A shock absorber performance optimization control method based on model fusion

Abstract

The present application relates to the technical field of industrial mechanism model, specifically to a shock absorber performance optimization control method based on model fusion, comprising: extracting low-frequency disturbance state variables and inputting variable topology industrial mechanism model to generate nominal reference state trajectory; using a deep state observation network with fusion energy passivity constraint to calculate nonlinear model mismatch compensation and adaptive weighting parameters; based on the adaptive weighting parameters, dynamically fusing the nominal reference state trajectory and the compensation to generate a generalized state estimate; based on the generalized state estimate, performing dynamic multi-objective optimization to generate hybrid modal control input acting on the execution end. The present application combines the physical consistency, calculation real-time performance and robustness of the control process through adaptive fusion of mechanism model and data-driven method.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial mechanism modeling, in particular to a shock absorber performance optimization control method based on model fusion. BACKGROUND

[0002] Vehicle shock control is a key link for improving the driving quality and safety performance of the whole vehicle, which can relieve the impact of road excitation on the vehicle body by actively adjusting the mechanical properties of the actuator. Under complex and variable driving conditions, a high-performance control strategy not only needs to effectively isolate the vertical vibration caused by road roughness to ensure passenger comfort, but also needs to ensure the continuous and stable contact between the tire and the road to maintain the steering stability.

[0003] In the prior art, although the general industrial mechanism model based on the first principle has clear physical interpretability, it is often difficult to accurately characterize the dynamic characteristics of the variable topology object with multi-stage stiffness switching or sudden change of damping characteristics with a single model. If a high-order refined model containing complex fluid thermodynamics and friction characteristics is used, it will bring huge computational load, and it is difficult to meet the millisecond-level real-time control requirements in the vehicle embedded unit. On the other hand, although the data-driven modeling method can fit complex nonlinear characteristics through learning, it is easy to output false control signals that violate physical consistency when encountering untrained conditions or sensor noise interference, leading to divergence and instability of the closed-loop system. In addition, the existing model fusion technology usually adopts static weight distribution, lacks online quantitative evaluation of model cognitive uncertainty, and cannot dynamically adjust the reliability of mechanism models and data models according to the change of working conditions, making it difficult to generate robust generalized state estimation.

[0004] Therefore, a shock absorber performance optimization control method based on model fusion is proposed, which can effectively compensate for the structural errors of mechanism models under the premise of ensuring that the control output strictly complies with the physical passivity constraint, and realize robust fusion based on uncertainty quantification, which is a technical problem to be solved at present.

[0005] Therefore, a shock absorber performance optimization control method based on model fusion is proposed, which can effectively compensate for the structural errors of mechanism models under the premise of ensuring that the control output strictly complies with the physical passivity constraint, and realize robust fusion based on uncertainty quantification, which is a technical problem to be solved at present. SUMMARY

[0006] The present application aims to provide a model fusion-based shock absorber performance optimization control method, which establishes a physical flow field benchmark of a system by constructing a variable topology industrial mechanism model embedded with a dynamic evolution equation, generates a nonlinear model mismatch compensation quantity orthogonal to a nominal reference state trajectory by using a physical constraint state observation operation based on a passivity constraint, and performs dynamic state correction and multi-objective Pareto optimization in a state space according to an adaptive weighting parameter representing prediction accuracy, thereby solving the problems of excessive real-time calculation load of a high-order mechanism model, insufficient robustness under unknown working conditions caused by a lack of physical consistency of a pure data-driven model, and difficulty of a static control strategy in adapting to dynamic performance requirements of a variable structure nonlinear object.

[0007] To achieve the above object, the present application provides the following technical scheme.

[0008] A model fusion-based shock absorber performance optimization control method, comprising:

[0009] Obtaining real-time running feedback signals of a controlled object and performing signal feature extraction and reconstruction to extract low-frequency disturbance state variables; inputting the low-frequency disturbance state variables into a variable topology industrial mechanism model embedded with an object dynamic equation to generate a nominal reference state trajectory and a linearized state deviation;

[0010] Constructing a deep state observation network fused with an energy passivity constraint and inputting the linearized state deviation into the deep state observation network; using a physical punishment mechanism in the deep state observation network to solve a nonlinear model mismatch compensation quantity under the condition of meeting the energy passivity constraint; simultaneously calculating and outputting an adaptive weighting parameter based on the cognitive uncertainty characteristics of a distribution output by the deep state observation network; performing a dynamic state correction operation in a state space by taking the adaptive weighting parameter as an adjustment factor, and adding the nonlinear model mismatch compensation quantity to the nominal reference state trajectory as a residual term to generate a generalized state estimation value;

[0011] Constructing a dynamic multi-objective optimization functional with the generalized state estimation value as a variable, searching for a control working point meeting a Pareto optimal condition in an object feasible region defined by the variable topology industrial mechanism model, and decoupling and converting the control working point into a hybrid modal control input acting on an execution end.

[0012] Preferably, the extraction step of the low-frequency disturbance state variable comprises:

[0013] The real-time running feedback signal is acquired, a multi-scale signal decomposition algorithm is used to process the real-time running feedback signal, the real-time running feedback signal is decomposed into modal components of different frequency bandwidths, and time-frequency distribution characteristics of each modal component are calculated; based on the signal characteristics of each modal component, effective modal selection is performed, high-frequency random texture noise components are filtered out, and effective disturbance energy components containing transient impact characteristics are retained and reconstructed; the reconstructed effective disturbance energy characteristics are subjected to data dimension reduction processing, principal components whose cumulative variance contribution rate meets a preset condition are extracted, and the principal components are reconstructed into low-frequency disturbance state variables with a dimension consistent with an input space of the variable topology industrial mechanism model.

[0014] Preferably, the variable topology industrial mechanism model is a multi-modal coupled model constructed based on discrete-continuous hybrid dynamics theory, and specifically includes:

[0015] The variable topology industrial mechanism model is composed of a discrete event state set and a continuous time dynamics equation set; the discrete event state set represents a topology configuration of an execution end, and covers geometric configuration states corresponding to different stiffness levels of elastic elements and damping characteristic states corresponding to different flow resistance characteristics of damping elements; the continuous time dynamics equation set includes a group of object dynamics equations one-to-one mapped with the discrete event state set, and is used to describe continuous dynamic behaviors of a controlled object under a corresponding topology configuration.

[0016] Preferably, the object dynamics equation includes:

[0017] a nonlinear elastic state equation for establishing a functional mapping relationship between an elastic element stroke, a medium state and a nonlinear elastic restoring force; and a damping dynamics equation for establishing a piecewise hybrid dynamics function mapping relationship between a damping adjustment opening, a medium flow state and a differential pressure-flow characteristic, and containing smooth transition logic between different flow patterns.

[0018] Preferably, the generation step of the nominal reference state trajectory and the linearized state deviation includes:

[0019] A rate of change of the low-frequency disturbance state variable of the input variable topology industrial mechanism model is monitored, when a preset condition is met, a target topology configuration at a current time is locked in the discrete event state set, and a target object dynamics equation corresponding to the target topology configuration is activated from the continuous time dynamics equation set; a discrete numerical integration algorithm is used to iteratively solve the activated target object dynamics equation, and a dynamic state evolution curve in a prediction time domain is generated as a nominal reference state trajectory; at each discrete time node of the nominal reference state trajectory, a first-order partial derivative matrix of the object dynamics equation with respect to a model state variable is calculated analytically, and a tangent space is defined using the first-order partial derivative matrix; a difference vector between an actual observed state at the current time and a projection of the nominal reference state trajectory in the tangent space is calculated as a linearized state deviation.

[0020] Preferably, the step of generating the nonlinear model mismatch compensation quantity comprises:

[0021] The deep state observer network adopts a deep neural network architecture with time series memory capability to perform time series feature extraction on the input linearized state deviation sequence; a compound loss functional is constructed by fusing a physical penalty term, which is constructed based on the stability derivative of the controlled object energy function, to impose loss penalty on the network predicted output that violates the energy passivity condition, guiding the model parameters to converge to the energy dissipation region; a complementary projection operator is configured in the output stage to construct a residual subspace projection matrix that is complementary to the tangent direction of the nominal reference state trajectory, forcing the output vector to project into the residual subspace, generating a nonlinear model mismatch compensation quantity containing vertical direction residuals that are not captured by the variable topology industrial mechanism model.

[0022] Preferably, the step of generating the adaptive weighting parameter comprises:

[0023] A probability inference mechanism is enabled in the calculation process of the deep state observer network to perform multiple random mask sampling and forward propagation on the network operation weights; the distribution variance of the multiple forward propagation output results is calculated and mapped to a normalized cognitive uncertainty indicator to output the adaptive weighting parameter.

[0024] Preferably, the step of generating the generalized state estimation value comprises:

[0025] A dynamic gain matrix is constructed with the adaptive weighting parameter as the weighting factor to define an uncertainty-based statistical weighting space; the nonlinear model mismatch compensation quantity located in the tangent space is subjected to amplitude modulation operation using the dynamic gain matrix, and based on the linear superposition criterion, the modulated compensation quantity and the nominal reference state trajectory are subjected to weighted fusion calculation to obtain the modified state vector as the generalized state estimation value.

[0026] Preferably, the step of constructing a dynamic multi-objective optimization functional and searching for a control working point comprises:

[0027] An output response fluctuation functional representing disturbance suppression performance and a contact stability functional representing stability maintenance performance are defined; the generalized state estimation value is used to identify the current external working condition category, and the weight coefficients of the two functionals are dynamically adjusted accordingly to synthesize a dynamic multi-objective optimization evaluation functional; the saturation characteristics of the execution end in the variable topology industrial mechanism model are used as boundary constraints, and a nonlinear constraint optimization algorithm is used to iteratively optimize the dynamic multi-objective optimization evaluation functional within the object feasible region to solve the optimal state derivative vector located on the instantaneous Pareto frontier as the control working point.

[0028] Preferably, the step of decoupling the control operating point into hybrid modal control inputs comprises:

[0029] Based on the differential flatness principle, an inverse dynamics operation model describing the input-output relationship of the execution end is constructed, the optimal state derivative vector corresponding to the control operating point is decoupled and mapped into a generalized control force demand and a generalized stiffness demand; according to the generalized stiffness demand, the target parameters of the elastic element and the corresponding stiffness adjustment instructions are calculated and determined by querying the pre-stored elastic element characteristic data; according to the generalized control force demand and the fluid characteristics under the current damping mode, the target parameters of the damping element are calculated by inverse solution, and the target parameters are converted into modulation control instructions for driving the actuators to act; the stiffness adjustment instructions and the modulation control instructions are combined to form the hybrid modal control inputs.

[0030] Compared with the prior art, the present application has the following beneficial effects:

[0031] 1. The present application extracts low-frequency disturbance state variables and inputs them into the variable topology industrial mechanism model embedded in the object dynamics equation, realizes the decomposition of the complex multi-stage stiffness switching and damping nonlinear process into the mechanism dominant nominal trajectory and linearization deviation which can be calculated in real time. This method avoids the problem of excessive load of traditional high-order model operation, and solves the technical problem of insufficient physical representation accuracy of the simplified model when facing the structure switching of the controlled object through the variable topology mechanism, ensuring the physical reference function of the nominal reference state trajectory.

[0032] 2. The present application fuses the deep state observation network with energy passivity constraint, uses the physical punishment mechanism to force the network output to meet the energy dissipation condition, fundamentally eliminates the risk of false signals generated by pure data models that violate physical laws (such as energy spontaneous increase). At the same time, based on the adaptive weighting parameters calculated by probability inference, the cognitive uncertainty of the model to the current working condition can be quantified in real time, and the fusion weight of the nonlinear model mismatch compensation is dynamically adjusted, effectively preventing the control instability caused by blindly relying on the data model in unobserved working conditions.

[0033] 3. The present application constructs a dynamic multi-objective optimization functional based on the generalized state estimate value, which can dynamically adjust the weight of disturbance suppression and stability maintenance according to the working condition (such as transient impact or smooth driving), and search for the optimal solution on the Pareto frontier. Combined with the inverse dynamics model, the control operating point is decoupled into hybrid modal control inputs (such as stiffness adjustment and damping modulation) acting on the execution end, solving the problem that the traditional static control strategy cannot meet the multi-objective demand and the difficulty of multi-actuator (airbag and shock absorber) collaborative control.

[0034] 4. This invention constructs a collaborative closed-loop logic that "ensures the stability of basic physics through a mechanistic model, enhances the accuracy of local nonlinearity through a physical constraint data model, and adjusts the fusion weights using uncertainty parameters." Within the state space, adaptive weighted parameters are used to dynamically fuse the nominal trajectory of the variable topology mechanistic model with the nonlinear compensation of the deep observation network, generating a high-precision generalized state estimate. This fusion architecture enables the system to simultaneously achieve control accuracy approaching that of a high-order mechanistic model and physical stability approaching that of a pure mechanistic model on a computationally limited embedded platform, significantly improving the overall performance of the shock absorber control system under complex road conditions. Attached Figure Description

[0035] Fig. 1 This is a flowchart illustrating a shock absorber performance optimization control method based on model fusion proposed in this invention.

[0036] Fig. 2 This is a schematic diagram of the constraint state observation and generalized state estimation process proposed in this invention;

[0037] Fig. 3 This is a schematic diagram illustrating the specific process of generating dynamic multi-objective optimization decision-making and hybrid modal control proposed in this invention. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] Please see Figs. 1 to 3 This invention provides a shock absorber performance optimization control method based on model fusion, the technical solution of which is as follows:

[0040] The system acquires real-time operational feedback signals from the controlled object and performs signal feature extraction and reconstruction to extract low-frequency disturbance state variables. The low-frequency disturbance state variables are then input into a variable topology industrial mechanism model with embedded object dynamic equations to generate a nominal reference state trajectory and a linearized state deviation.

[0041] A deep state observation network with energy passivity constraint is constructed, and a linearized state deviation is input into the deep state observation network; a physical punishment mechanism in the deep state observation network is used to solve a nonlinear model mismatch compensation under the condition of meeting the energy passivity constraint; meanwhile, based on the cognitive uncertainty characteristics of the output distribution of the deep state observation network, an adaptive weighting parameter is calculated and output; the adaptive weighting parameter is used as an adjustment factor to perform a dynamic state correction operation in a state space, and the nonlinear model mismatch compensation is added as a residual term to a nominal reference state trajectory to generate a generalized state estimation value;

[0042] A dynamic multi-objective optimization functional is constructed with the generalized state estimation value as a variable, a control working point meeting the Pareto optimal condition is searched in an object feasible region defined by a variable topology industrial mechanism model, and the control working point is decoupled and converted into a mixed mode control input acting on an execution end.

[0043] The technical solutions of the present application will be further described in detail below in combination with specific embodiments.

[0044] Embodiment one:

[0045] The embodiment provides a shock absorber performance optimization control method based on model fusion, which is configured and run on a vehicle equipped with a four-corner independent electronic control pneumatic vibration isolation system and a continuous damping control semi-active support system as a controlled object, aiming to realize optimization control of vehicle body vertical vibration acceleration and tire dynamic load by adjusting the air chamber pressure of the elastic element (specifically an air spring) and the flow area of the damping element (specifically a shock absorber damping valve) in real time.

[0046] In the execution scenario set in this embodiment, the vehicle is driving at a constant speed of 60 km / h on a section of asphalt gravel test road, and the road section contains a trapezoidal deceleration strip with a height of 40 mm and a bottom width of 300 mm.

[0047] Further, the real-time running feedback signals of the controlled object are acquired and signal feature extraction and reconstruction are performed, and low-frequency disturbance state variables are extracted.

[0048] Specifically, first, the sensor values at the four wheels of the vehicle are read synchronously through the vehicle-mounted communication bus at a sampling frequency of 1000 Hz, and real-time running feedback signals including the sprung mass vertical acceleration signal, the relative motion stroke signal and the rotating part speed signal are established.

[0049] Specifically, the real-time running feedback signal is processed by using a multi-scale signal decomposition algorithm (for example, a variational mode decomposition VMD); in the operation configuration, the number of mode decomposition K is set to 5, and the original signal is decomposed into 5 mode components with limited bandwidth; then, the time-frequency distribution characteristics of each component are calculated, and according to the set 20Hz threshold, the high-frequency random noise component is removed, and only the low-frequency transient disturbance energy characteristics representing the road surface long-wave undulation and the speed bump impact are retained.

[0050] Specifically, the retained low-frequency transient disturbance energy characteristics are subjected to data dimension reduction processing (for example, principal component analysis PCA); the first three principal components with a cumulative variance contribution rate of 85% are selected; and the three principal components are reconstructed into low-frequency disturbance state variables with a dimension consistent with the input space of the subsequent mechanism model through inverse transformation operation, and the variables are determined as a three-dimensional vector including road surface vertical input displacement, road surface vertical input speed and under-spring mass vertical motion speed.

[0051] Further, the low-frequency disturbance state variables are input into the variable topology industrial mechanism model embedded with the object dynamics equation, and the nominal reference state trajectory and the linearized state deviation are generated through model evolution calculation.

[0052] Specifically, the pre-constructed variable topology industrial mechanism model is called, which embeds the object dynamics equation, specifically including a nonlinear elastic state equation (a multivariable process gas state equation) and a damping dynamics equation (a throttle orifice flow equation); wherein in the calculation of the elastic state, a multivariable index of 1.3 is used in combination with real-time environmental temperature correction to calculate the nonlinear elastic restoring force corresponding to the compression stroke of the gas chamber; in the calculation of the damping dynamics, the differential pressure-flow characteristics corresponding to the damping adjustment opening and the fluid flow state are calculated according to the segmented mixed fluid model.

[0053] Specifically, the input low-frequency disturbance state variables are monitored in real time during the model running process, and when it is detected that the road surface vertical input speed exceeds 0.5m / s, it is determined that the state transition condition is met; at this time, the target topology configuration of the current time in the discrete event state set of the model is locked as the compression damping dominant mode, and the object dynamics equation corresponding to the mode is activated from the continuous time dynamics equation set.

[0054] Specifically, a discretized numerical integration algorithm (for example, a fourth-order Runge-Kutta method) is used to iteratively solve the activated object dynamics equation, to deduce the controlled object dynamics state in a future 50ms prediction time domain, and to generate a nominal reference state trajectory; the difference between the actual observed state vector at the current time and the numerical value at the corresponding time in the nominal reference state trajectory is calculated, and a first-order linearization processing (for example, calculating the Jacobian matrix) is performed on the difference to obtain a linearized state deviation.

[0055] Further, a deep state observer network integrating energy passivity constraint is constructed, and the linearized state deviation is input into the deep state observer network; a physical penalty mechanism is used in the deep state observer network to solve the nonlinear model mismatch compensation under the condition of satisfying the energy passivity constraint; and the adaptive weighting parameter is calculated and output based on the cognitive uncertainty characteristics of the output distribution of the deep state observer network.

[0056] Specifically, the deep state observer network adopts a deep neural network architecture with time series memory capability (e.g., long short-term memory network LSTM) to perform time series feature extraction on the input linearized state deviation sequence; a compound loss functional integrating a physical penalty term is constructed, and the physical penalty term is constructed based on the stability derivative of the energy function of the controlled object; the specific calculation logic of the physical penalty term is as follows: the total energy Hamilton function of the controlled object is constructed, and the derivative of the Hamilton function with respect to time is calculated; when the calculated energy change rate violates the passivity condition of energy dissipation (i.e., the energy of the controlled object does not naturally increase), a high penalty weight is applied in the loss function, so as to constrain the energy dissipation direction of the calculated gradient to satisfy the energy passivity condition and prevent the divergence of the calculation result.

[0057] Specifically, a complementary projection operator is configured in the output stage; first, the tangent vector of the nominal reference state trajectory at the current time is calculated, and a complementary residual subspace orthogonal to the tangent vector is constructed; then, the output vector is forced to project into the complementary residual subspace, and the nonlinear model mismatch compensation quantity containing the vertical direction residual which is not captured by the variable topology industrial mechanism model is generated. This step ensures that the generated compensation quantity only contains the dynamic residual (such as complex friction torque, nonlinear leakage, etc.) perpendicular to the prediction direction of the mechanism model, avoiding the interference of the data-driven model on the part accurately predicted by the mechanism model.

[0058] Specifically, a probability inference calculation mechanism (e.g., Monte Carlo dropout method) is enabled in the calculation process of the deep state observer network. The random dropout rate is set to 0.3, the network operation weight is randomly masked and sampled multiple times (e.g., 20 times), and forward propagation is performed; the distribution variance of the output results of multiple forward propagations is counted, and the distribution variance is mapped to a normalized cognitive uncertainty index (e.g., the normalized value is 0.2); the adaptive weighting parameter negatively correlated with the cognitive uncertainty index (in this calculation, the parameter value is determined as 0.67) is output, so as to quantify the credibility of the current data-driven model.

[0059] Further, the nonlinear model mismatch compensation quantity is superimposed on the nominal reference state trajectory by using the adaptive weighting parameter to calculate and generate the generalized state estimation value.

[0060] Specifically, the dynamic gain matrix is constructed with the calculated adaptive weighting parameter (0.67) to define the statistical weighting space based on uncertainty; the amplitude modulation operation is performed on the nonlinear model mismatch compensation quantity by using the dynamic gain matrix, and the weighted fusion calculation is performed on the modulated compensation quantity and the nominal reference state trajectory based on the linear superposition criterion. This process is mathematically equivalent to dynamically adjusting the fusion ratio of the mechanism model prediction value and the data-driven compensation value according to the current cognitive uncertainty, thereby obtaining a more robust generalized state estimation value.

[0061] Further, a dynamic multi-objective optimization functional is constructed with the generalized state estimation value as a variable, and a control working point satisfying the Pareto optimal condition is searched in the object feasible region defined by the variable topology industrial mechanism model, and the control working point is decoupled and converted into a hybrid modal control input acting on the execution end.

[0062] Specifically, a dynamic multi-objective optimization evaluation functional is constructed, which is composed of an output response fluctuation functional (specifically, a frequency domain weighted vertical acceleration functional) representing disturbance suppression performance and a contact stability functional (specifically, a tire dynamic load functional) representing stability maintenance performance; according to the generalized state estimation value, it is identified that the current is in the deceleration zone climbing condition, and the weight coefficient is dynamically adjusted (for example, the contact stability weight is set to 0.7, and the output response fluctuation weight is set to 0.3).

[0063] Specifically, the boundary constraint conditions of the object feasible region (solenoid current 0-2.0A, air chamber pressure 5-12bar) are set; a nonlinear constraint optimization algorithm (such as sequential quadratic programming SQP) is used to optimize the evaluation functional in the object feasible region, and the optimal state derivative vector located on the instantaneous Pareto frontier is solved as the control working point.

[0064] Further, based on the differential flatness principle, an inverse dynamics operation model describing the input-output relationship of the execution end is constructed, and the optimal state derivative vector corresponding to the control working point is decoupled and mapped into a generalized control force demand and a generalized stiffness demand; according to the generalized stiffness demand, it is determined by table lookup calculation that the target parameter of the elastic element (air chamber target pressure) needs to be increased by 0.5bar, and the intake air solenoid needs to be opened for 120ms; according to the generalized control force demand, the target parameter of the damping element (damping valve target flow resistance) is calculated by combining the current fluid characteristics inverse solution, and it is converted into the corresponding modulation control instruction (solenoid drive current 1.6A and pulse width modulation duty cycle 65%); finally, the stiffness adjustment instruction (intake valve opening logic instruction) and the modulation control instruction (PWM duty cycle instruction) are generated synchronously to form a hybrid modal control input, which drives the actuator to act.

[0065] Embodiment two:

[0066] This embodiment details the specific execution steps of high-precision signal reconstruction and variable topology mechanism modeling in a shock absorber performance optimization control method based on model fusion; the method is configured and run on a specific controlled object (i.e., a vehicle equipped with a four-corner independent electronically controlled pneumatic vibration isolation system and a continuous damping control semi-active support system).

[0067] In the execution scenario set in this embodiment, the vehicle travels at a constant speed of 60 km / h on a section of asphalt gravel test road surface, which includes a trapezoidal speed bump with a height of 40 mm and a bottom width of 300 mm.

[0068] Furthermore, regarding the extraction of low-frequency disturbance state variables, the method includes: synchronously acquiring the vertical acceleration signal of the sprung mass of the controlled object (vehicle), the relative motion stroke signal, and the rotational speed signal of the rotating parts to construct a real-time operation feedback signal; and using a multi-scale signal decomposition algorithm to process the real-time operation feedback signal within the current moment and the past preset time window.

[0069] Specifically, sensor values ​​at the four wheels of the controlled object are synchronously acquired at a sampling frequency of 1000Hz to construct a real-time operation feedback signal matrix that includes the vertical acceleration of the sprung mass, the relative motion stroke, and the rotational speed of the rotating parts.

[0070] Specifically, multi-scale signal decomposition (e.g., using variational mode decomposition algorithm) is performed on the real-time operation feedback signal matrix. In a preferred parameter configuration, the number of mode decompositions K is set to 5. The physical basis for selecting this value is that the vibration energy of the controlled object in the vertical direction is mainly distributed in five main frequency bands: the rigid body mode of the vehicle body, the damping support system mode, the engine idling coupled vibration, the macroscopic undulation of the road surface, and the microscopic texture of the road surface. Setting the corresponding number of decomposition layers can ensure that the algorithm can effectively separate the above physical frequency bands and avoid mode aliasing. Setting an appropriate bandwidth limiting parameter (e.g., a quadratic penalty factor α=2000) aims to balance the bandwidth constraints of each modal component and prevent signal distortion caused by excessively narrow bandwidth or noise residue caused by excessively wide bandwidth.

[0071] Furthermore, the time-frequency distribution characteristics of each modal component are calculated; high-frequency random noise components (represented as high-frequency texture noise in this embodiment) are filtered out based on a preset frequency threshold, and low-frequency transient disturbance energy characteristics (represented as low-frequency transient impact energy in this embodiment) are extracted.

[0072] Specifically, the modal components obtained by decomposition are analyzed one by one to construct analytic signals (for example, by Hilbert transform), and the instantaneous frequency and instantaneous amplitude are calculated; the threshold of the cutoff frequency is set to 20 Hz, and the reason for selecting 20 Hz as the threshold is that the human body is mainly sensitive to vertical vibration in the interval of 4 Hz to 8 Hz, and vibration signals above 20 Hz are mainly excited by the rough texture of the asphalt pavement, which belongs to the category of high-frequency random noise and is not the target frequency band of active posture control.

[0073] Specifically, the time-frequency energy spectrum obtained by traversal calculation is processed to perform effective bandwidth truncation. The effective control bandwidth threshold is set to 20 Hz, and the physical basis for selecting this threshold is that the physical action response cutoff frequency of the controlled object actuator (air spring gas circuit and shock absorber electromagnetic valve) is about 20 Hz, and the main energy of the deceleration strip impact causing the vehicle body to move greatly is concentrated in the low-frequency interval of the vehicle body mode and the wheel bounce mode; the high-frequency components above 20 Hz are mainly road texture noise, which not only has no contribution to vehicle body posture control, but also causes ineffective high-frequency vibration of the actuator.

[0074] Specifically, the components with a frequency coordinate greater than 20 Hz are considered as uncontrolled texture noise and are suppressed, and only the effective control energy in the interval of 0 Hz to 20 Hz is retained; this processing step not only ensures the matching of the control command and the response capability of the actuator, but also effectively removes the high-frequency random noise caused by the gravel pavement and retains the main wave energy characteristics caused by the deceleration strip impact.

[0075] Further, data dimension reduction processing is performed on the low-frequency transient disturbance energy characteristics, the principal components whose cumulative variance contribution rate meets the preset condition are extracted, and a low-frequency disturbance state variable is reconstructed.

[0076] Specifically, principal component analysis (PCA) is performed on the filtered energy feature matrix; the cumulative variance contribution rate threshold (for example, 85%) is set, and the reason for selecting this threshold is that in the analysis of object dynamics signals, the first 85% of the energy usually contains the main motion modes of the vehicle body and the wheels, and the remaining part is usually measurement noise or secondary nonlinear coupling terms.

[0077] Specifically, the calculation result shows that the cumulative variance contribution rate of the first three principal components meets the requirement, so the first three principal components are selected for inverse transformation and reconstruction; the reconstructed output low-frequency disturbance state variable is determined as a three-dimensional vector, and the three components of the vector correspond to the road vertical input displacement, the road vertical input velocity, and the sprung mass vertical motion velocity, respectively, which are directly used as input variables of the subsequent mechanism model.

[0078] Further, a variable topology industrial mechanism model is constructed, which is defined as a hybrid dynamic mathematical description containing a discrete event state set and a continuous time dynamic equation set; the discrete event state set represents the topology configuration of the execution end.

[0079] Specifically, the variable topology industrial mechanism model is constructed in a hierarchical architecture, the top layer is a topology decision layer, which is used to determine the current discrete event state according to the input signal; the middle layer is a dynamics evolution layer, which contains continuous-time differential equation sets under different topological structures; and the bottom layer is a parameter mapping layer, which describes the nonlinear mapping of physical parameters with the change of state.

[0080] Specifically, for the combined actuator using a damping element (continuous damping control shock absorber) and an elastic element (air spring), a discrete event state set is defined to contain three mutually exclusive topological states:

[0081] Compression inflation state: corresponding to the physical configuration that the elastic element inlet valve is opened and the damping element compression cavity flow valve is opened; at this time, the control logic is in the variable stiffness and compression damping cooperative working mode.

[0082] Recovery deflation state: corresponding to the physical configuration that the elastic element exhaust valve is opened and the damping element recovery cavity flow valve is opened; at this time, the control logic is in the variable stiffness and recovery damping cooperative working mode.

[0083] Pressure holding locking state: corresponding to the physical configuration that all air valves are closed and the fluid is in quasi-static state; at this time, the control logic is in the constant stiffness and basic damping working mode.

[0084] The input variables of the topology decision layer include the external excitation signal provided by the low-frequency disturbance state variable, specifically the road vertical input velocity, and the internal state feedback composed of the current air chamber pressure and the current relative stroke. Using a logic threshold in the calculation process, the layer generates an impact detection flag bit as an intermediate variable, which is in an effective state when the absolute value of the road vertical input velocity exceeds 0.5 m / s and the relative stroke is less than 10 mm; at the same time, the flow direction determination flag bit is determined according to the difference between the current pressure and the supply pressure. The output variable of the layer is the topology mode index, which outputs the integer index corresponding to the compression inflation state, the recovery deflation state and the pressure holding locking state respectively.

[0085] The input variables of the parameter mapping layer include the topology mode index from the topology decision layer, the relative motion state including the relative displacement and the relative velocity, and the internal bristle state from the last time dynamic friction model (LuGre model). The intermediate variables calculated by the layer through interpolation table lookup include the Reynolds number calculated based on the relative velocity and the viscosity of hydraulic oil, which is used to determine the laminar or turbulent state; also includes the effective cross-sectional area of the air spring obtained based on the relative stroke table, and the instantaneous dynamic friction coefficient calculated based on the internal bristle state and the relative velocity. The output variable of the layer is the instantaneous physical parameter set, including the instantaneous equivalent stiffness, the instantaneous equivalent damping coefficient and the nonlinear friction force.

[0086] The input variables of the dynamic evolution layer include the instantaneous physical parameter set from the parameter mapping layer, the system full state vector at the last time, and the control input at the current time. The intermediate variables calculated by the layer include the resultant force vector, i.e., the algebraic sum of the elastic force, the damping force, the friction force, and the inertial force; and the state derivatives calculated according to the second Newton's law and the differential form of the gas state equation, such as the acceleration and the pressure rate of change. The final output variables solved by the layer through numerical integration include the nominal reference state trajectory in the future prediction time domain obtained through numerical integration, and the linearized Jacobian matrix calculated synchronously in the solving process.

[0087] Further, the object dynamic equation includes a nonlinear elastic state equation (specifically embodied as a polytropic process gas state equation) for establishing a functional mapping relationship between the elastic element stroke, the medium state, and the nonlinear elastic restoring force.

[0088] Specifically, in the calculation of the nonlinear elastic restoring force, the polytropic process gas state equation is adopted; wherein the value range of the polytropic index n is generally between 1.0 (isothermal process) and 1.4 (adiabatic process). In the embodiment, considering the actual heat exchange characteristics of the rubber airbag, the polytropic index n is set to 1.3 (empirical value). The basis for selecting this value instead of 1.4 (adiabatic process) or 1.0 (isothermal process) is that the deformation frequency of the rubber airbag is fast during the operation of the controlled object, the heat exchange is insufficient but not completely adiabatic, and this value can better reflect the actual material characteristics of the rubber airbag. For other industrial dampers using different materials (such as metal bellows) or different media (such as nitrogen), those skilled in the art can select the corresponding value within the range according to the thermodynamic property table of the medium without changing the equation structure of the model.

[0089] Specifically, a temperature correction factor calculation step is introduced, the real-time collected gas temperature and the standard reference temperature are read for ratio operation, and the initial pressure parameter is corrected by using the ratio to compensate for the stiffness thermal decay effect caused by the heating of the gas chamber due to long-time operation.

[0090] Specifically, when in any discrete state in the discrete event state set, the continuous-time dynamic equation group called includes the aerodynamic equation and the friction force model. For the aerodynamic equation, the rate of change of the gas chamber pressure with time is described as the sum of two terms: the first term is the product of the gas chamber pressure, the effective cross-sectional area, and the stroke speed, and the second term is the input control term related to the gas constant, the working temperature, and the gas mass flow, both of which are divided by the current gas chamber volume. For the friction force model, in order to make up for the deficiency of the linear model, the LuGre model commonly used in the field is introduced into the basic equation to describe the piston rod friction, which describes the friction state as a nonlinear function related to the relative motion speed and the internal bristle deformation state, and provides a basic physically consistent description.

[0091] Furthermore, the object dynamics equation also includes: a damping dynamics equation (specifically embodied in the orifice flow equation), used to establish a functional mapping relationship between the damping adjustment opening, the medium flow state, and the pressure difference-flow characteristics.

[0092] Specifically, the damping force is calculated based on Bernoulli's equation, employing a piecewise hybrid modeling strategy: a critical Reynolds number is set, typically between 2000 and 2500 depending on the channel diameter; in this embodiment, 2300 is used. When the real-time calculated Reynolds number is less than the critical value, the fluid is determined to be in a laminar state, and a linear laminar flow model is used. When the Reynolds number is greater than or equal to the critical value, it is determined to be turbulent, and a nonlinear turbulent flow model is used. A smoothing weighting function is used to achieve a continuous transition between the two flow regimes near the critical point. This piecewise function mapping ensures that the model can accurately describe the nonlinear switching characteristics of the shock absorber between low-speed high-damping and high-speed low-damping.

[0093] Furthermore, regarding the generation of the nominal reference state trajectory and the linearized state deviation, the steps include: monitoring the rate of change of the low-frequency disturbance state variables of the input variable topology industrial mechanism model, and when the preset conditions are met, locking the target topology configuration and activating the corresponding object dynamics equation.

[0094] Specifically, the time derivative of the vertical input velocity component of the road surface (i.e., the vertical acceleration of the road surface) in the aforementioned low-frequency disturbance state variables is calculated in real time, and a migration threshold (e.g., 9.8 m / s²) is set. When the absolute value of the road surface acceleration exceeds the threshold, it is determined that the controlled object has been impacted by the speed bump, and the state migration is forcibly triggered. The object immediately switches from the pressure holding and locking state to the compression and inflation state, and the object dynamics equation corresponding to this state, which is based on the air chamber pressure and the compression chamber flow rate, is activated.

[0095] Furthermore, a discretized numerical integration algorithm is used to iteratively solve the dynamic equations of the activated object to generate a nominal reference state trajectory; at each discrete time node, the linearized state deviation is calculated.

[0096] Specifically, a numerical integration algorithm (such as the fourth-order Runge-Kutta method) is adopted, and the prediction time domain is set to 50ms (e.g., step size 1ms, prediction 50 steps). The basis for selecting a prediction time domain of 50ms is that the physical response time of actuators (solenoid valves and air pumps) is usually between 10 and 20ms. 50ms is sufficient to cover the response delay of the actuator and reserve sufficient control margin, while avoiding excessive cumulative error due to excessive prediction time.

[0097] Specifically, at each prediction time step, a first-order partial derivative matrix (Jacobian matrix) of the object dynamics equation at the current nominal state point is calculated; the actual observed state is subtracted from the nominal state to calculate the state deviation, and the deviation is verified to satisfy the linearization assumption, and finally the deviation is output as the linearized state deviation for subsequent input into the deep state observation network for calculation.

[0098] Embodiment Three:

[0099] In this embodiment, the construction of the deep state observation network and the generation of the nonlinear model mismatch compensation quantity in the shock absorber performance optimization control method based on model fusion are described in detail.

[0100] Further, regarding the generation of the nonlinear model mismatch compensation quantity, the method comprises: the deep state observation network adopts a deep neural network architecture with time series memory capability (such as a recurrent neural network) to perform time series feature extraction on the input linearized state deviation sequence.

[0101] Specifically, the linearized state deviation calculated in Embodiment Two is received as input. Due to the strong historical dependence of the damping force hysteresis effect of the controlled object (shock absorber) and the viscoelasticity of the elastic element, a single-time error is not enough to infer the transient change of the controlled object, so a network with long short-term memory operation logic is used to process the deviation sequence.

[0102] Specifically, in a preferred implementation configuration, the deep state observation network in this embodiment adopts a physically-aware long short-term memory network architecture, and the input layer dimension of the network is set to 6 (corresponding to the bias dimension of state quantities such as displacement and velocity), specifically including: the bias of the vertical displacement on the spring, the vertical velocity on the spring, the vertical displacement under the spring, the vertical velocity under the spring, the pressure in the air chamber, and the friction internal state; the network includes 2 hidden layers, each hidden layer is configured with 64 neural units; the engineering basis for selecting 64 neural units is that the effective modal number of the controlled object dynamics is usually small (generally less than 10 orders), 64 units are enough to capture high-dimensional nonlinear features, while avoiding excessive computational load, ensuring that the operation process can be completed within the preset control period (e.g. 5ms).

[0103] Specifically, the time step is set to 10, i.e. the network reads the linearized state deviation sequence in the past 10ms as input window each time; the sequence information is filtered and memorized by using the gating mechanism (such as the forget gate, the input gate and the output gate) inside the network, the hidden state vector is calculated, and finally the preliminary nonlinear compensation vector is obtained by mapping back to the 6-dimensional space through the fully connected output layer. The activation function in the hidden layer adopts a variant of linear rectifier unit (ReLU) (such as Leaky ReLU), and the output layer adopts linear or hyperbolic tangent function to adapt to the large amplitude compensation output in both positive and negative directions.

[0104] Further, a composite loss functional is constructed with a physical penalty term, which is constructed based on the stability derivative of the energy function of the controlled object, and the physical penalty term is used to impose a high loss penalty on the network prediction output that violates the energy passivity condition, guiding the model parameters to converge to the energy dissipation region.

[0105] Specifically, in the network operation solving or weight updating process, the composite loss functional is defined as the weighted sum of the numerical approximation term and the physical penalty term. The numerical approximation term (such as mean square error) is used to quantify the numerical approximation degree between the network output and the true observation residual.

[0106] Specifically, the physical penalty term is constructed to guide the network to follow the physical energy conservation law; first, the total energy function (Hamilton-like function) of the controlled object is constructed, which is defined as the sum of the translational kinetic energy of the mass on the spring and the mass under the spring, plus the integral of the nonlinear elastic potential energy. In the composite loss functional, in addition to the data-driven loss term that calculates the mean square error between the predicted value and the true value, a physical penalty term is also included. The calculation logic of the physical penalty term is: the rate of change of the total energy of the system with respect to time corresponding to the network prediction state is calculated in real time; if the rate of change is positive, it indicates that the internal energy of the system is increasing unnaturally, which violates the passivity constraint of the physical system, at this time the positive value is truncated by the rectified linear unit and multiplied by a preset penalty weight coefficient (for example, 10.0) to activate the penalty term, so that it produces a high loss value and is added to the composite loss functional; if the rate of change is negative or zero, it indicates that the system is in an energy dissipation or conservation state, which conforms to the physical law, at this time the penalty term is zero. The preset penalty weight coefficient is determined based on the statistical characteristics of the offline training stage, and the setting principle is to balance the order of magnitude difference between the numerical approximation term and the physical penalty term, so as to ensure that the gradient component generated by the physical constraint has sufficient correction effect in the back propagation process, and prevent the physical constraint mechanism from being ineffective due to the small weight proportion.

[0107] Specifically, the derivative of the energy function with respect to time (i.e. Lyapunov derivative) is calculated, which represents the total energy change rate of the controlled object; according to the passivity theory, the energy change rate of the controlled object under no external work should be less than or equal to zero (i.e. energy dissipation). When the energy change rate corresponding to the model prediction state is positive, it is determined that the current operation state violates the passivity dissipation condition; at this time, the positive value is multiplied by a preset penalty weight coefficient (for example, 0.1) as a physical penalty term added to the loss function.

[0108] Specifically, by minimizing the composite loss functional, the deep state observer is forced to learn the data characteristics while its weight update direction is strictly limited to the physical dissipation region; this mechanism mathematically guarantees that the compensation quantity output by the network will not cause the energy of the controlled object to diverge unnaturally, thereby ensuring the stability of the closed-loop control system.

[0109] Further, a complementary projection operator is configured in the output operation stage, and a tangent vector of the nominal reference state trajectory is used to construct a residual subspace projection matrix, so that the output vector is forced to project into a residual subspace complementary to the tangent direction of the nominal reference state trajectory.

[0110] Specifically, the physical mechanism of this step is that the nominal reference trajectory generated by the variable topology mechanism model mainly describes the main motion trend (i.e. the tangent component) conforming to the physical law; and the complex friction, high-frequency leakage and other nonlinear disturbances ignored due to model simplification are regarded as orthogonal or complementary residuals superimposed on the main motion trend. Through the projection operation, the data-driven model can be prevented from repeatedly learning the dynamics law mastered by the mechanism model, realizing the decoupling of "mechanism governing the main part and data governing the residual part".

[0111] Specifically, the state derivative vector of the nominal reference state trajectory generated in Embodiment Two at the current time is extracted, i.e. the tangent vector; the tangent vector is normalized to obtain a unit tangent vector.

[0112] Specifically, a complementary projection matrix is constructed, and the construction logic of the matrix is based on the self-multiplication (outer product) of the unit tangent vector subtracted from the unit matrix; the geometric meaning of the matrix is to define a linear subspace complementary to the tangent direction of the nominal trajectory.

[0113] Specifically, the preliminary nonlinear compensation vector output by the deep state observer network is subjected to matrix multiplication operation with the complementary projection matrix; this operation filters out the components parallel to the mechanism model prediction direction in the network output, and only the complementary components are retained.

[0114] Specifically, the final vector output after the projection operation is the nonlinear model mismatch compensation quantity. In the physical sense, the compensation quantity is strictly limited to the secondary effects that the mechanism model cannot describe, including the complex friction between the piston rod and the seal, the nonlinear leakage of the valve gap, etc.; these components are accurately separated through the above projection operation and are superimposed into the control loop as pure residual signals.

[0115] Embodiment Four:

[0116] In this embodiment, the generation of the adaptive weighting parameter and the fusion calculation step of the generalized state estimation value in the shock absorber performance optimization control method based on model fusion are described in detail.

[0117] In the traditional training process, the validation set is used to evaluate the performance of the model on unseen data; in this embodiment, the validation set is defined as a set of prediction distributions generated by performing multiple Monte Carlo random masking inferences on the same input signal within a single control cycle.

[0118] The construction of the validation set does not rely on manually labeled true values, but is verified by statistical methods. The specific training and validation logic closed loop includes:

[0119] The first step is the offline pre-training phase, using the standard road data collected historically as the training set, performing gradient descent training on the deep state observation network until the loss function converges, and determining the initial baseline parameters of the network weights.

[0120] The second step is the online running phase, freeze the network weights, but activate the random inactivation layer, for each real-time sampled input state sequence, real-time generate no less than twenty prediction samples, this group of prediction samples constitute the current instantaneous validation set;

[0121] The third step is to calculate the statistical dispersion in the instantaneous validation set, which is used as a criterion to evaluate whether the model is suitable for the current working condition, thereby realizing the validation of the model effectiveness without true value reference.

[0122] Further, regarding the generation step of the adaptive weighting parameter, the method includes: enabling the probability inference mechanism in the calculation process of the deep state observation network, and performing forward propagation by randomly masking sampling (this embodiment uses Monte Carlo dropout method) the network operation weights multiple times.

[0123] Specifically, in the inference calculation process of the aforementioned deep neural network with timing memory capability, the probability inference logic is activated; the dropout rate of the Monte Carlo dropout method is set to 0.3, and 0.3 is selected as the parameter value. The basis for selecting 0.3 is that in the regularization calculation of the deep neural network, a dropout rate of 0.2 to 0.5 can effectively simulate the sparsity change of the internal connection structure of the neural network, and 0.3 is the balance point between maintaining the convergence of the prediction mean and introducing sufficient random disturbance variance after offline verification.

[0124] Further, in the generation step of the adaptive weighting parameter, for the linearized state deviation sequence input at the same time, the random masking sampling strategy is executed, 30% of the connection weights in the neural network hidden layer are randomly masked, and 20 to 50 times of forward propagation operation are performed in parallel to construct the output sample set.

[0125] Specifically, the network structure is randomly disturbed while keeping the input data unchanged; 20 to 50 times are selected as the number of parallel forward propagation, which is determined by balancing the computing power boundary of the vehicle-mounted computing unit within the preset real-time control period (for example, 5 to 10 milliseconds) and the requirement of the law of large numbers. The sampling size can ensure that the output sample set constructed is sufficient to meet the Gaussian distribution assumption in statistics, thereby realizing effective approximation of the distribution characteristics of the network output, and will not cause calculation timeout.

[0126] In order to ensure the effectiveness of the verification mechanism, the deep state observation network needs to complete the basic offline training step before deployment. The specific training parameter settings are as follows:

[0127] The small batch random gradient descent algorithm is used for optimization, the batch size is set to 128; the initial learning rate is set to 0.001, and the cosine annealing strategy is used for decay; the training iteration round is set to 500 rounds, and the mean square error is used as the loss function; in the offline training process, a standard data set containing asphalt road, cement road and speed bump is pre-defined, and the training set and the validation set are divided according to the proportion of 7:3. The standard data set is specifically composed of vehicle driving data of three different road conditions: one is the asphalt road driving data meeting the ISO8608A road standard, which is used to represent high-frequency low-amplitude random texture noise; the second is the cement road driving data containing periodic joint impact, which is used to represent medium-frequency structured vibration; the third is the transient impact driving data of crossing trapezoidal speed bump, which is used to represent low-frequency large-amplitude nonlinear sudden road conditions.

[0128] The validation set is used to calibrate the uncertainty threshold of the system. Specifically, in the offline environment, the validation set is input into the trained deep state observation network, the above-mentioned multiple random mask reasoning is performed, the variance distribution range when the model prediction is accurate is counted, and the statistical boundary value with a distribution confidence of 95% is determined as the variance saturation threshold in the subsequent online control (i.e. 0.1 mentioned in the embodiment).

[0129] This strategy solves the technical problem that online uncertainty evaluation of deep learning model usually requires huge calculation overhead and is difficult to deploy on embedded end. Through fine calibration of the sampling number (20-50 times), high confidence quantification of cognitive uncertainty is realized under low calculation load, which prevents the loss of robustness caused by blindly relying on data model in unobserved working conditions, and realizes the balance between real-time performance and safety.

[0130] Further, the distribution variance of the output results of multiple forward propagations is counted, the distribution variance is mapped to a normalized cognitive uncertainty index, and an adaptive weighting parameter negatively correlated with the cognitive uncertainty index is output.

[0131] In this process, the distribution variance is defined as the pass rate index of the model online verification. The verification discrimination logic is as follows:

[0132] The mean value of the variance of all predicted samples in the preceding instantaneous verification set is calculated, denoted as the cognitive uncertainty metric value at the current time; the metric value is compared with the variance saturation threshold (numerical value 0.1); if the metric value is less than or equal to the variance saturation threshold, it is determined that the model passes the verification, indicating that the current working condition is within the known knowledge domain of the model, and the adaptive weighting parameter close to 1 is generated, and the control strategy is mainly based on the deep network output; if the metric value is greater than the variance saturation threshold, it is determined that the model does not pass the verification, indicating that the current working condition belongs to an unseen abnormal distribution (such as extreme impact or sensor failure), at this time, no manual intervention is needed, and the adaptive weighting parameter is automatically adjusted to 0 according to the preset exponential decay formula.

[0133] This mechanism is equivalent to constructing a real-time changing verification set, which can capture the drift of the input data distribution, thereby improving the defect that the fixed verification set cannot cover the infinite unknown working conditions, and meeting the verification demand for the generalization ability of the model.

[0134] Specifically, statistical operations are performed on the 50 nonlinear model mismatch compensation quantity vectors generated by the above 50 forward propagations, the variance values of the 50 vectors in each state dimension are calculated, and the arithmetic mean of the variances of all dimensions is taken as the overall prediction variance (i.e. distribution variance) at the current time; in this calculation instance (for example, for the deceleration strip disturbance working condition), due to the sharp mutation of the external disturbance leading to the deviation of the input data distribution from the training data distribution, the overall prediction variance value calculated is 0.04.

[0135] Specifically, the overall prediction variance is mapped to the cognitive uncertainty index; the variance saturation threshold is set to 0.1, which is determined based on the maximum error distribution boundary of the network in the offline training phase under unseen working conditions; the calculated 0.04 is divided by 0.1 to obtain the normalized cognitive uncertainty index of 0.4; the index quantifies the cognitive deviation degree of the deep state observation network for the current input working condition in numerical value: the larger the value, the lower the confidence region of the network.

[0136] Specifically, the cognitive uncertainty index is substituted into the preset adaptive decay function to calculate the adaptive weighting parameter, and the decay function is defined as an exponential function with the natural constant e as the base and the negative value of the cognitive uncertainty index as the index; by substituting 0.4 into the calculation, the adaptive weighting parameter is about 0.67; the parameter defines the fusion algorithm's acceptance weight of the nonlinear model mismatch compensation quantity in numerical value, i.e. in the current high dynamic working condition, the algorithm logic tends to retain about 67% of the data-driven compensation component to reduce the risk of model illusion caused by data distribution extrapolation.

[0137] Further, regarding the step of generating a generalized state estimate, the method comprises: linearly superimposing the nonlinear model mismatch compensation quantity and the nominal reference state trajectory by using an adaptive weighting parameter as a weight factor.

[0138] Specifically, a dynamic gain matrix is constructed by using the calculated adaptive weighting parameter 0.67; in this embodiment, a diagonal matrix with the adaptive weighting parameter as the diagonal elements is constructed as an "amplitude modulation operator". The function of this operator is to dynamically adjust (scale) the amplitude of the data-driven compensation quantity according to the current cognitive uncertainty of the model.

[0139] Specifically, the operation objects are defined: the state point of the nominal reference state trajectory generated in the foregoing embodiment at the current time is defined as the reference state, and the generated nonlinear model mismatch compensation quantity is defined as the residual correction quantity.

[0140] Specifically, an amplitude modulation operation is performed, in which the nonlinear model mismatch compensation quantity is multiplied by the dynamic gain matrix (i.e., multiplied by 0.67) to obtain a modulated compensation vector; this step completes the confidence-based signal amplitude scaling, which mathematically ensures that the output residual of the network is completely adopted only when the network confidence is high.

[0141] Specifically, a linear superimposition fusion operation is performed, and the calculation logic is: the state vector corresponding to the nominal reference state trajectory is subjected to a vector addition operation with the modulated compensation vector to obtain a corrected state vector. This process realizes the optimal state estimation based on statistical uncertainty in the Euclidean space.

[0142] Specifically, to ensure that the fused state variables meet the object feasible region constraints, a truncation process is performed after the superimposition operation; for example, the key state variable (such as the gas chamber pressure) of the controlled object is checked, and if the superimposed value is lower than the physical lower limit (such as atmospheric pressure), it is forced to be truncated to the minimum physical boundary value; the finally output generalized state estimate value is a high-precision state vector that contains both the stable reference trend provided by the mechanism model and the nonlinear detail compensation weighted by the uncertainty.

[0143] This embodiment realizes online verification of the optimization method by constructing a verification set and calculating a variance index. This process essentially constitutes a real-time verification and adaptive closed loop of the robustness of the control algorithm, solving the problem that the traditional static verification set cannot cover unknown working conditions.

[0144] Embodiment Five:

[0145] This embodiment details a step of generating a dynamic multi-objective optimization decision and a hybrid modal control input in a shock absorber performance optimization control method based on model fusion.

[0146] Further, the step of constructing a dynamic multi-objective optimization functional and searching for a control working point comprises: defining an output response fluctuation functional representing disturbance suppression performance and a contact stability functional representing stability maintenance performance; identifying the current external working condition category using the generalized state estimation value, and dynamically adjusting the weight coefficients of the two functionals according to the category to synthesize a dynamic multi-objective optimization evaluation functional.

[0147] Specifically, first, two independent physical performance evaluation sub-functions are constructed; the first sub-function calculates the weighted root mean square value of the output response (such as the vertical acceleration of the mass on the spring) component in the generalized state estimation value, and the weight function preferably adopts a human comfort sensitivity curve (assigning a high weight to a specific frequency band such as 4Hz to 8Hz) to quantify the disturbance suppression performance; the second sub-function calculates the square integral of the contact stability index (such as the tire dynamic load or contact force fluctuation) in the generalized state estimation value to quantify the contact stability of the controlled object with the environment.

[0148] Further, in the construction of the dynamic multi-objective optimization functional, when it is monitored that the external disturbance input speed exceeds a preset threshold and the contact deformation increases sharply, the proportion of the stability maintenance performance weight is forced to be raised to a dominant position (such as 0.7).

[0149] The generalized state estimation value is characterized and discriminated, and when a transient large impact working condition such as deceleration zone climbing is identified (characterized by sudden change of input speed to 1.5m / s and accompanied by large deformation), the weight configuration of the optimization functional is dynamically adjusted; the weight distribution in the basic configuration focusing on comfort (such as comfort 0.6 / stability 0.4) is quickly reversed, and the stability weight is set to 0.7 and the comfort weight is set to 0.3; the engineering basis of 0.7 is the experience of dynamic calibration, and this weight is sufficient to force the stroke to be within the physical limit in optimization.

[0150] This mechanism solves the technical problem that excessive pursuit of comfort in extreme working conditions may lead to multi-objective optimization control. Through weight dynamic reversal based on working condition identification, the physical safety of the mechanical structure and the tire grip are prioritized in the moment of impact, realizing the intelligent trade-off of "comfort first, safety first in danger".

[0151] Further, with the saturation characteristics of the execution end in the variable topology industrial mechanism model as a boundary constraint, a nonlinear constraint optimization algorithm is used to optimize the dynamic multi-objective optimization evaluation functional within the feasible region of the object, and the optimal state derivative vector located on the instantaneous Pareto frontier is solved as the control working point.

[0152] Specifically, the boundary constraint condition of the feasible region of the object is set, and the physical constraint index of the model is defined: the physical constraint index is determined according to the physical limit of the actuator, the search range (for example, 0A to 2.0A) of the driving control quantity (for example, electromagnetic valve current) is set, the search range (for example, 5bar to 12bar) of the elastic element parameter (for example, air chamber pressure) is set, and the upper and lower limits of the parameter change rate (limited by the fluid pipeline and power source) are set. When implementing the control method, the specific values of the boundary index should be directly determined according to the rated specification parameters of the actuators actually equipped in the controlled object.

[0153] Specifically, a nonlinear constraint optimization algorithm (for example, sequential quadratic programming SQP) is executed in the above-mentioned feasible region of the object; the synthesized dynamic multi-objective optimization evaluation functional is taken as the objective function, the generalized state estimation value at the current time is taken as the optimization starting point, and the optimal state derivative (that is, the expected speed and acceleration of the controlled object) at the next time is searched; after several iterations of the algorithm, the optimal state derivative vector calculated makes the value of the evaluation functional minimum; the optimal state derivative vector is the control working point satisfying the Pareto optimal condition, and its physical meaning represents the best motion trend balance point that the controlled object can theoretically reach under the premise of meeting the constraints under the current working condition.

[0154] Further, the step of decoupling the control working point into hybrid mode control input includes: constructing an inverse dynamics operation model of the execution end based on the differential flatness principle, and mapping the optimal state derivative vector as input into generalized control force demand and generalized stiffness demand.

[0155] Specifically, the inverse dynamics operation model is configured to receive the optimal state derivative vector calculated by the dynamic multi-objective optimization functional as input data, and the optimal state derivative vector includes the expected vertical speed and expected vertical acceleration of the controlled object at the next time; the inverse dynamics operation model internally performs structured operation based on the principle of mechanical equilibrium to realize signal conversion:

[0156] The product of the expected vertical acceleration and the preset on-mass inertia parameter of the controlled object is calculated to obtain a theoretical inertia force demand; the theoretical inertia force demand is subtracted by the gravity component of the controlled object at the current time and the system inertia force component generated by the current motion state to calculate the required generalized force demand of the execution end; the generalized force demand is received and decoupled by using a frequency domain complementary filtering strategy: a low-pass filter is used to extract the component with a frequency lower than a set cutoff frequency (e.g. 1 to 2 Hz) in the generalized force demand, which is taken as a low-frequency support component borne by the elastic element; a high-pass logic is used to extract the component with a frequency higher than the cutoff frequency but within the effective control bandwidth (e.g. the cutoff frequency to 20 Hz) of the system, which is taken as a high-frequency dissipation component borne by the damping element; finally, the low-frequency support component is divided by the current relative motion stroke of the controlled object to calculate the generalized stiffness demand, and the high-frequency dissipation component is directly mapped as the generalized control force demand; the generalized stiffness demand and the generalized control force demand are taken as output data, which are respectively used for subsequent elastic element characteristic query step and damping element inverse calculation step, thereby establishing an explicit association relationship from the kinematic target to the dynamic execution target in vehicle shock absorption control.

[0157] Further, according to the generalized stiffness demand, the target parameters of the elastic element and the corresponding stiffness adjustment instructions are determined by table lookup calculation.

[0158] Specifically, according to the pre-stored elastic element characteristic data (such as the stiffness-stroke-pressure three-dimensional pulse spectrum), the current measured stroke and the calculated stiffness demand are taken as inputs to inversely calculate the target parameters (such as the target pressure of the air chamber).

[0159] Specifically, the adjustment action timing is calculated. The difference between the source end parameters and the current target end parameters (such as a pressure difference of 6.0 bar) is known, and according to the medium flow characteristic formula (such as the Saint-Venant formula), the medium mass flow through the effective flow area of the regulating valve under the current difference is calculated; the medium mass required to reach the target parameter is calculated, and the flow is combined to obtain the opening duration of the regulating valve (e.g. 120 ms).

[0160] Specifically, the stiffness adjustment instruction is generated: since the target parameter is higher than the current parameter, the logic judges that it is the pressure increasing mode, and outputs the "open the regulating valve" instruction, and the opening duration is set to 120 ms.

[0161] Further, according to the generalized control force demand and the fluid characteristics under the current damping mode, the target parameters of the damping element are calculated by inverse solution, and the target parameters are converted into modulation control instructions for driving the actuator to act; the stiffness adjustment instructions and the modulation control instructions are combined to form the hybrid modal control input. The detailed steps are as follows: calculating the target damping parameters: dividing the generalized control force demand (for example, 2500 Newton) by the current relative motion speed in the generalized state estimation value to obtain the required damping coefficient (target flow resistance); converting the control signal: according to the input-output characteristic map of the damping element (for example, the current-damping force F-V curve), the corresponding control amount (for example, the current 1.6A) that can generate the target damping force is found on the characteristic curve of the current relative speed; generating the modulation control instruction: according to the actuator driving circuit parameters (such as the maximum peak current 2.5A), the target control amount is converted into the duty cycle (for example, 65%). Finally, the final output is executed: the logical instruction of "opening the regulating valve for 120 milliseconds" and the modulation instruction of "duty cycle 65%" are synchronously sent; the two groups of instructions are used as the hybrid modal control input to cooperatively drive the actuator, so that the controlled object can instantly increase the stiffness to prevent stroke saturation and increase the damping force to quickly attenuate vibration when suffering from disturbance, thereby achieving the dynamic balance of disturbance suppression and stability.

[0162] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and variations can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A model fusion-based shock absorber performance optimization control method, characterized by, The method comprises the following steps: acquiring real-time running feedback signals of the controlled object and performing signal feature extraction and reconstruction, and extracting low-frequency disturbance state variables; inputting the low-frequency disturbance state variables into a variable topology industrial mechanism model embedded with object dynamic equations to generate a nominal reference state trajectory and a linearized state deviation; constructing a deep state observation network integrating energy passivity constraints, inputting the linearized state deviation into the deep state observation network, and using a physical punishment mechanism in the deep state observation network to solve a nonlinear model mismatch compensation under the condition of satisfying the energy passivity constraints; meanwhile, based on the cognitive uncertainty characteristics of a distribution output by the deep state observation network, adaptive weighting parameters are calculated and output; the adaptive weighting parameters are used as adjustment factors to perform dynamic state correction operations in a state space, and the nonlinear model mismatch compensation is used as a residual term to be superimposed to the nominal reference state trajectory to generate a generalized state estimation value; defining an output response fluctuation functional representing disturbance suppression performance and a contact stability functional representing stability maintenance performance; taking the generalized state estimation value as a variable to identify a current external working condition category, dynamically adjusting weight coefficients of the two functionals according to the external working condition category, constructing a dynamic multi-objective optimization functional, and using a nonlinear constraint optimization algorithm to iteratively optimize the dynamic multi-objective optimization functional in an object feasible region defined by the variable topology industrial mechanism model to search for a control working point satisfying a Pareto optimal condition, and decoupling and converting the control working point into hybrid mode control inputs acting on an execution end.

2. The model fusion-based shock absorber performance optimization control method according to claim 1, characterized in that, The extraction step of the low-frequency disturbance state variables comprises the following steps: acquiring real-time running feedback signals, processing the real-time running feedback signals by using a multi-scale signal decomposition algorithm, decomposing the real-time running feedback signals into modal components of different frequency bandwidths, and calculating time-frequency distribution characteristics of the modal components; performing effective mode selection based on the signal characteristics of the modal components, filtering out high-frequency random texture noise components, retaining and reconstructing effective disturbance energy components containing transient impact characteristics; performing data dimension reduction processing on the reconstructed effective disturbance energy characteristics, extracting principal components whose cumulative variance contribution rates satisfy a preset condition, and reconstructing the principal components into low-frequency disturbance state variables with a dimension consistent with an input space of the variable topology industrial mechanism model.

3. The model fusion-based shock absorber performance optimization control method according to claim 1, characterized by, The variable topology industrial mechanism model is a multi-modal coupled model constructed based on discrete-continuous hybrid dynamics theory, and specifically comprises the following steps: The variable topology industrial mechanism model is composed of a discrete event state set and a continuous-time dynamic equation set; the discrete event state set represents a topology configuration of an execution end, and includes geometric configuration states corresponding to different stiffness levels of elastic elements and damping characteristic states corresponding to different flow resistance characteristics of damping elements; the continuous-time dynamic equation set includes a group of object dynamic equations one-to-one mapped with the discrete event state set, and is used to describe continuous dynamic behaviors of the controlled object under corresponding topology configurations.

4. The model fusion-based shock absorber performance optimization control method according to claim 1, characterized by, The object dynamic equations comprise the following steps: The nonlinear elastic state equation establishes a functional mapping relationship among the stroke of the elastic element, the medium state and the nonlinear elastic restoring force; and the damping dynamics equation establishes a segmented mixed dynamics function mapping relationship among the damping adjustment opening, the medium flow state and the differential pressure-flow characteristic, and contains a smooth transition logic between different flow patterns.

5. The model fusion-based shock absorber performance optimization control method according to claim 1, characterized by, The generating step of the nominal reference state trajectory and the linearized state deviation comprises: The change rate of the low-frequency disturbance state variable input into the variable topology industrial mechanism model is monitored, and when a preset condition is met, the target topology configuration at the current time is locked in the discrete event state set, and the target object dynamics equation corresponding to the target topology configuration is activated from the continuous time dynamics equation set; a discrete numerical integration algorithm is used to iteratively solve the activated target object dynamics equation, and a dynamics state evolution curve in a prediction time domain is generated as a nominal reference state trajectory; at each discrete time node of the nominal reference state trajectory, a first-order partial derivative matrix of the object dynamics equation with respect to the model state variable is calculated analytically, and a tangent space is defined using the first-order partial derivative matrix; a difference vector between the actual observed state at the current time and the projection of the nominal reference state trajectory in the tangent space is calculated as a linearized state deviation.

6. The model fusion-based shock absorber performance optimization control method according to claim 1, characterized by, The generating step of the nonlinear model mismatch compensation quantity comprises: The deep state observation network adopts a deep neural network architecture with time sequence memory capability to perform time sequence feature extraction on the input linearized state deviation sequence; a compound loss functional that fuses a physical penalty term is constructed, the physical penalty term is constructed based on the stability derivative of the controlled object energy function, the loss penalty is applied to the network prediction output that violates the energy passivity condition by using the physical penalty term, and the model parameters are guided to converge to the energy dissipation area; a complementary projection operator is configured in the operation output stage, a residual subspace projection matrix that is complementary to the tangent direction is constructed by using the tangent vector of the nominal reference state trajectory, the output vector is forced to project into the residual subspace, and a nonlinear model mismatch compensation quantity containing a vertical direction residual that is not captured by the variable topology industrial mechanism model is generated.

7. The model fusion-based shock absorber performance optimization control method according to claim 1, characterized by, The generating step of the adaptive weighting parameter comprises: A probability inference mechanism is enabled in the calculation process of the deep state observation network, the network operation weight is sampled multiple times by random masking and forward propagation is performed; the distribution variance of the output results of multiple forward propagations is counted, the distribution variance is mapped to a normalized cognitive uncertainty index, and the adaptive weighting parameter is output.

8. The model fusion-based shock absorber performance optimization control method according to claim 1, characterized by, The generating step of the generalized state estimation value comprises: A dynamic gain matrix is constructed by using the adaptive weighting parameter as a weight factor, a statistical weighting space based on uncertainty is defined; the nonlinear model mismatch compensation quantity located in the tangent space is subjected to amplitude modulation operation by using the dynamic gain matrix, and based on the linear superposition criterion, the compensated quantity and the nominal reference state trajectory are subjected to weighted fusion calculation, and the modified state vector is obtained as the generalized state estimation value.

9. The model fusion-based shock absorber performance optimization control method according to claim 1, wherein The step of decoupling the control operating point into a mixed modal control input comprises: Based on the differential flatness principle, an inverse dynamics model is constructed to describe the input-output relationship of the execution end, and the optimal state derivative vector corresponding to the control work point is decoupled and mapped into the generalized control force demand and the generalized stiffness demand; according to the generalized stiffness demand, the target parameters of the elastic element and the corresponding stiffness adjustment instruction are calculated and determined by querying the pre-stored elastic element characteristic data; according to the generalized control force demand and the fluid characteristics under the current damping mode, the target parameters of the damping element are calculated by inverse solution, and the target parameters are converted into the modulation control instruction for driving the actuator to act; the stiffness adjustment instruction and the modulation control instruction are combined to form the hybrid modal control input.

Images