended state observer. The training dataset contains real-time sensor data time series under different operating conditions and corresponding true total disturbance labels.

[0105] Initialize the encoder module, the first feature decoupling branch, the second feature decoupling branch, the third feature decoupling branch, and the network parameters of each fully connected regression layer;

[0106] The training data is input into the physical information neural network extended state observer for forward propagation to obtain the estimated values ​​of the vibration disturbance components transmitted by the sample seat, the estimated values ​​of the nonlinear disturbance components of the sample seat damper, and the estimated values ​​of the unmodeled dynamic disturbance components of the sample seat-human coupling system.

[0107] The composite loss function is calculated. The composite loss function is composed of a weighted sum of a data fitting loss term, a physical equation constraint loss term, and a decoupling regularization loss term. The data fitting loss term constrains the sum of the estimated values ​​of the sample seat transmitted vibration disturbance component, the sample seat damper nonlinear disturbance component, and the sample seat-human coupling system unmodeled dynamic disturbance component to be consistent with the true total disturbance label. The physical equation constraint loss term forces the calculation of the equation residuals obtained by substituting the estimated values ​​of the sample seat transmitted vibration disturbance component, the sample seat damper nonlinear disturbance component, and the sample seat-human coupling system unmodeled dynamic disturbance component into the vehicle seat-human coupling dynamic equation. The decoupling regularization loss term constrains the correlation between the estimated values ​​of the sample seat transmitted vibration disturbance component, the sample seat damper nonlinear disturbance component, and the sample seat-human coupling system unmodeled dynamic disturbance component.

[0108] Based on the composite loss function value, the network parameters of the encoder module, the first feature decoupling branch, the second feature decoupling branch, the third feature decoupling branch, and each fully connected regression layer are updated through the error backpropagation algorithm until the network converges.

[0109] After training convergence, the network parameters corresponding to the feature extraction layer in the first feature decoupling branch, the second feature decoupling branch, and the third feature decoupling branch are fixed as the first set of filter weights, the second set of filter weights, and the third set of filter weights, respectively.

[0110] In this embodiment, the physical information-guided training method is a machine learning approach that incorporates domain knowledge (physical laws) as constraints into the neural network training process. The training dataset can be constructed by collecting real-time sensor data (time series) of vehicle seats operating under different conditions (seat load, occupant posture parameters, vehicle ride comfort level, and seat adjustment position) in a high-fidelity simulation environment or real-vehicle tests. The true total disturbance label can be calculated by substituting measured data such as seat vertical acceleration and seat telescopic displacement into a known nominal dynamic model of the vehicle seat suspension system and solving the model's inverse problem. It represents the resultant force or equivalent acceleration of all disturbances experienced by the seat-human coupling system at the current moment. The purpose of constructing this dataset is to provide the network with the input-output sample pairs required for learning the disturbance decoupling mapping.

[0111] Initializing network parameters refers to assigning initial values ​​to all adjustable weights and biases in the encoder module, each feature decoupling branch, and the fully connected regression layer. Random initialization methods, such as Xavier initialization or He initialization, are usually used to ensure effective gradient propagation in the early stages of training.

[0112] The forward propagation process involves the training data sequentially passing through the encoder module, each feature decoupling branch, and its corresponding fully connected regression layer, performing operations such as matrix multiplication and activation function transformation in turn, ultimately outputting estimates of the three sample perturbation components. This process simulates the actual workflow of the observer after training.

[0113] In the composite loss function, the data fitting loss term typically uses the mean squared error function. Its role is to ensure that the sum of the three sample perturbation component estimates output by the network is as close as possible to the true total perturbation label, which is a basic requirement of supervised learning. The physics equation constraint loss term takes the three estimates output by the network as known quantities, substitutes them into the differential equations describing the motion of the vehicle seat suspension (e.g., second-order equations containing mass, spring, and damping terms), calculates the residuals on both sides of the equation, and minimizes the norm of these residuals. This constraint forces the decoupling results learned by the network to conform to basic physical laws, ensuring that the estimated perturbation components are dynamically self-consistent, which significantly improves the model's generalization ability and physical interpretability. The decoupling regularization loss term is used to encourage statistical independence among the three sample perturbation component estimates, for example, by minimizing their mutual information or correlation coefficients, thereby prompting each decoupling branch to focus on extracting different types of perturbation features and avoiding feature confusion. These three losses are balanced by a weighted sum, and the weight coefficients can be determined empirically or through cross-validation.

[0114] The backpropagation algorithm calculates the total error based on the composite loss function and uses the chain rule to calculate the gradient of each parameter with respect to the loss layer by layer from the output layer to the input layer. Then, gradient descent or its variants (such as the Adam optimizer) are used to update all network parameters according to the gradient direction. This process is repeated iteratively until the loss function value decreases to a stable range or the preset number of iterations is reached, at which point the network is considered to have converged.

[0115] After network convergence, the parameters of the network layers (i.e., feature extraction layers) in the first, second, and third feature decoupling branches responsible for extracting specific features from high-dimensional mixed features are no longer updated. These fixed parameters, having been optimized during training to address seat-transmitted vibrations, seat damper nonlinearity, and the unmodeled dynamic sensitivity of the seat-human coupling system, respectively, constitute the first, second, and third sets of filter weights used in subsequent embodiments.

[0116] This embodiment constructs a composite loss function that incorporates physical equation constraints and decoupling regularization. This training method not only requires the network output to be accurate in terms of data but also forces its decoupling results to conform to physical laws and ensure that each component is independent. This physically-guided training mechanism fundamentally ensures that each decoupling branch of the observer can learn feature extraction capabilities with clear physical meaning, resulting in reliable frequency or time-domain filtering characteristics for the final fixed filter weights. This embodiment overcomes the shortcomings of purely data-driven methods that may produce physically unreliable solutions or incomplete decoupling, providing a robust and interpretable disturbance observation basis for the entire vehicle seat adaptive vibration suppression scheme, which is a key step in ensuring the effectiveness of the method.

[0117] In some embodiments, based on the seat-transmitted vibration disturbance component, the seat damper nonlinear disturbance component, the unmodeled dynamic disturbance component of the seat-human coupling system, and real-time sensing data, a real-time control current for the seat damper is generated by a linear active disturbance rejection controller to perform vibration suppression, including:

[0118] The total disturbance of the system is estimated by summing the vibration disturbance component transmitted by the seat, the nonlinear disturbance component of the seat damper, and the unmodeled dynamic disturbance component of the seat-human coupling system.

[0119] The total disturbance estimate of the system, the seat vertical acceleration and seat telescopic displacement in the real-time sensor data, and the preset seat vertical acceleration reference value are all input into the linear state error feedback law of the linear active disturbance rejection controller.

[0120] In the linear state error feedback law, the preliminary control quantity is calculated based on the difference between the seat vertical acceleration and the reference value of the seat vertical acceleration, the seat extension displacement and its derivative, and in combination with the preset controller gain parameters.

[0121] The total control quantity is obtained by superimposing the initial control quantity with the estimated total system disturbance after feedforward compensation.

[0122] Based on the current-force mapping relationship between the total control quantity and the seat damper, the real-time control current of the seat damper is calculated and output.

[0123] In this embodiment, the estimated total system disturbance is a scalar or vector obtained by directly adding the three decoupled disturbance components. It represents the total disturbance of the seat-human coupling system that the linear active disturbance rejection controller needs to compensate for. The summation operation mathematically achieves a unified quantification of multi-source disturbances, providing a single input for feedforward compensation.

[0124] The linear state error feedback law is the core algorithm module of the linear active disturbance rejection controller. A preset reference value for the seat's vertical acceleration is typically set to zero, representing the ideal state of a stationary, vibration-free seat, aligning with the core objective of occupant comfort. This feedback law receives the estimated total system disturbance, real-time seat vertical acceleration, real-time seat extension / retraction displacement, and the reference value as inputs. The real-time seat extension / retraction displacement is the extension / retraction deformation of the vehicle seat suspension system. Internally, it first calculates the deviation between the measured seat vertical acceleration and the reference value, as the primary control error. Simultaneously, the seat extension / retraction displacement and its derivative (i.e., the suspension system's velocity) are also introduced as information reflecting the system state. The preset controller gain parameters are a set of pre-defined coefficients, including proportional gain and derivative gain. Based on a specific control algorithm (e.g., proportional-derivative control), the feedback law multiplies these state errors by the corresponding gain parameters and sums them to calculate the initial control quantity aimed at eliminating errors, stabilizing the vehicle seat vibration control system, and improving occupant comfort. This process achieves feedback adjustment based on the deviation of the current seat state from the desired comfort state.

[0125] Feedforward compensation involves multiplying the estimated total system disturbance by a feedforward gain coefficient, or directly passing it through a dynamic compensator. The goal is to generate a control component equal in magnitude but opposite in direction to the disturbance, which is then used to cancel it out before it affects the system output. Superimposing this processed value with the initial control input yields the final total control input acting on the vehicle seat vibration control system. This combination of feedforward and feedback allows the controller to simultaneously handle both known (estimated) disturbances and unknown model errors.

[0126] The current-force mapping relationship of the seat damper describes the static or dynamic correspondence between the input current and the output damping force. This relationship can be obtained by fitting experimental data of the seat damper's characteristics, and is usually expressed as a nonlinear function or a lookup table. The solution process involves using this mapping relationship to back-calculate the input current value that generates the desired damping force based on the total control quantity. This current value is the real-time control current. The output stage sends this current command to the power driver of the seat damper.

[0127] In this embodiment, the linear active disturbance rejection controller generates a control current using decoupled seat-related disturbance information. By summing and then feeding forward compensation, it fully utilizes the fine disturbance information provided by the observer. Through linear state error feedback, it ensures the stability and dynamic performance of the vehicle seat vibration control system, ensuring that seat vibration remains within a comfortable range. Finally, it outputs executable control commands through a clear physical mapping relationship, adapting to the actual driving requirements of the seat damper. This design makes the controller structure clear and the parameter tuning relatively simple. At the same time, by integrating feedforward and feedback, it significantly enhances the ability to suppress complex seat-related disturbances and the robustness of the system. It is a key execution link for achieving efficient vibration suppression of vehicle seats and ensuring occupant comfort.

[0128] In some embodiments, in the linear state error feedback law, based on the difference between the seat vertical acceleration and the reference value of the seat vertical acceleration, the seat extension displacement and its derivative, and in conjunction with preset controller gain parameters, a preliminary control quantity is calculated, including:

[0129] The difference between the seat vertical acceleration and the reference value of the seat vertical acceleration is taken as the first tracking error;

[0130] The derivative of the seat's telescopic displacement is used as the second tracking error;

[0131] The seat extension and retraction displacement is used as the third tracking error;

[0132] The first tracking error, the second tracking error, and the third tracking error are multiplied by the proportional gain, the derivative gain, and the displacement feedback gain in the preset controller gain parameters, respectively.

[0133] Sum the three results after multiplication to obtain the preliminary control quantity;

[0134] Based on the current-force mapping relationship between the total control quantity and the seat damper, the real-time control current of the seat damper is calculated and output, including:

[0135] The total control quantity is used as the desired output damping force of the seat damper;

[0136] Based on the current-force mapping relationship of the seat damper, the desired output damping force is solved by inverse mapping to obtain the corresponding desired control current value.

[0137] The desired control current value is limited to keep it within the safe current range of the seat damper actuator.

[0138] The desired control current value after limiting is used as the real-time control current output to the driver of the seat damper.

[0139] In this embodiment, the first tracking error, the second tracking error, and the third tracking error quantify the deviation of the seat's vertical acceleration, the vehicle seat suspension's velocity, and the seat's telescopic displacement from their desired state, respectively. The differential of the seat's telescopic displacement, i.e., the vehicle seat suspension's velocity, can be calculated using a displacement sensor or directly obtained from a velocity sensor. Both the displacement sensor and the velocity sensor are supporting sensing components of the vehicle seat vibration control system. These error terms collectively constitute a multidimensional error vector reflecting the dynamic state of the seat-human coupling system, accurately capturing various deviations in seat vibration from a comfortable state.

[0140] The preset controller gain parameters, including proportional gain, derivative gain, and displacement feedback gain, are a set of pre-defined positive coefficients. Their values ​​are adapted to the dynamic requirements of vehicle seat vibration suppression and align with occupant comfort goals. The proportional gain acts on the first tracking error (acceleration error), generating a control force proportional to the error magnitude to quickly reduce it. The derivative gain acts on the second tracking error (velocity error), providing damping to suppress oscillations in the vehicle seat suspension, improving the stability of the seat-human coupling system, and preventing exacerbated vibrations from affecting comfort. The displacement feedback gain acts on the third tracking error (displacement error), providing additional positional restoring force to help maintain the vehicle seat suspension near its equilibrium position, ensuring seat posture stability. Multiplying each error term by its corresponding gain essentially assigns different control weights to seat vibration-related errors with different physical meanings. Summing the weighted results yields a comprehensive preliminary control quantity, which physically represents a desired primary force aimed at simultaneously correcting deviations in seat acceleration, suspension velocity, and displacement, achieving precise seat vibration suppression.

[0141] The desired output damping force is the force that the seat damper is expected to generate in real time, as specified by the total control quantity of the controller. The current-force mapping relationship of the seat damper is typically a nonlinear function, potentially including hysteresis. This relationship can be obtained through experimental fitting of the seat damper's characteristics, closely reflecting its inherent operating characteristics. The inverse mapping solution involves using this relationship to inversely solve for the required input current from the known desired output damping force. In practice, this can be achieved through interpolation using a pre-established lookup table reflecting the current-force correspondence, or by calculation using the fitted inverse model function.

[0142] Limiting is used to ensure the physical feasibility of control commands and the safety of the vehicle seat vibration control system, preventing damage to seat components, impacting vibration suppression, and affecting occupant comfort due to abnormal current. The seat damper driver has a maximum permissible output current, determined by the driver hardware specifications and adapted to the rated operating parameters of the seat damper. Limiting compares the calculated desired control current value with this safe current range. If it exceeds the upper limit, the upper limit value is used; if it falls below the lower limit, the lower limit value is used. This ensures that the final output current command always remains within the safe operating range of the driver, guaranteeing stable and reliable operation of the seat damper.

[0143] This embodiment reveals the composition of the error term and the physical role of the gain in the linear state error feedback law, clarifies the conversion and safety protection process from the total control quantity to the final drive current, and focuses entirely on the core objectives of vehicle seat vibration suppression and occupant comfort. By defining three tracking errors with clear physical meanings and configuring their gains, the controller can finely adjust multiple dynamic dimensions of the seat-human coupling system to adapt to the seat's vibration suppression requirements under all operating conditions. Combined with precise inverse mapping and necessary amplitude limiting protection, the control commands are ensured to be both accurate and effective, as well as safe and reliable. This robustly converts the output of the control algorithm into the physical control quantity that actually acts on the vehicle seat suspension device, providing an indispensable detail for achieving high-performance vibration suppression of vehicle seats and ensuring occupant comfort.

[0144] In some embodiments, the current seat vibration comfort performance index is calculated based on real-time sensing data, and combined with the current operating condition feature vector, an event triggering rule is used to determine whether to initiate the parameter optimization process, including:

[0145] Based on the seat vertical acceleration in the real-time sensing data within a preset time window, the current seat vibration comfort performance index is calculated. The current seat vibration comfort performance index is the root mean square value of the seat vertical acceleration.

[0146] The current seat vibration comfort performance index is compared with the dynamic performance threshold, which is adaptively determined based on the historical best performance index and the current operating condition feature vector.

[0147] The similarity between the current operating condition feature vector and the operating condition feature vector recorded during the last start of the parameter optimization process is calculated to obtain the operating condition drift degree.

[0148] Determine whether the current seat vibration comfort performance index is worse than the dynamic performance threshold, or whether the operating condition drift exceeds the preset operating condition drift threshold;

[0149] If the current seat vibration comfort performance index is worse than the dynamic performance threshold, or the operating condition drift exceeds the operating condition drift threshold, then the optimization conditions are met, and the parameter optimization process is triggered.

[0150] In this embodiment, the length of the preset time window can be determined based on the sampling frequency and main frequency components of the vibration of the vehicle seat vibration control system. For example, it can be set to include hundreds of sampling points to cover several vibration cycles. The root mean square value of the seat vertical acceleration is calculated by squaring all acceleration sampling values ​​within the window, averaging them, and finally taking the square root. This value can effectively characterize the average level of vibration intensity over a period of time and is a key quantitative indicator for measuring occupant comfort.

[0151] The dynamic performance threshold is not a fixed value; its adaptive determination process can be based on historical operating data of the vehicle seat vibration control system. For example, the historical best performance index could be the best root mean square value ever achieved under similar current operating conditions (represented by the operating condition feature vector). The dynamic performance threshold can be set as this historical best value multiplied by a relaxation coefficient greater than 1, or by adding a fixed margin to it. This mechanism allows performance requirements to be dynamically adjusted according to operating conditions and the system's historical performance, avoiding false triggering caused by a single fixed threshold being too stringent or too lenient.

[0152] Operating condition drift is used to quantify the degree of difference between the current operating environment and the environment at the time of the last optimization. Similarity calculation can use metrics such as Euclidean distance and cosine similarity to calculate a scalar value from the feature vectors of two operating conditions. The operating condition drift threshold is a preset constant representing the maximum allowable change in operating conditions; exceeding this threshold is considered a significant change in operating conditions.

[0153] The judgment logic employs an "OR" relationship, meaning that optimization is deemed necessary if either a decrease in control performance or a change in the operating environment is met. Performance falling below a threshold directly reflects a deterioration in control effectiveness, while operating condition drift anticipates the risk of potential performance degradation under current parameters. This dual judgment criterion balances actual performance with potential risks, ensuring that optimization triggers are both timely and forward-looking.

[0154] This embodiment constructs an intelligent and robust event-triggered rule by specifically defining the calculation method of performance indicators, the adaptive mechanism of dynamic thresholds, and the quantification method of operating condition drift. This rule not only relies on real-time performance feedback but also introduces early warnings of operating condition changes, ensuring that the parameter optimization process is only initiated when control performance genuinely needs improvement or when a significant shift in operating conditions has occurred. This maintains the high vibration suppression performance of vehicle seats and ensures passenger comfort while minimizing unnecessary optimization computational overhead, achieving an efficient balance between computational resources and vibration suppression effectiveness.

[0155] In some embodiments, when the event triggering rule determines that the optimization conditions are met, the meta-learning optimizer is invoked. The meta-learning optimizer outputs the initial hyperparameters of the projection iterative optimization algorithm based on the current operating condition feature vector, including:

[0156] Input the current working condition feature vector into the pre-trained meta-learning optimizer;

[0157] The forward propagation of the meta-learning optimizer outputs the initial hyperparameters of the projection iterative optimization algorithm. The initial hyperparameters include the center and covariance of the population initialization distribution, the initial step size factor of each projection operator, and the initial scale parameter of the Lévy fly operator.

[0158] The initialization hyperparameters are loaded into the corresponding configuration variables of the projection iterative optimization algorithm to complete the initialization settings of the projection iterative optimization algorithm.

[0159] In this embodiment, the pre-trained meta-learning optimizer is a pre-trained neural network model trained to learn the mapping relationship between the vehicle seat operating condition features and the optimal initial configuration of the optimization algorithm. This optimizer is invoked after an event is triggered, receiving the current operating condition feature vector as input.

[0160] Forward propagation refers to the process where the input vector sequentially passes through each network layer of the optimizer, undergoing linear transformations and nonlinear activations, ultimately yielding a set of values ​​at the output layer. These values ​​constitute the initialization hyperparameters. The center and covariance of the initial population distribution determine the approximate location and distribution range of the initial candidate solution population seat-related parameter space (linear active disturbance rejection controller control parameters, physical information neural network extended state observer network weights) in the subsequent manifold projection iterative optimization algorithm. The initial step size factors of each projection operator control the initial step size for different search directions such as gradient projection and random projection. The initial scale parameter of the Lévy flight operator affects the jump range of its long-range exploration. These hyperparameters collectively set the starting point and initial search behavior for the optimization algorithm, and their quality directly affects the optimization efficiency.

[0161] The loading process assigns the values ​​of the initial hyperparameters output above to predefined variables within the projection iterative optimization algorithm that control population initialization, step size, and scale. For example, the center value is assigned to the variable representing the average position of the population, and the covariance matrix is ​​assigned to the variable representing the distribution shape of the population. After the assignment is completed, the projection iterative optimization algorithm has an initial configuration customized for the current working condition and can immediately begin efficient parameter search.

[0162] In this embodiment, the meta-learning optimizer transforms abstract seat condition information into a set of operable, high-quality optimization algorithm initialization parameters within the vehicle seat adaptive vibration suppression adjustment framework. By providing intelligent initialization that matches the seat condition, subsequent manifold projection iterative optimization can start searching from a region closer to the optimal solution, thereby significantly reducing the number of iterations and computation time required to achieve excellent seat vibration suppression performance. This is a key step in realizing rapid online adaptive adjustment of the entire vehicle seat vibration control system, effectively improving the real-time performance and practicality of the parameter optimization process, ensuring that the system can quickly adapt to changing seat conditions and continuously maintain occupant comfort.

[0163] In some embodiments, the training process of the pre-trained meta-learning optimizer is as follows:

[0164] Collect a historical training task set containing feature vectors of various vehicle seat operating conditions and corresponding optimal controller parameter sets;

[0165] Initialize the network parameters of the meta-learning optimizer;

[0166] A batch of training tasks are sampled from the historical training task set. For each training task, the working condition feature vector is input into the meta-learning optimizer to obtain the corresponding sample initialization hyperparameters.

[0167] Based on the sample initialization hyperparameters, the projection iterative optimization algorithm is initialized and run for a fixed number of iterations to obtain the optimized sample controller parameter set.

[0168] Calculate the performance loss of the optimized sample controller parameter set on the vehicle seat-human coupled real system or high-fidelity simulation model corresponding to the training task.

[0169] Based on the performance loss of each training task, the network parameters of the meta-learning optimizer are updated using the meta-learning backpropagation algorithm.

[0170] Repeat the above sampling, optimization, and update process until the meta-learning optimizer converges, resulting in a trained meta-learning optimizer.

[0171] In this embodiment, the historical training task set serves as the data foundation for training the meta-learning optimizer. This set is constructed by collecting a large amount of historical operating data from the vehicle seat vibration control system. Each training task includes a specific operating condition feature vector (such as a specific vehicle speed and load combination) and the optimal set of controller parameters obtained under that operating condition through offline fine-tuning or expert experience. These optimal parameter sets serve as the target reference for training.

[0172] Initializing network parameters involves assigning initial values ​​to the weights and biases within the meta-learning optimizer, typically using a random initialization method suitable for deep learning.

[0173] After sampling a batch of training tasks, for each task, the meta-learning optimizer outputs sample initialization hyperparameters in the manner described above. Subsequently, these hyperparameters are used to configure the projection iterative optimization algorithm, and starting from a random or default starting point, a fixed number of iterations (e.g., 50) are performed. This optimization process is conducted in a high-fidelity simulation environment of vehicle seat-human coupling, with the goal of adjusting controller parameters to optimize a certain performance metric. After a fixed number of iterations, a set of optimized sample controller parameters is obtained.

[0174] The performance loss is used to evaluate the quality of this set of parameters. During calculation, these parameters are substituted into a high-fidelity simulation model of the vehicle-seat-human coupled real system corresponding to the training task. The simulation is run, and performance metrics such as the root mean square value of the seat's vertical acceleration are calculated. This metric is compared with the ideal value (or the known best performance for the task), and the difference or related metric constitutes the performance loss. The smaller the loss, the better the results obtained by the initial hyperparameters provided by the meta-learning optimizer guiding subsequent optimization.

[0175] The performance loss calculated by the meta-learning backpropagation algorithm is not directly used to update the parameters within the projection iterative optimization algorithm. Instead, through a chain rule, the gradient of the loss with respect to the final optimized controller parameters is backpropagated to the initial hyperparameters of the samples that influence the starting point of the optimization, and then further propagated to the network parameters of the meta-learning optimizer that generates these hyperparameters. In this way, the parameters of the meta-learning optimizer are updated so that when faced with similar conditions in the future, it can output initial hyperparameters that guide the projection iterative optimization algorithm to find a better solution more quickly.

[0176] Repeat batch sampling, simulation optimization, loss calculation, and parameter updates until the performance of the meta-learning optimizer stabilizes on the validation set, at which point it is considered converged. The trained meta-learning optimizer then possesses the ability to quickly infer high-quality initialization hyperparameters based on the characteristics of new operating conditions.

[0177] In this embodiment, the meta-learning optimizer is trained by simulating a "rapid adaptation" process. Its training objective is not to directly predict the optimal parameters, but rather to learn and predict a set of initialization configurations that enable subsequent optimization algorithms to work efficiently. This training method allows the meta-learning optimizer to extract cross-task commonalities from historical tasks, thus providing an intelligent starting point for optimization when facing new but similar operating conditions. This ensures that in actual online applications, the initialization hyperparameters output by the meta-learning optimizer are high-quality parameters validated by extensive historical experience. This provides a reliable knowledge base for the entire vehicle seat adaptive vibration suppression system to quickly and accurately adjust parameters, ensuring that the system can quickly adapt to changing seat operating conditions and continuously maintain excellent seat vibration suppression and occupant comfort.

[0178] In some embodiments, initialization hyperparameters are loaded into the corresponding configuration variables of the projection iterative optimization algorithm to complete the initialization settings of the projection iterative optimization algorithm, including:

[0179] The center of the population initialization distribution in the initialization hyperparameters is assigned to the population center vector of the projection iterative optimization algorithm;

[0180] The covariance of the population initialization distribution in the initialization hyperparameters is assigned to the population covariance matrix of the projection iterative optimization algorithm.

[0181] The initial step size factors of each projection operator in the initialization hyperparameters are assigned to the step size control variables of the corresponding gradient projection operator, random projection operator, and Lévy flight projection operator in the projection iterative optimization algorithm, respectively.

[0182] The initial scale parameter of the Lévy flight operator in the initialization hyperparameters is assigned to the scale control variable of the Lévy flight projection operator in the projection iterative optimization algorithm.

[0183] In this embodiment, the assignment operation is a key step in the program implementation. It writes the numerical values ​​output by the meta-learning optimizer into the storage space of the internal variables of the projection iterative optimization algorithm. The values ​​are the initialization hyperparameters adapted to the vehicle seat operating conditions, ensuring that the assignment operation fits the seat parameter optimization requirements. The population center vector is a multi-dimensional vector whose dimension is the same as the total number of parameters of the vehicle seat vibration control system to be optimized (linear active disturbance rejection controller control parameters, physical information neural network extended state observer network weights). After assignment, this vector defines the center point of the initial candidate solution population distribution in the parameter space. The population covariance matrix is ​​a symmetric positive definite matrix. After assignment, this matrix, together with the center vector, defines the dispersion range and correlation of each solution in the initial population around the center point. For example, a larger covariance value means a wider initial exploration range.

[0184] Gradient projection, random projection, and Lévy flight projection are three different strategies in projection iterative optimization algorithms used to update solutions on low-dimensional manifolds of vehicle seat-human coupled dynamics manifold models. The gradient projection operator tends to move along the direction of fastest performance improvement (the gradient direction); the random projection operator introduces random perturbations to explore local regions; and the Lévy flight projection operator simulates the Lévy flight process, allowing occasional large step jumps to explore regions far from the current point. Assigning the initial step size factors from the initialization hyperparameters to the step size control variables of these operators sets the initial movement step size for each search strategy; for example, the gradient projection step size factor controls the magnitude of movement along the gradient direction.

[0185] The scale parameter of the Lévy flight operator controls the distribution characteristics of its jump step size during random walks. A larger scale parameter usually means a greater probability of long-distance exploration. Assigning this initial scale parameter to the corresponding scale control variable sets the initial intensity of the Lévy flight exploration behavior, ensuring that the exploration range matches the parameter optimization requirements for vehicle seat vibration suppression.

[0186] This embodiment clearly demonstrates how the output of the meta-learning optimizer precisely configures the initial behavior of subsequent optimization algorithms by specifically illustrating the correspondence between each initialization hyperparameter and the key control variables within the projection iterative optimization algorithm. The entire process revolves around the core objective of adaptive vibration suppression for vehicle seats. This refined configuration enables the optimization algorithm to execute an efficient search strategy tailored to the current seat conditions from the very first generation, avoiding blind searches. This significantly improves the convergence speed and final performance of the entire online parameter optimization process, ensuring that the optimized parameters can quickly adapt to the seat conditions and improve occupant comfort. This is a crucial guarantee for achieving rapid adaptive adjustment of the vehicle seat vibration control system.

[0187] In some embodiments, based on the initialized hyperparameters, a manifold projection iterative optimization algorithm is executed within the parameter subspace defined by the vehicle seat-human coupled dynamics manifold model to search for and update the control parameter set of the linear active disturbance rejection controller and the network weight fine-tuning of the physical information neural network extended state observer, including:

[0188] Based on the center and covariance of the population initialization distribution in the initialization hyperparameters, an initial candidate solution population is generated in the joint parameter space composed of the control parameter set of the linear active disturbance rejection controller and the network weight fine-tuning of the physical information neural network extended state observer.

[0189] Each candidate solution in the initial candidate solution population is input into the encoder of the vehicle seat-human body coupled dynamics manifold model and mapped to a low-dimensional manifold space to obtain the corresponding manifold coordinates.

[0190] In the low-dimensional manifold space, based on the initial step size factor of each projection operator in the initialization hyperparameters, an iterative projection update operation is performed on the manifold coordinates. The iterative projection update operation includes manifold gradient-based projection, random direction-based projection, and long-range exploration projection based on the Lévy fly operator.

[0191] The manifold coordinates obtained after each iteration are reconstructed back into the joint parameter space through the decoder of the vehicle seat-human body coupled dynamics manifold model to obtain the updated candidate solution;

[0192] Using the performance predictor built into the vehicle seat-human body coupled dynamics manifold model, the updated candidate solutions are quickly evaluated to obtain the predicted performance index.

[0193] Based on the predicted performance index, the candidate solution with the best performance is selected in the joint parameter space. The subset of control parameters and the set of network weight fine-tuning quantum contained in the candidate solution with the best performance are used as the control parameter set of the linear active disturbance rejection controller to be updated and the network weight fine-tuning amount of the physical information neural network extended state observer, respectively.

[0194] In this embodiment, the joint parameter space is a high-dimensional Euclidean space whose dimension is equal to the sum of all adjustable parameters of the linear active disturbance rejection controller (e.g., proportional gain, differential gain, etc.) and the network weight parameters specified as online fine-tunable in the physical information neural network extended state observer. These parameters are all core adjustable parameters of the vehicle seat vibration control system. Each point in this space corresponds to a complete set of parameter configurations. The initial candidate solution population is a set of such points generated by random sampling from a multivariate normal distribution. The center and covariance of the initial population distribution determine the clustering location and dispersion pattern of these initial points in the space.

[0195] In this embodiment, the vehicle seat-human coupled dynamics manifold model is further concretized as a deep generative model with an encoder-decoder structure, such as a variational autoencoder. This model learns a bidirectional mapping from a high-dimensional joint parameter space to a low-dimensional latent space (i.e., manifold space) through offline training. Its training data comes from a large amount of parameter configurations and performance data obtained from historical operation or simulation of the vehicle seat vibration control system. The encoder's role is to compress a high-dimensional parameter vector (candidate solution) into a low-dimensional, dense manifold coordinate vector, which captures the essential characteristics and structure of the parameter configuration. The decoder's role is to reconstruct the manifold coordinate vector back to the original high-dimensional parameter space. The core assumption of this model is that all high-performance or feasible parameter configurations are not uniformly distributed in the high-dimensional space, but rather concentrated on one or more low-dimensional nonlinear manifold structures.

[0196] Iterative projection update operations performed within the low-dimensional manifold space are crucial for efficient searching in this continuous and smooth latent space. Manifold gradient-based projection utilizes gradient information provided by the performance predictor to calculate the direction of the fastest performance improvement within the manifold tangent space and move the coordinates along this direction. Random-direction-based projection randomly selects a unit vector within the tangent space of the current coordinates to explore the neighborhood of the current position. Long-range exploration projection based on the Lévy fly operator simulates a stochastic process with a heavy-tailed distribution, allowing the search coordinates to jump significantly with a low probability, thus enabling the exploration of other distant regions on the manifold and enhancing global search capabilities. The initial step size factor of each projection operator directly controls the step size of these movements.

[0197] The performance predictor is a regression network integrated into the vehicle seat-human body coupled dynamics manifold model. It accepts high-dimensional candidate solutions reconstructed by the decoder (or directly accepts their low-dimensional manifold coordinates) as input, and outputs a scalar value as a predicted performance metric, such as the predicted root mean square value of the seat vertical acceleration, after forward propagation. By undergoing supervised training on a dataset containing a large number of parameter-performance sample pairs, the predictor learns an approximation of the complex nonlinear mapping relationship between parameters and performance, thereby enabling rapid, batch performance evaluation of candidate solutions at a computational cost far lower than that of high-fidelity simulation. Preferably, the training dataset is constructed by testing a widely sampled set of controller parameters and observer network weights covering different operating conditions on a high-fidelity simulation model or bench test system of the vehicle seat suspension device, and recording the corresponding seat vibration comfort performance metrics (such as the root mean square value of the seat vertical acceleration), thus forming parameter combination-performance metric sample pairs.

[0198] The selection process is based on the predicted performance indices calculated by the performance predictor for all candidate solutions, with the core objective of optimizing occupant comfort. The algorithm compares the index values ​​corresponding to all candidate solutions and selects the candidate solution with the best predicted performance (i.e., the smallest index value, representing the best vibration suppression effect). Finally, from the high-dimensional vector of this optimal candidate solution, according to a predefined indexing rule, the portion corresponding to the controller parameters is separated as the set of control parameters to be updated, and the portion corresponding to the observer network weights is separated as the network weight fine-tuning amount.

[0199] In this embodiment, the manifold projection iterative optimization algorithm transforms the optimization problem from a high-dimensional space to a low-dimensional manifold by utilizing a pre-trained manifold model, significantly reducing the search dimensionality. On the low-dimensional manifold, a balance between exploration and extraction is achieved by combining gradient-guided local refinement search, random perturbation, and the Levy flight strategy with long-range exploration capabilities. The integrated lightweight performance predictor avoids costly real-world performance evaluation, making it possible to screen large-scale candidate solutions within a limited online computation time. This combined strategy enables the algorithm to quickly and accurately locate parameter update schemes that significantly improve vehicle seat vibration suppression performance and ensure occupant comfort, serving as the core and efficiency guarantee for the online adaptive adjustment of the vehicle seat vibration control system.

[0200] By adopting the above technical solutions, this invention differs from existing technologies and possesses the following beneficial effects: It refines and estimates three types of disturbances—road excitation, seat damper nonlinearity, and unmodeled system dynamics—by expanding the state observer through a physical information neural network. This provides more targeted feedforward compensation information for the linear active disturbance rejection controller, thereby improving the accuracy and robustness of vibration suppression. Furthermore, this method constructs a complete online adaptive parameter adjustment mechanism. It intelligently determines optimization timing through event-triggered rules and calls a meta-learning optimizer to provide high-quality initial hyperparameters for the manifold projection iterative optimization algorithm based on the current operating condition feature vector. This allows for efficient searching and updating of the controller and observer parameters within the low-dimensional parameter subspace defined by the vehicle seat-human body coupled dynamics manifold model. This mechanism enables the control system to autonomously perceive performance degradation and operating condition transitions, and proactively and rapidly optimize and deploy parameters, thus overcoming the performance degradation defects of traditional fixed-parameter active disturbance rejection control under varying operating conditions. The above technical solutions achieve continuous optimization and stable maintenance of the vehicle seat's vibration suppression performance across the entire operating range, significantly improving ride comfort and the overall adaptive capability of the system.

[0201] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0202] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0203] The above description is only a part of the embodiments of the present invention and does not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made based on the content of the present invention specification and drawings, or direct or indirect application in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. An adaptive vibration suppression method for vehicle seats based on projection iterative disturbance rejection, characterized in that, include: The system acquires real-time sensor data and current operating condition feature vectors of the vehicle seats. The real-time sensor data includes seat vertical acceleration, seat telescopic displacement, and seat damper drive current. The current operating condition feature vectors include vehicle seat load, occupant posture parameters, vehicle ride comfort level, and seat adjustment position. The real-time sensing data is input into the physical information neural network extended state observer to decouple and estimate the vibration disturbance component transmitted by the seat, the nonlinear disturbance component of the seat damper, and the unmodeled dynamic disturbance component of the seat-human coupling system. Based on the vibration disturbance component transmitted by the seat, the nonlinear disturbance component of the seat damper, the unmodeled dynamic disturbance component of the seat-human coupling system, and the real-time sensing data, a real-time control current for the seat damper is generated by a linear active disturbance rejection controller to perform vibration suppression. The current seat vibration comfort performance index is calculated based on the real-time sensing data, and combined with the current working condition feature vector, the parameter optimization process is determined by the event triggering rules. The performance index includes the root mean square value of the seat vertical weighted acceleration, the vibration amplitude in the human body sensitive frequency range, and the seat vibration decay rate. When the event triggering rule determines that the optimization conditions are met, the meta-learning optimizer is invoked. The meta-learning optimizer outputs the initialization hyperparameters of the projection iterative optimization algorithm based on the current working condition feature vector. Based on the initial hyperparameters, within the parameter subspace defined by the vehicle seat-human body coupled dynamics manifold model, a manifold projection iterative optimization algorithm is executed to search and update the control parameter set of the linear active disturbance rejection controller and the network weight fine-tuning of the physical information neural network extended state observer. The control parameter set and the network weight fine-tuning amount are respectively deployed to the linear active disturbance rejection controller and the physical information neural network extended state observer to complete the adaptive adjustment of the vehicle seat.

2. The adaptive vibration suppression method for vehicle seats based on projection iterative disturbance rejection according to claim 1, characterized in that, The real-time sensing data is input into a physical information neural network extended state observer to decouple and estimate the seat-transmitted vibration disturbance component, the seat damper nonlinear disturbance component, and the unmodeled dynamic disturbance component of the seat-human coupling system, including: The time window data, consisting of the real-time sensing data and its historical sequence, is input into the encoder module of the physical information neural network extended state observer to extract high-dimensional hybrid perturbation features. The high-dimensional hybrid perturbation features are input into the parallel first feature decoupling branch, the second feature decoupling branch, and the third feature decoupling branch. The first feature decoupling branch extracts the first feature related to the frequency domain characteristics of the seat-transmitted vibration through the first set of filter weights. The second feature decoupling branch extracts the second feature related to the hysteresis nonlinearity of the seat damper through the second set of filter weights. The third feature decoupling branch extracts the remaining dynamic third features of the seat-human coupling system that are not modeled through the third set of filter weights. The first feature, the second feature, and the third feature are respectively input into the corresponding fully connected regression layer to obtain the estimated values ​​of the seat-transmitted vibration disturbance components, the estimated values ​​of the seat damper nonlinear disturbance components, and the estimated values ​​of the unmodeled dynamic disturbance components of the seat-human coupling system. The estimated values ​​of the seat-transmitted vibration disturbance components, the estimated values ​​of the seat damper nonlinear disturbance components, and the estimated values ​​of the unmodeled dynamic disturbance components of the seat-human coupling system are used as the final decoupling output of the physical information neural network extended state observer.

3. The adaptive vibration suppression method for vehicle seats based on projection iterative disturbance rejection according to claim 2, characterized in that, The first set of filter weights, the second set of filter weights, and the third set of filter weights are implemented through a physical information-guided training method, including: A training dataset for the physical information neural network extended state observer is constructed. The training dataset contains the time series of the real-time sensing data under different operating conditions and the corresponding true total disturbance label. The real-time sensing data are seat vertical acceleration, seat telescopic displacement and seat damper drive current. The different operating conditions are the operating conditions corresponding to seat load, occupant sitting posture parameters, vehicle ride comfort level and seat adjustment position. Initialize the network parameters of the encoder module, the first feature decoupling branch, the second feature decoupling branch, the third feature decoupling branch, and each fully connected regression layer; The training dataset is input into the physical information neural network extended state observer for forward propagation to obtain the estimated values ​​of the vibration disturbance components transmitted by the sample seat, the estimated values ​​of the nonlinear disturbance components of the sample seat damper, and the estimated values ​​of the unmodeled dynamic disturbance components of the sample seat-human coupling system. The composite loss function value is calculated, which is composed of a weighted sum of a data fitting loss term, a physical equation constraint loss term, and a decoupling regularization loss term. The data fitting loss term constrains the sum of the estimated values ​​of the sample seat transmitted vibration disturbance components, the sample seat damper nonlinear disturbance components, and the sample seat-human coupling system unmodeled dynamic disturbance components to be consistent with the true total disturbance label. The physical equation constraint loss term forces the minimization of the residuals obtained after substituting the estimated values ​​of the sample seat transmitted vibration disturbance components, the sample seat damper nonlinear disturbance components, and the sample seat-human coupling system unmodeled dynamic disturbance components into the vehicle seat-human coupling dynamic equation. The decoupling regularization loss term constrains the correlation between the estimated values ​​of the sample seat transmitted vibration disturbance components, the sample seat damper nonlinear disturbance components, and the sample seat-human coupling system unmodeled dynamic disturbance components. Based on the composite loss function value, the network parameters of the encoder module, the first feature decoupling branch, the second feature decoupling branch, the third feature decoupling branch, and each fully connected regression layer are updated using the error backpropagation algorithm until the network converges. After training convergence, the network parameters corresponding to the feature extraction layer in the first feature decoupling branch, the second feature decoupling branch, and the third feature decoupling branch are fixed as the first set of filter weights, the second set of filter weights, and the third set of filter weights, respectively.

4. The adaptive vibration suppression method for vehicle seats based on projection iterative disturbance rejection according to claim 1, characterized in that, Based on the seat-transmitted vibration disturbance component, the seat damper nonlinear disturbance component, the unmodeled dynamic disturbance component of the seat-human coupling system, and the real-time sensing data, a real-time control current for the seat damper is generated through a linear active disturbance rejection controller to perform vibration suppression, including: The total system disturbance estimate is obtained by summing the vibration disturbance component transmitted by the seat, the nonlinear disturbance component of the seat damper, and the unmodeled dynamic disturbance component of the seat-human coupling system. The total disturbance estimate of the system, the seat vertical acceleration and seat telescopic displacement in the real-time sensing data, and the preset seat vertical acceleration reference value are all input into the linear state error feedback law of the linear active disturbance rejection controller. In the linear state error feedback law, the preliminary control quantity is calculated based on the difference between the vertical acceleration of the seat and the reference value of the vertical acceleration of the seat, the extension and retraction displacement of the seat and its derivative, and in combination with the preset controller gain parameters. The initial control quantity is superimposed with the estimated total system disturbance after feedforward compensation to obtain the total control quantity; Based on the total control quantity and the current-force mapping relationship of the seat damper, the real-time control current of the seat damper is calculated and output.

5. The adaptive vibration suppression method for vehicle seats based on projection iterative disturbance rejection according to claim 4, characterized in that, In the linear state error feedback law, based on the difference between the seat vertical acceleration and the reference value of the seat vertical acceleration, the seat extension / retraction displacement and its derivative, and in conjunction with preset controller gain parameters, the preliminary control quantity is calculated, including: The difference between the vertical acceleration of the seat and the reference value of the vertical acceleration of the seat is taken as the first tracking error; The derivative of the seat's telescopic displacement is taken as the second tracking error; The seat extension / retraction displacement is taken as the third tracking error; The first tracking error, the second tracking error, and the third tracking error are multiplied by the proportional gain, the derivative gain, and the displacement feedback gain in the preset controller gain parameters, respectively. The three results after multiplication are summed to obtain the preliminary control quantity; Based on the total control quantity and the current-force mapping relationship of the seat damper, the real-time control current of the seat damper is calculated and output, including: The total control quantity is used as the desired output damping force of the seat damper; Based on the current-force mapping relationship of the seat damper, the desired output damping force is solved by inverse mapping to obtain the corresponding desired control current value. The desired control current value is limited to ensure it is within the safe current range of the seat damper driver. The desired control current value after limiting is output as the real-time control current of the seat damper to the driver of the seat damper.

6. The adaptive vibration suppression method for vehicle seats based on projection iterative disturbance rejection according to claim 1, characterized in that, The current seat vibration comfort performance index is calculated based on the real-time sensor data, and combined with the current operating condition feature vector, the parameter optimization process is determined through event triggering rules, including: Based on the seat vertical acceleration in the real-time sensing data within a preset time window, the current seat vibration comfort performance index is calculated, where the current seat vibration comfort performance index is the root mean square value of the seat vertical acceleration. The current seat vibration comfort performance index is compared with the dynamic performance threshold, which is adaptively determined based on the historical best performance index and the current operating condition feature vector. The similarity between the current operating condition feature vector and the operating condition feature vector recorded during the previous parameter optimization process is calculated to obtain the operating condition drift degree. Determine whether the current seat vibration comfort performance index is worse than the dynamic performance threshold, or whether the operating condition drift exceeds the preset operating condition drift threshold. If the current seat vibration comfort performance index is worse than the dynamic performance threshold, or the operating condition drift exceeds the operating condition drift threshold, then the optimization conditions are met, and the parameter optimization process is triggered.

7. The adaptive vibration suppression method for vehicle seats based on projection iterative disturbance rejection according to claim 1, characterized in that, When the event triggering rule determines that the optimization conditions are met, the meta-learning optimizer is invoked. The meta-learning optimizer outputs the initialization hyperparameters of the projection iterative optimization algorithm based on the current working condition feature vector, including: The current working condition feature vector is input into a pre-trained meta-learning optimizer; Through the forward propagation of the meta-learning optimizer, the initialization hyperparameters of the projection iterative optimization algorithm are output. The initialization hyperparameters include the center and covariance of the population initialization distribution, the initial step size factor of each projection operator, and the initial scale parameter of the Lévy flight operator. The initialization hyperparameters are loaded into the corresponding configuration variables of the projection iterative optimization algorithm to complete the initialization settings of the projection iterative optimization algorithm.

8. The adaptive vibration suppression method for vehicle seats based on projection iterative disturbance rejection according to claim 7, characterized in that, The training process of the pre-trained meta-learning optimizer is as follows: Collect a historical training task set containing feature vectors of various vehicle seat operating conditions and corresponding optimal controller parameter sets; Initialize the network parameters of the meta-learning optimizer; A batch of training tasks are sampled from the historical training task set. For each training task, the current working condition feature vector is input into the meta-learning optimizer to obtain the corresponding sample initialization hyperparameters. Based on the sample initialization hyperparameters, the projection iterative optimization algorithm is initialized and run for a fixed number of iterations to obtain the optimized sample controller parameter set. Calculate the performance loss of the optimized sample controller parameter set on the vehicle seat-human coupled real system or high-fidelity simulation model corresponding to the training task; Based on the performance loss of each training task, the network parameters of the meta-learning optimizer are updated using the meta-learning backpropagation algorithm. Repeat the above sampling, optimization and update process until the meta-learning optimizer converges, and obtain the trained meta-learning optimizer.

9. The adaptive vibration suppression method for vehicle seats based on projection iterative disturbance rejection according to claim 7, characterized in that, The initialization hyperparameters are loaded into the corresponding configuration variables of the projection iterative optimization algorithm to complete the initialization settings of the projection iterative optimization algorithm, including: The center of the population initialization distribution in the initialization hyperparameters is assigned to the population center vector of the projection iterative optimization algorithm. The population center vector is initially adapted to the parameter optimization benchmark of the vehicle seat-human coupling system. Assign the covariance of the population initialization distribution in the initialization hyperparameters to the population covariance matrix of the projection iterative optimization algorithm; The initial step size factors of each projection operator in the initialization hyperparameters are assigned to the step size control variables of the corresponding gradient projection operator, random projection operator, and Lévy flight projection operator in the projection iterative optimization algorithm, respectively. The initial scale parameter of the Lévy flight operator in the initialization hyperparameters is assigned to the scale control variable of the Lévy flight projection operator in the projection iterative optimization algorithm.

10. The adaptive vibration suppression method for vehicle seats based on projection iterative disturbance rejection according to claim 1, characterized in that, Based on the initial hyperparameters, within the parameter subspace defined by the vehicle seat-human coupled dynamics manifold model, a manifold projection iterative optimization algorithm is executed to search for and update the control parameter set of the linear active disturbance rejection controller and the network weight fine-tuning of the physical information neural network extended state observer, including: Based on the center and covariance of the population initialization distribution in the initialization hyperparameters, an initial candidate solution population is generated in the joint parameter space formed by the control parameter set of the linear active disturbance rejection controller and the network weight fine-tuning of the physical information neural network extended state observer. Each candidate solution in the initial candidate solution population is input into the encoder of the vehicle seat-human body coupled dynamics manifold model and mapped to a low-dimensional manifold space to obtain the corresponding manifold coordinates; Within the low-dimensional manifold space, an iterative projection update operation is performed on the manifold coordinates according to the initial step size factor of each projection operator in the initialization hyperparameters. The iterative projection update operation includes manifold gradient-based projection, random direction-based projection, and long-range exploration projection based on the Lévy fly operator. The manifold coordinates obtained after each iteration are reconstructed back into the joint parameter space through the decoder of the vehicle seat-human body coupled dynamics manifold model to obtain the updated candidate solution; Using the performance predictor built into the vehicle seat-human body coupled dynamics manifold model, the updated candidate solution is quickly evaluated to obtain the predicted performance index. Based on the predicted performance index, the candidate solution with the best performance is selected in the joint parameter space. The subset of control parameters and the set of network weight fine-tuning quantum contained in the candidate solution with the best performance are used as the control parameter set of the linear active disturbance rejection controller to be updated and the network weight fine-tuning amount of the physical information neural network extended state observer, respectively.

Images