A semi-active suspension bottom-layer closed-loop force tracking control method and system based on data-driven forward and inverse models

Abstract

The application discloses a semi-active suspension bottom-layer closed-loop force tracking control method and system based on data-driven forward and inverse models, and relates to the technical field of vehicle control. The method comprises the following steps: determining the piston rod speed and the target damping force of the shock absorber by using a state estimation-upper layer control module; performing constraint processing on the target damping force by using a damping force boundary characteristic curve; inputting the piston rod displacement, the piston rod speed and the input current into the forward model to obtain the damping force estimation value of the shock absorber; performing damping force tracking based on the target constraint damping force and the damping force estimation value to determine the control damping instruction containing an error term; inputting the piston rod displacement, the piston rod speed and the control damping instruction containing the error term into the inverse model to obtain the actual control current instruction of the shock absorber after constraint processing of the compensation control current of the shock absorber. The application improves the force tracking accuracy of the semi-active suspension by constructing data-driven forward and inverse models.

Description

Technical Field

[0001] This application relates to the field of vehicle control technology, and in particular to a semi-active suspension bottom closed-loop force tracking control method and system based on a data-driven forward and inverse model. Background Technology

[0002] Accurate tracking of damping force is the core control objective of the underlying actuator of a semi-active suspension. However, in the process of rapidly and accurately converting the target damping force command issued by the upper controller into the control current input to the shock absorber and then into the actual physical output force of the shock absorber, the existing technology reveals the following significant shortcomings in terms of the accuracy and robustness of force tracking when facing dynamic boundary constraints and execution delays of damping force under complex working conditions.

[0003] Existing semi-active suspension systems mostly utilize PID or sliding mode control (SMC) to dynamically adjust the control output for damping force closed-loop tracking. The principle is often to adjust the control rate by back-calculating the actual damping force using a low-order damper model. Some methods also characterize the dynamic characteristics of the damping force using RC operators and perform secondary optimization. However, due to the strong nonlinear time-varying characteristics of semi-active suspension systems, traditional mechanistic models cannot accurately characterize the internal hysteresis and fluid yield delay, failing to meet the requirements for real-time and high-precision force tracking. For example, when using Bouc-Wen parametric models to construct a force tracking inverse solver, the complex hysteresis nonlinearity of the damper and the differential equation characteristics of the parametric model itself lead to severe multivaluedness and singularity problems in the inverse mapping process. This makes it difficult to quickly calculate the unique instantaneous current required to track the target force. Furthermore, the parameters of the physical model are highly susceptible to drift with operating temperature and component wear, causing the force tracking error of the initially well-identified static inverse model to amplify sharply after continuous driving. Traditional shallow data-driven models are prone to current spikes and timing error accumulation in force tracking applications. Existing neural networks such as BP, FNN, or NARX lack efficient long-range memory of input sequences when used as inverse force tracking models. When the desired force is near the yield zone or changes rapidly, it is easy to output control current with high-frequency jump spikes. This high-frequency flutter not only fails to achieve smooth force tracking, but also excites high-frequency resonance in the suspension structure.

[0004] Besides the limitations of the damper model, some existing force tracking strategies habitually embed the dynamic characteristics of damping force directly into the suspension controller. This fusion of upper and lower level controllers introduces the uncertainties and calculation deviations of the lower-level actuator model into the upper-level suspension control decisions, causing model error divergence and significantly affecting the overall vehicle ride comfort. Furthermore, damping force and suspension state are coupled, and the suspension state is constrained by unknown random road surface excitations. Existing force tracking control methods employ an open-loop feedforward architecture without physical sensors, where the lower-level control system can only blindly issue commands, lacking an anti-disturbance closed-loop compensation mechanism. In engineering applications, due to the constraints of suspension structure and engineering economics, it is difficult to incorporate physical damping force sensors. When faced with unknown instantaneous road impacts or temperature drift, traditional strategies cannot meet the tracking accuracy requirements under complex dynamic environments, easily generating current spikes and accumulating timing errors, causing the actual force generated by the actuator to deviate significantly from the target force. Summary of the Invention

[0005] The purpose of this application is to provide a semi-active suspension bottom-level closed-loop force tracking control method and system based on data-driven forward and inverse models. It constructs a forward model and an inverse model based on a CNN-LSTM architecture to replace the measurement of the force sensor and the work of the feedback controller, respectively. A dynamic error compensation loop is introduced into the control layer to automatically learn the spatiotemporal dependency between system states, thereby solving the technical problems of inaccurate force tracking and lack of closed-loop feedback under complex working conditions, and thus improving the force tracking accuracy of the semi-active suspension.

[0006] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a semi-active suspension bottom-level closed-loop force tracking control method based on a data-driven forward and inverse model, the method being applied to a bottom-level closed-loop force tracking control architecture of a semi-active suspension system; The architecture includes: a sensor group, a filtering module, a state estimation-upper-layer control module, and a lower-layer force tracking module; The underlying force tracking module encapsulates a forward model and an inverse model; both the forward model and the inverse model are obtained by training the CNN-LSTM architecture using a training sample set. The state estimation-upper-level control module includes: a differential controller and an upper-level controller; The method includes: The displacement state of the semi-active suspension system at the current moment, as well as the input current and current constraint conditions of the damper, are obtained; the damper is a magnetorheological damper; the displacement state includes: sprung mass displacement, unsprung mass displacement, and road surface displacement. Based on the displacement state of the semi-active suspension system at the current moment and the input current of the shock absorber, the piston rod speed and the target damping force of the shock absorber are determined using the state estimation-upper-level control module. The target damping force is constrained using the damping force boundary characteristic curve to obtain the target constrained damping force; the damping force boundary characteristic curve is constructed using a forward model. The piston rod displacement, piston rod velocity, and input current are all input into the positive model to obtain the estimated damping force of the shock absorber; the piston rod displacement is determined based on the displacement state quantity. Damping force tracking is performed based on the target constraint damping force and the estimated damping force value to determine the control damping command containing the error term; The piston rod displacement and piston rod velocity, as well as the control damping command including the error term, are all input into the inverse model to obtain the compensation control current of the damper. The compensation control current is constrained using current constraints to obtain the actual control current command for the vibration damper.

[0007] Optionally, before determining the piston rod speed and the target damping force of the shock absorber using the state estimation-upper-level control module based on the displacement state of the semi-active suspension system at the current moment and the input current of the shock absorber, the following steps are also included: The input current and the displacement state quantity are both filtered using a filtering module.

[0008] Optionally, the filtering module includes a low-pass filter and a high-pass filter.

[0009] Optionally, based on the displacement state of the semi-active suspension system at the current moment and the input current of the shock absorber, the piston rod speed and the target damping force of the shock absorber are determined using the state estimation-upper-level control module, specifically including: Based on the displacement state quantity, the motion state of the semi-active suspension system is estimated using a differential controller to obtain the sprung mass velocity and the unsprung mass velocity. The difference between the sprung mass velocity and the unsprung mass velocity is determined as the piston rod velocity; The upper-level controller determines the target damping force of the vibration damper based on the sprung mass velocity, the unsprung mass velocity, the filtered input current, and the filtered displacement state quantity using a control strategy.

[0010] Optionally, the control strategy is an optimal control strategy.

[0011] Optionally, the control damping command including the error term is: ; in, This is a control damping command that includes an error term; The target is constrained by damping force; This is the estimated value of the damping force at the current moment.

[0012] Optionally, before obtaining the displacement state of the semi-active suspension system at the current moment, the following steps are also included: Constructing a CNN-LSTM architecture; The magnetorheological damper was subjected to characteristic tests to construct a training sample set. The training sample set consisted of time series data including piston rod displacement, piston rod velocity, damper input current, and damper output damping force. The damper output damping force was obtained by a damping force sensor in the sensor array. Using piston rod displacement, piston rod velocity, and damper input current as inputs, and damper output damping force as output, the CNN-LSTM architecture is trained using the training sample set to obtain a forward model; Using piston rod displacement, piston rod velocity, and damper output damping force as inputs, and damper input current as output, the CNN-LSTM architecture is trained using the training sample set to obtain the inverse model.

[0013] Optionally, the CNN-LSTM architecture specifically includes: a two-branch convolutional neural network and a long short-term memory network connected in sequence; The dual-branch convolutional neural network includes a first convolutional branch, a second convolutional branch, and a splicing layer; The outputs of both the first convolutional branch and the second convolutional branch are connected to the splicing layer; The first convolutional branch is used to slide along the time dimension based on the current data to generate local features; The second convolutional branch is used to slide along the time dimension based on historical data to generate local features; The Long Short-Term Memory network includes: a first LSTM layer, a second LSTM layer, and an output layer connected in sequence; The output layer of the splicing layer is connected to the input layer of the first LSTM layer.

[0014] Optionally, the working principle of the first convolutional branch and the second convolutional branch is as follows: ; in, Represents the convolutional branch in model s l Output local features; Indicates the activation function; Represents the convolutional branch in model s l The convolution kernel; Represents the input matrix; Represents the convolutional branch in model s l The bias term; The working principles of the first LSTM layer and the second LSTM layer are as follows: ; ; ; ; In the formula, , and These represent the forget gate, input gate, and output gate, respectively. and Indicates a hidden unit; This represents the feature matrix output by the dual-branch convolutional neural network at time t; , and All are bias terms; and Indicates the cell state; Indicates candidate values; This represents element-wise multiplication; The output of the forward model is: ; in, This represents the output of the forward model at time t; Represents the forward model The weight matrix of the output layer; Represents the forward model The output of the hidden unit at time t in the second LSTM layer; Represents the forward model The bias term of the output layer; The output of the inverse model is: ; in, This represents the output of the inverse model at time t; Representing the inverse model The weight matrix of the output layer; Representing the inverse model The output of the hidden unit at time t in the second LSTM layer; Representing the inverse model The bias term of the output layer.

[0015] Secondly, this application provides a semi-active suspension bottom-level closed-loop force tracking control system based on a data-driven forward and inverse model. The system applies the aforementioned semi-active suspension bottom-level closed-loop force tracking control method based on a data-driven forward and inverse model. The system includes: The data acquisition module is used to acquire the displacement state of the semi-active suspension system at the current moment, as well as the input current and current constraint conditions of the damper; the damper is a magnetorheological damper; the displacement state includes: sprung mass displacement, unsprung mass displacement, and road surface displacement; The target damping force determination module is used to determine the piston rod speed and the target damping force of the shock absorber based on the displacement state of the semi-active suspension system at the current moment and the input current of the shock absorber, using the state estimation-upper control module. The damping force constraint module is used to constrain the target damping force using the damping force boundary characteristic curve to obtain the target constrained damping force; the damping force boundary characteristic curve is constructed using a forward model; The damping force estimation module is used to input the piston rod displacement, the piston rod velocity, and the input current into the positive model to obtain the estimated damping force of the shock absorber; the piston rod displacement is determined based on the displacement state quantity. The damping force tracking module is used to track the damping force based on the target constraint damping force and the estimated damping force, and to determine the control damping command including the error term; The current estimation module is used to input the piston rod displacement and piston rod velocity, as well as the control damping command including the error term, into the inverse model to obtain the compensation control current of the vibration damper. The current constraint module is used to constrain the compensation control current using current constraint conditions to obtain the actual control current command of the vibration damper.

[0016] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a semi-active suspension bottom-level closed-loop force tracking control method and system based on data-driven forward and inverse models. Independent forward observation models and constrained inverse dynamic models are constructed using a CNN-LSTM architecture. The forward model outputs high-precision actual observed force in real time, avoiding the problem of error propagation from traditional mechanistic models to control decisions and replacing physical sensors. The inverse model uses historical and current displacement, velocity, and reconstructed target damping force as a joint input time window, utilizing long-range memory to mine the hysteresis physical evolution direction, and calculating and outputting a unique and smooth optimal control current in real time. Furthermore, this application constructs a bottom-level closed-loop architecture combining forward observation and inverse feedforward compensation, using the deviation between the observed damping force and the target force as feedback for dynamic superposition correction. Without changing hardware costs, this enables the system to have extremely strong anti-disturbance and dynamic error self-adjustment capabilities, achieving high-precision closed-loop force tracking under all operating conditions. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a semi-active suspension bottom closed-loop force tracking control method based on a data-driven forward and reverse model in one embodiment of this application. Figure 2 This is a schematic diagram of a semi-active suspension bottom closed-loop force tracking control system based on a data-driven forward and reverse model in one embodiment of this application. Figure 3 The static and dynamic boundary characteristics of the shock absorber in one embodiment of this application are shown. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] Terminology Explanation: ① Data-Driven Model: Relying on experimental data to capture the dynamic physical characteristics of the vibration damper and establish a deep learning model. ② Force Tracking Control: A control process that continuously corrects the input current by comparing the feedback of the actual damping force with the target force, so that the actual output force of the vibration damper accurately follows the target damping force.

[0022] Accurate tracking of damping force is a core task of the underlying control of semi-active suspension, and it is of paramount importance to the dynamic characterization of the vehicle and the optimization of ride comfort. The damping force of magnetorheological dampers exhibits strong nonlinear, hysteretic, and frequency-varying characteristics depending on the operating conditions and system state. There is a significant response delay between the establishment of the magnetic field and the yielding of the fluid, resulting in a strong time-dependent characteristic. In practical engineering applications, without the implantation of high-precision physical force sensors within the suspension system, the underlying controller struggles to obtain the actual output damping force of the damper. This leads to a significant deviation of the actual output force from the target force in traditional pure inverse open-loop control systems when faced with the complex hysteresis characteristics of magnetorheological dampers. Existing physical parameterization modeling or shallow neural networks cannot simultaneously solve the multivalued problem of inverse solution and the error divergence problem of open-loop control. To address this, this application proposes a data-driven modeling approach, constructing a forward model and an inverse model based on a CNN-LSTM architecture to replace the force sensor for measurement and the feedback controller for operation, respectively. By introducing a dynamic error compensation loop in the control layer, the complex spatiotemporal dependencies between system states are automatically learned, thus solving the technical problems of inaccurate force tracking and lack of closed-loop feedback under complex working conditions and improving the force tracking accuracy of the semi-active suspension.

[0023] In an exemplary embodiment, a data-driven forward and inverse model-based low-level closed-loop force tracking control method for a semi-active suspension is provided. This method is applied to a low-level closed-loop force tracking control architecture for a semi-active suspension system. The architecture includes: a sensor group, a filtering module, a state estimation-upper-level control module, and a low-level force tracking module. The low-level force tracking module encapsulates both a forward model and an inverse model. Both the forward and inverse models are obtained by training a CNN-LSTM architecture using a training sample set. The state estimation-upper-level control module includes a differential controller and an upper-level controller. The filtering module includes a low-pass filter and a high-pass filter.

[0024] like Figure 1 As shown, the semi-active suspension bottom-level closed-loop force tracking control method based on data-driven forward and inverse models includes: Step 101: Obtain the displacement state quantities of the semi-active suspension system at the current moment, as well as the input current and current constraint conditions of the damper. The damper is a magnetorheological damper. The displacement state quantities include: sprung mass displacement, unsprung mass displacement, and road surface displacement.

[0025] Step 102: Based on the displacement state variables of the semi-active suspension system at the current moment and the input current of the shock absorber, the piston rod velocity and the target damping force of the shock absorber are determined using the state estimation-upper-level control module. Step 102 specifically includes: estimating the motion state of the semi-active suspension system using a differential controller based on the displacement state variables to obtain the sprung mass velocity and unsprung mass velocity. The difference between the sprung mass velocity and the unsprung mass velocity is determined as the piston rod velocity. The upper-level controller determines the target damping force of the shock absorber using a control strategy based on the sprung mass velocity, unsprung mass velocity, filtered input current, and filtered displacement state variables. The control strategy is the optimal control strategy.

[0026] Step 103: Constrain the target damping force using the damping force boundary characteristic curve to obtain the target constrained damping force. The damping force boundary characteristic curve is constructed using a forward model.

[0027] The process of using the positive model of the vibration damper to constrain boundary characteristics is as follows: Step 1. Input the information sequence of the vibration damper's motion state (displacement and velocity of the piston rod) at the current moment and in the historical state into the forward model of the vibration damper.

[0028] Step 2. The forward model of the vibration damper calculates the maximum / minimum damping force in the current state based on the input sequence and the maximum / minimum input control current.

[0029] Step 3. Limit the target damping force to ensure that the target damping force value at the current moment is within the boundary constraints established in step 2 of the vibration damper.

[0030] The boundary characteristic curve represents the maximum and minimum damping force data points of the shock absorber under various motion states. Figure 3 (As shown by the yellow / green scatter plots), a closed curve enclosed in the velocity-damping force plane, as shown... Figure 3 As shown ( Figure 3 The static boundary characteristics of the black solid / dashed lines are commonly used in current applications, but they have shortcomings. This application uses a forward model of the vibration damper to construct a different model. Figure 3 (The dynamic boundary characteristic curve is shown by the red / blue line).

[0031] Step 104: Input the piston rod displacement, piston rod velocity, and input current into the forward model to obtain the estimated damping force of the shock absorber. The piston rod displacement is determined based on the displacement state variables.

[0032] Step 105: Based on the target constraint damping force and the estimated damping force, perform damping force tracking to determine the control damping command including the error term. The control damping command including the error term is: .

[0033] in, This is a control damping command that includes an error term. The target is constrained by damping force. This is the estimated value of the damping force at the current moment.

[0034] Step 106: Input the piston rod displacement and piston rod velocity, as well as the control damping command including the error term, into the inverse model to obtain the compensation control current of the damper.

[0035] Step 107: Use current constraint conditions to constrain the compensation control current to obtain the actual control current command of the vibration damper.

[0036] Before step 102, the method further includes: using a filtering module to filter both the input current and the displacement state quantity.

[0037] Before step 101, the following is also included: Construct the CNN-LSTM architecture.

[0038] The magnetorheological damper was subjected to characteristic tests to construct a training sample set. The training sample set consisted of time series data including piston rod displacement, piston rod velocity, damper input current, and damper output damping force. The damper output damping force was obtained using a damping force sensor in the sensor array.

[0039] Using piston rod displacement, piston rod velocity, and damper input current as inputs, and damper output damping force as output, the CNN-LSTM architecture is trained using a training sample set to obtain a forward model.

[0040] Using piston rod displacement, piston rod velocity, and damper output damping force as inputs, and damper input current as output, the CNN-LSTM architecture is trained using a training sample set to obtain the inverse model.

[0041] The CNN-LSTM architecture specifically includes a two-branch convolutional neural network and a long short-term memory network connected in sequence.

[0042] A dual-branch convolutional neural network consists of a first convolutional branch, a second convolutional branch, and a splicing layer.

[0043] The outputs of both the first and second convolutional branches are connected to the splicing layer.

[0044] The first convolutional branch is used to slide along the time dimension based on the current data to generate local features.

[0045] The second convolutional branch is used to slide along the time dimension based on historical data to generate local features.

[0046] The Long Short-Term Memory (LSTM) network consists of a first LSTM layer, a second LSTM layer, and an output layer connected in sequence.

[0047] The output layer of the splicing layer is connected to the input layer of the first LSTM layer.

[0048] The working principle of the first and second convolution branches is as follows: .

[0049] in, Represents the convolutional branch in model s l Output local features. This represents the activation function. Represents the convolutional branch in model s l The convolution kernel. This represents the input matrix. Represents the convolutional branch in model s l The bias term.

[0050] The working principles of the first LSTM layer and the second LSTM layer are as follows: ; ; ; ; In the formula, , and These represent the forget gate, input gate, and output gate, respectively. and Indicates a hidden unit; This represents the feature matrix output by the dual-branch convolutional neural network at time t; , and All are bias terms; and Indicates the cell state; Indicates candidate values; This indicates element-wise multiplication.

[0051] The output of the forward model is: .

[0052] in, This represents the output of the forward model at time t. Represents the forward model The weight matrix of the output layer. Represents the forward model The output of the hidden unit at time t in the second LSTM layer. Represents the forward model The bias term of the output layer.

[0053] The output of the inverse model is: .

[0054] in, This represents the output of the inverse model at time t. Representing the inverse model The weight matrix of the output layer. Representing the inverse model The output of the hidden unit at time t in the second LSTM layer. Representing the inverse model The bias term of the output layer.

[0055] The specific implementation steps are summarized as follows: Step 1: Based on MTS tests, displacement, velocity, current, and damping force data of the magnetorheological damper under multiple operating conditions are collected and preprocessed. Then, a data-driven network based on a combination of a dual-branch CNN and LSTM is constructed. The CNN extracts local spatiotemporal features of the sequence, and the LSTM is used to process long-range dynamic evolution. This allows for the construction of a forward observation model that estimates the actual damping force using velocity, displacement, and current as inputs, and an inverse dynamic model that calculates the control current using velocity, displacement, and damping force as inputs.

[0056] Step 2: Divide the preprocessed dataset into training, validation, and test sets to adjust network weights and monitor hyperparameters to prevent overfitting. After model convergence, initially evaluate the fitting accuracy on the test set. Finally, introduce new, unused working condition data for independent validation. By comparing the root mean square error between predicted and measured values, comprehensively evaluate the mechanical mapping performance and robustness of the forward and inverse models under unknown and complex dynamic conditions.

[0057] Step 3: The trained forward and inverse models are encapsulated into the chassis controller to build a real-time closed-loop force tracking system. The system uses the forward model to estimate the actual output damping force of the current shock absorber in real time and compares it with the expected damping force of the upper controller to calculate the control command with error terms. Then, this command and the real-time motion state of the suspension are input into the inverse model to solve for the optimal compensation control current. This current drives the magnetorheological shock absorber after being constrained by physical boundaries of 0-3A, thereby completing the high-precision closed-loop control at the bottom layer.

[0058] In one exemplary embodiment, a semi-active suspension bottom-level closed-loop force tracking control system based on a data-driven forward and inverse model is provided. The system applies the aforementioned semi-active suspension bottom-level closed-loop force tracking control method based on a data-driven forward and inverse model, such as... Figure 2 The system includes: The data acquisition module is used to acquire the displacement state quantities of the semi-active suspension system at the current moment, as well as the input current and current constraint conditions of the damper. The damper is a magnetorheological damper. The displacement state quantities include: sprung mass displacement, unsprung mass displacement, and road surface displacement.

[0059] The target damping force determination module is used to determine the piston rod speed and the target damping force of the shock absorber based on the displacement state of the semi-active suspension system at the current moment and the input current of the shock absorber, using the state estimation-upper-level control module.

[0060] The damping force constraint module is used to constrain the target damping force using the damping force boundary characteristic curve, thus obtaining the target constrained damping force. The damping force boundary characteristic curve is constructed using a forward model.

[0061] The damping force estimation module is used to input piston rod displacement, piston rod velocity, and input current into the forward model to obtain the estimated damping force of the vibration damper. The piston rod displacement is determined based on the displacement state variables.

[0062] The damping force tracking module is used to track the damping force based on the target constraint damping force and the estimated damping force, and to determine the control damping command that includes the error term.

[0063] The current estimation module is used to input the piston rod displacement and piston rod velocity, as well as the control damping command including the error term, into the inverse model to obtain the compensation control current of the vibration damper.

[0064] The current constraint module is used to constrain the compensation control current using current constraint conditions to obtain the actual control current command of the vibration damper.

[0065] The steps for building the semi-active suspension bottom-level closed-loop force tracking control system based on data-driven forward and inverse models in this embodiment are as follows: The first step is to build forward and inverse models driven by magnetorheological vibration dampers based on CNN-LSTM.

[0066] Based on the characteristic test of the magnetorheological vibration damper, the original characteristic data of the magnetorheological vibration damper under multiple working conditions were collected synchronously by high-frequency sensors, including control current, piston displacement, piston speed and corresponding actual damping force. After noise reduction and normalization of the collected sequence data, a training sample set with time-series correlation characteristics was constructed using sliding window technology, and the input and output variables were divided into different combinations for the forward and inverse models. This application employs a combined CNN and LSTM network as a data-driven model. The CNN uses a two-layer network with a one-dimensional convolutional layer sliding along the time dimension. The two-layer network performs local feature scanning and analysis on the input sequences at historical and current times, identifying and extracting spatiotemporal features from the original signal, and converting the original high-dimensional noisy time-series signal into a more discriminative high-level feature sequence. The LSTM receives the high-level feature sequence output by the CNN, and its internal precise gating mechanism enables the model to dynamically select and remember, learn and model how the current system output is affected by multiple spatiotemporal factors, thereby more comprehensively and accurately predicting the input-output characteristics of the magnetorheological damper under complex excitation, and constructing a data-driven model that can both perceive local mechanical morphology and understand long-term dynamic evolution.

[0067] This application employs CNN-LSTM-based data-driven forward and inverse models to obtain the actual damping force and control current output. For the forward model, the piston rod... v and x and the input current of the vibration damper I As input variables for this data-driven model, they can be represented as: .

[0068] In the formula: n represents the time span of the corresponding input sequence. Corresponding to the forward model, the input variable of the inverse model is the piston rod... v and x and the output damping force of the shock absorber F As input variables for this data-driven model, they can be represented as: .

[0069] This application employs a two-branch convolutional neural network (CNN), where branch one extracts features from current data and branch two extracts features from historical data. The convolution operation slides along the time dimension to generate local features, represented as: .

[0070] In the formula: It is an activation function. It is a convolution kernel. It is a bias term. s Indicates the model index.f and r These represent the forward model and the inverse model, respectively. Indicates the branch index, where 1 and 2 represent the first and second branches, respectively. The characteristics of the merged branch 1 and branch 2 are: .

[0071] The input sequence matrix, after undergoing multiple convolutional kernels and pooling operations, is encoded into a more advanced and compact temporal feature sequence, which is then input into the first LSTM unit. The network's internal state updates are as follows: .

[0072] .

[0073] .

[0074] .

[0075] In the formula, , and These represent the forget gate, input gate, and output gate, respectively. It is a hidden unit. , and It is a bias term. It is a unit state. It is a candidate value. This indicates element-wise multiplication, where the time feature sequences of the forward and inverse models will be input into the LSTM network of the corresponding models, respectively.

[0076] The feature sequence output from the first LSTM layer is input into the second LSTM layer, and then the state of the last time step is... Returning to the output layer, the signal is linearized again through a fully connected layer to calculate the model's output signal. For a forward model, the output is: .

[0077] The output of the inverse model is: .

[0078] In the formula: It is the weight matrix of the output layer. It is a bias term.

[0079] The second step is to train and evaluate the forward and inverse models based on CNN-LSTM.

[0080] The preprocessed multi-condition experimental data were divided into training, validation, and test sets according to a ratio of 70%:15%:15% and different input-output variable combinations for the forward and inverse models. The weights of the CNN convolutional kernels and LSTM units were adjusted using the training set, and hyperparameters were monitored using the validation set to prevent overfitting. After the model converged, its fitting accuracy for known working conditions was initially evaluated on the test set. Finally, to verify the model's robustness, independent testing was conducted using entirely new working condition data that was not used in the training. The RMSE deviation between predicted and measured values ​​was compared to evaluate the model's mechanical mapping performance under complex and unknown dynamic conditions.

[0081] The third step is to build a real-time closed-loop force tracking underlying control system based on the forward and inverse models of CNN-LSTM.

[0082] To achieve higher precision force tracking control at the lower level, specifically to calculate more accurately the deviation between the real-time output damping force of the shock absorber and the desired damping force of the upper-level controller under the current state, and to reduce this deviation by inputting current commands, a closed-loop force tracking lower-level control system is designed as follows: Figure 2 As shown, the trained forward and inverse models are first encapsulated in the vehicle chassis controller. The sensor acquisition module acquires the current motion state of the suspension system in real time, including the sprung mass displacement. Unsprung mass displacement and road surface displacement And read the input current from the vibration damper. To evaluate the performance of the damping force estimation model, a damping force sensor was added to the experimental rig. This sensor was used only as a performance evaluation reference and not as a system variable. The signals collected from the sensor were input to a filtering module containing low-pass and high-pass filters to eliminate sensor noise and other influences. These signals were then input to the state estimation and upper-level controller. and The speeds of the sprung mass and unsprung mass will be obtained separately through the differentiator calculation. and Information such as the suspension's motion state and the shock absorber's input current is collected and sent to the upper-level controller. This controller typically employs widely used engineering control strategies such as ceiling control and optimal control to calculate the shock absorber's output damping force at the current moment, based on the suspension system's motion state. .

[0083] Together with the upper-level controller, the desired target damping force required to suppress vehicle body vibration is calculated. and suspension dynamic speed and the current input current of the vibration damper Together, these parameters are input to the underlying force tracking module. Since the output force of the semi-active suspension is strictly limited by its real-time motion state, such as piston speed, the system cannot generate control forces exceeding boundary limits. Based on the constructed damping force estimation model, the maximum and minimum damping force boundary characteristic curves are further constructed using a forward model. The input parameters are maximum 3 A / minimum 0 A current and piston speed. and expectation Output force After boundary constraints. Additionally, the suspension dynamic speed... and the current input current of the vibration damper The input is fed into a CNN-LSTM forward model of the vibration damper, which outputs an estimate of the damper's damping force in the current state. The system compares the target force with the estimated force to calculate the control damping force command, which includes an error term. .

[0084] The controller inputs the above instructions, along with the current and historical displacement and velocity states, into the inverse model of the vibration damper to derive the optimal compensation control current. This current, after being constrained by a maximum current of 3A and a minimum current of 0A, becomes the actual control current command. The input is fed to the magnetorheological damper to drive the MR damper, thereby forming a bottom-level closed-loop force tracking control system to ensure the control accuracy of the semi-active suspension.

[0085] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0086] This application uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. In summary, the content of this specification should not be construed as a limitation of this application.

Claims

1. A semi-active suspension bottom-level closed-loop force tracking control method based on data-driven forward and inverse models, characterized in that, The method is applied to a low-level closed-loop force tracking control architecture for a semi-active suspension system; The architecture includes: a sensor group, a filtering module, a state estimation-upper-layer control module, and a lower-layer force tracking module; The underlying force tracking module encapsulates a forward model and an inverse model; both the forward model and the inverse model are obtained by training the CNN-LSTM architecture using a training sample set. The state estimation-upper-level control module includes: a differential controller and an upper-level controller; The method includes: The displacement state of the semi-active suspension system at the current moment, as well as the input current and current constraint conditions of the damper, are obtained; the damper is a magnetorheological damper; the displacement state includes: sprung mass displacement, unsprung mass displacement, and road surface displacement. Based on the displacement state of the semi-active suspension system at the current moment and the input current of the shock absorber, the piston rod speed and the target damping force of the shock absorber are determined using the state estimation-upper-level control module. The target damping force is constrained using the damping force boundary characteristic curve to obtain the target constrained damping force; the damping force boundary characteristic curve is constructed using a forward model. The piston rod displacement, piston rod velocity, and input current are all input into the positive model to obtain the estimated damping force of the shock absorber; the piston rod displacement is determined based on the displacement state quantity. Damping force tracking is performed based on the target constraint damping force and the estimated damping force value to determine the control damping command containing the error term; The piston rod displacement and piston rod velocity, as well as the control damping command including the error term, are all input into the inverse model to obtain the compensation control current of the damper. The compensation control current is constrained using current constraints to obtain the actual control current command for the vibration damper.

2. The semi-active suspension bottom-level closed-loop force tracking control method based on data-driven forward and inverse models according to claim 1, characterized in that, Before determining the piston rod speed and the target damping force of the shock absorber using the state estimation-upper-level control module based on the displacement state of the semi-active suspension system at the current moment and the input current of the shock absorber, the following steps are also included: The input current and the displacement state quantity are both filtered using a filtering module.

3. The semi-active suspension bottom-level closed-loop force tracking control method based on data-driven forward and inverse models according to claim 1, characterized in that, The filtering module includes a low-pass filter and a high-pass filter.

4. The semi-active suspension bottom-level closed-loop force tracking control method based on data-driven forward and inverse models according to claim 1, characterized in that, Based on the displacement state of the semi-active suspension system at the current moment and the input current of the shock absorber, the piston rod speed and the target damping force of the shock absorber are determined using the state estimation-upper-level control module, specifically including: Based on the displacement state quantity, the motion state of the semi-active suspension system is estimated using a differential controller to obtain the sprung mass velocity and the unsprung mass velocity. The difference between the sprung mass velocity and the unsprung mass velocity is determined as the piston rod velocity; The upper-level controller determines the target damping force of the vibration damper based on the sprung mass velocity, the unsprung mass velocity, the filtered input current, and the filtered displacement state quantity using a control strategy.

5. The semi-active suspension bottom-level closed-loop force tracking control method based on data-driven forward and inverse models according to claim 4, characterized in that, The control strategy described is the optimal control strategy.

6. The semi-active suspension bottom-level closed-loop force tracking control method based on data-driven forward and inverse models according to claim 1, characterized in that, The control damping command that includes an error term is: ； in, This is a control damping command that includes an error term; The target is constrained by damping force; This is the estimated value of the damping force at the current moment.

7. The semi-active suspension bottom-level closed-loop force tracking control method based on data-driven forward and inverse models according to claim 1, characterized in that, Before obtaining the displacement state of the semi-active suspension system at the current moment, the following steps are also included: Constructing a CNN-LSTM architecture; The magnetorheological damper was subjected to characteristic tests to construct a training sample set. The training sample set consisted of time series data including piston rod displacement, piston rod velocity, damper input current, and damper output damping force. The damper output damping force was obtained by a damping force sensor in the sensor array. Using piston rod displacement, piston rod velocity, and damper input current as inputs, and damper output damping force as output, the CNN-LSTM architecture is trained using the training sample set to obtain a forward model; Using piston rod displacement, piston rod velocity, and damper output damping force as inputs, and damper input current as output, the CNN-LSTM architecture is trained using the training sample set to obtain the inverse model.

8. The semi-active suspension bottom-level closed-loop force tracking control method based on data-driven forward and inverse models according to claim 7, characterized in that, The CNN-LSTM architecture specifically includes: a dual-branch convolutional neural network and a long short-term memory network connected in sequence; The dual-branch convolutional neural network includes a first convolutional branch, a second convolutional branch, and a splicing layer; The outputs of both the first convolutional branch and the second convolutional branch are connected to the splicing layer; The first convolutional branch is used to slide along the time dimension based on the current data to generate local features; The second convolutional branch is used to slide along the time dimension based on historical data to generate local features; The Long Short-Term Memory network includes: a first LSTM layer, a second LSTM layer, and an output layer connected in sequence; The output layer of the splicing layer is connected to the input layer of the first LSTM layer.

9. The semi-active suspension bottom-level closed-loop force tracking control method based on data-driven forward and inverse models according to claim 8, characterized in that, The working principle of the first convolutional branch and the second convolutional branch is as follows: ； in, Represents the convolutional branch in model s l Output local features; Indicates the activation function; Represents the convolutional branch in model s l The convolution kernel; Represents the input matrix; Represents the convolutional branch in model s l The bias term; The working principles of the first LSTM layer and the second LSTM layer are as follows: ； ； ； ； In the formula, , and These represent the forget gate, input gate, and output gate, respectively. and Indicates a hidden unit; This represents the feature matrix output by the dual-branch convolutional neural network at time t; , and All are bias terms; and Indicates the cell state; Indicates candidate values; This represents element-wise multiplication; The output of the forward model is: ； in, This represents the output of the forward model at time t; Represents the forward model The weight matrix of the output layer; Represents the forward model The output of the hidden unit at time t in the second LSTM layer; Represents the forward model The bias term of the output layer; The output of the inverse model is: ； in, This represents the output of the inverse model at time t; Representing the inverse model The weight matrix of the output layer; Representing the inverse model The output of the hidden unit at time t in the second LSTM layer; Representing the inverse model The bias term of the output layer.

10. A semi-active suspension bottom-level closed-loop force tracking control system based on a data-driven forward and inverse model, characterized in that, The system employs a semi-active suspension bottom-level closed-loop force tracking control method based on a data-driven forward and inverse model as described in any one of claims 1-9, and the system comprises: The data acquisition module is used to acquire the displacement state of the semi-active suspension system at the current moment, as well as the input current and current constraint conditions of the damper; the damper is a magnetorheological damper; the displacement state includes: sprung mass displacement, unsprung mass displacement, and road surface displacement; The target damping force determination module is used to determine the piston rod speed and the target damping force of the shock absorber based on the displacement state of the semi-active suspension system at the current moment and the input current of the shock absorber, using the state estimation-upper control module. The damping force constraint module is used to constrain the target damping force using the damping force boundary characteristic curve to obtain the target constrained damping force; the damping force boundary characteristic curve is constructed using a forward model; The damping force estimation module is used to input the piston rod displacement, the piston rod velocity, and the input current into the positive model to obtain the estimated damping force of the shock absorber; the piston rod displacement is determined based on the displacement state quantity. The damping force tracking module is used to track the damping force based on the target constraint damping force and the estimated damping force, and to determine the control damping command including the error term; The current estimation module is used to input the piston rod displacement and piston rod velocity, as well as the control damping command including the error term, into the inverse model to obtain the compensation control current of the vibration damper. The current constraint module is used to constrain the compensation control current using current constraint conditions to obtain the actual control current command of the vibration damper.

Images