Method for reconstructing parameters of spring damping system of electric vehicle under data loss

Abstract

The invention relates to a method for reconstructing parameters of a spring damping system of an electric vehicle under data loss, which comprises the following steps of: acquiring sensor data of the spring damping system of the electric vehicle, normalizing the data, and complementing the missing data by using part of known data; secondly, the dynamic constraint of a physical system is embedded into the neural network through a differential equation to form a differential neural network prediction parameter value; and finally, establishing a dichotomy neural network to evaluate whether an estimation result accords with the real data distribution, and if not, continuing optimization iteration until an optimal parameter estimation value of the spring damping system is obtained and output. According to the method, the problems of insufficient fusion, weak data loss processing capability, incapability of realizing online optimization and the like in the aspect of parameter estimation of the spring damping system in the prior art are solved, and the method has remarkable technical advantages and practical value.

Description

Technical Field

[0001] The invention relates to a method for reconstructing parameters of a spring damping system of an electric vehicle under data loss, and belongs to the field of new energy vehicle design and manufacturing. Background Art

[0002] In the suspension system, the vehicle's spring damping system connects the chassis frame to the wheels through an elastic structure to absorb vibration and impact from the ground. Its performance directly affects the vehicle's driving comfort and handling stability. Therefore, accurate estimation of the spring damping system parameters in the suspension system is of great significance for evaluating the reliability of the vehicle system and driving safety.

[0003] Existing estimation methods combine the noise variation characteristics and model adaptability, and adopt the fusion method of adaptive unscented Kalman filter and interactive multi-model to improve the adaptability and accuracy of estimation. The multi-model fusion is the direct fusion of the results of physical model and data model. However, this kind of method has shortcomings in dealing with the problem of system parameter estimation in the case of data loss. In the case of data loss, the adaptive unscented Kalman filter (UKF) cannot obtain complete observation data to update the state estimation. The interactive multi-model (IMM) estimates the system parameters based on the fusion of different models. When data is lost, the basis for selecting the appropriate model becomes incomplete. At the same time, if the direct fusion of the results of physical model and data model is adopted, the results provided by both models may be affected in the case of data loss; for example, the physical model may produce inaccurate results due to the lack of some input data (due to data loss), and the data model may also output biased results due to incomplete training data; directly fusing these two inaccurate results will make the final fusion result unreliable and unable to accurately estimate the parameters of the spring damping system.

[0004] The invention with application number 202411336607.7 discloses a method, equipment, medium and product for dynamic reconstruction of suspension parameters of a vehicle, which relates to the field of vehicle design and manufacturing. The method includes: establishing a state space model of the vehicle suspension based on a pre-constructed road suspension coupling model, and iteratively processing the state space model using a nonlinear function probability distribution approximate filtering algorithm to obtain a priori prediction values ​​of the suspension parameters; establishing a new information variance based on the measured variable sequence collected by the sensors on the vehicle and the observed variable sequence generated during the iterative processing, and using the new information variance to correct the priori prediction value to obtain the posterior prediction value of the suspension parameter; using an interactive multi-model to weight the posterior prediction value and the prediction value output by the pre-trained suspension parameter prediction model to obtain the suspension parameter prediction value. Although the aforementioned application provides a fusion method, it is obvious that it also simply fuses the output results and cannot achieve online evaluation of the accuracy of the estimation results.

[0005] Therefore, it is necessary to design a new method for reconstructing the spring-damper system parameters of electric vehicles based on data loss, which can also estimate the spring-damper system parameters well under harsh road conditions, complex traffic environments and other working conditions with a higher risk of data loss, thereby improving the versatility and practicality of this technology in the actual vehicle engineering field. Summary of the invention

[0006] The present invention provides a method for reconstructing the parameters of the spring damping system of an electric vehicle under data loss, which solves the problems of insufficient fusion and weak ability to handle data loss in the prior art in estimating the parameters of the spring damping system, and has significant technical advantages and practical value.

[0007] The technical solution adopted by the present invention to solve its technical problem is:

[0008] A method for reconstructing parameters of a spring damping system of an electric vehicle under data loss, comprising the following steps:

[0009] Step S1, obtaining a road input signal, sprung displacement and velocity information with data loss, and unsprung displacement and velocity information with data loss through a sensor installed on a spring damping system of an electric vehicle, normalizing the three types of information by means, constructing a reconstruction filter to complete the generation of missing data using part of the known data, and realizing data filling;

[0010] Step S2, establishing a differential equation of the spring damping system, based on the generated and completed real data obtained in step S1, using the constructed differential neural network to predict the parameter values ​​of the spring damping system of the electric vehicle, and outputting the parameter prediction values;

[0011] Step S3, constructing a binary classification neural network, inputting the parameter prediction value outputted by step S2 and the real data completed by step S1 into the binary classification neural network to judge the authenticity of the parameter prediction value, and optimizing the differential neural network according to the judgment result, until the parameter prediction value outputted by the differential neural network meets the termination condition, and outputting the latest differential neural network parameter prediction value;

[0012] Furthermore, in step S1, the steps of implementing data filling are:

[0013] Step S11, define the road surface input signal as Z r , the sprung displacement information is the vertical displacement z of the sprung mass s , the sprung velocity information is the velocity of the sprung mass The unsprung displacement is the vertical displacement of the unsprung mass z w , the unsprung velocity information is the velocity of the unsprung mass The road input signal, the vertical displacement of the sprung mass, the speed of the sprung mass, the vertical displacement of the unsprung mass, and the speed of the unsprung mass are normalized, that is,

[0014]

[0015] Step S12, define the data sequence obtained by the sensor as X n (obs), the missing data sequence is X n (miss), design a three-layer neural network, where the input layer is Contains 6 neurons, the middle layer is 1 layer, and the output layer is Contains 4 neurons, namely

[0016] X n (miss) = G(X n (obs),γ) (2)

[0017] In formula (2), γ is random noise and G is the neural network;

[0018] Step S13, from the obtained data sequence X n Select the displacement and velocity data sequence in (obs) The completed data set is

[0019]

[0020] Furthermore, in step S2, the step of outputting the parameter prediction value is:

[0021] Step S21, establish the spring damping system differential equation, specifically:

[0022]

[0023] In formula (4) and formula (5), M is the wheel mass, is the acceleration of the sprung mass, k s is the spring stiffness, z s is the vertical displacement of the sprung mass, z w is the vertical displacement of the unsprung mass, c s is the damping coefficient of the suspension system, is the velocity of the sprung mass, is the velocity of the unsprung mass, m is the sprung mass, k t is the stiffness of the tire, z r Input for road surface;

[0024] Step S22, initializing the neural network;

[0025] Step S23, based on step S21, calculate the loss function based on physical constraints for each training sample, and the calculation formula is:

[0026]

[0027] In formula (6), is the loss function based on physical constraints, N is the total number of training samples;

[0028] Step S24, calculating the data-driven loss function according to the sprung displacement data in the completed data set, the calculation formula is:

[0029]

[0030] In formula (7), is the data-driven loss function, z w_pre is the predicted vertical displacement of the unsprung mass, z s_pre is the predicted vertical displacement of the sprung mass;

[0031] Step S25, based on the physical constraint-based loss function obtained in step S23 and the data-driven loss function obtained in step S24, a total loss function is calculated, and the calculation formula is:

[0032]

[0033] In formula (8), is the total loss function, λ1 is the weight loss hyperparameter;

[0034] Step S26, using a gradient descent optimization algorithm to update the parameters of the neural network to minimize the total loss function, training until the total loss function converges, and outputting parameter prediction values, the parameter prediction values ​​including the spring stiffness estimation value Suspension system damping coefficient estimate Estimated tire stiffness And the predicted values ​​of displacement and velocity of the spring damping system of electric vehicles

[0035] Furthermore, in step S3, the steps of binary classification evaluation and optimization are:

[0036] Step S31, establish a binary classification neural network, where the input is the parameter prediction value output by the differential neural network in step S2 and the data set completed in step S1, and the output is the probability value D, and D∈(0,1), then

[0037]

[0038] Step S32, fix the differential neural network, optimize the binary classification neural network, the binary classification neural network is With the completed dataset X n (com) comparison, training, so that the binary classification neural network can accurately distinguish between real data and generated data;

[0039] Step S33, fix the binary classification neural network, optimize the differential neural network, and minimize the total loss function The new total loss function obtained by optimization is:

[0040]

[0041] Step S34, repeating steps S32-S33, the differential neural network is iteratively trained until the output of the binary classification neural network is close to the real data. When the binary classification neural network cannot further distinguish the generated data from the real data, the training is stopped and the system parameter estimation result of the differential neural network at this time is used as the optimal output.

[0042] Through the above technical solution, compared with the prior art, the present invention has the following beneficial effects:

[0043] 1. The method for reconstructing the parameters of the spring damping system of an electric vehicle under data loss provided by the present invention innovatively considers embedding physical knowledge in the training process of the data-driven method, and integrates the physical characteristics of the spring damping system based on the principle of mechanics. Through such a nested approach, the regularity of physical knowledge and the fitting ability of the data-driven method to complex situations can be fully utilized, making the parameter estimation more accurate and reasonable, thus filling the gap in the deep integration and collaborative optimization of physical and data methods in the prior art;

[0044] 2. The electric vehicle spring damping system parameter reconstruction method under data loss provided by the present invention can perform quality evaluation on the parameter estimation results obtained after the physical and data methods are nested in real time, and establish a result evaluation method to realize the nesting and online optimization of the physical and data methods, further enhancing the scientificity and reliability of the entire estimation process;

[0045] 3. The method for reconstructing the parameters of the spring damping system of an electric vehicle under data loss provided by the present invention is a data-driven method combined with physical knowledge, which can better understand and adapt to the changes of the system under different working conditions according to the physical essence. Physical knowledge provides a priori constraints for the data-driven method, reducing the estimation errors caused by factors such as data noise and outliers;

[0046] 4. The method for reconstructing the parameters of the spring damping system of an electric vehicle under missing data provided by the present invention and the method for filling the missing measurement data established can effectively deal with the data loss situations occurring under different working conditions, greatly broadening its scope of application and improving the versatility and practicality of the technology in the actual vehicle engineering field. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0048] Figure 1 It is a flow chart of a method for reconstructing parameters of a spring damping system of an electric vehicle under data loss provided by the present invention;

[0049] Figure 2 It is a schematic diagram of a spring damping system according to a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0050] The present invention will now be described in further detail with reference to the accompanying drawings. In the description of the present application, it should be understood that the orientation or positional relationship indicated by the terms "left side", "right side", "upper part", "lower part", etc. is based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and "first", "second", etc. do not indicate the importance of the components, and therefore cannot be understood as a limitation on the present invention. The specific dimensions used in this embodiment are only for illustrating the technical solution by example, and do not limit the scope of protection of the present invention.

[0051] As described in the background technology, the existing multi-model fusion used for spring damping system parameter estimation is direct fusion. This fusion method causes unreliable fusion results in dealing with data loss problems and cannot accurately estimate the parameters of the spring damping system.

[0052] In order to solve the above problems, this application provides a method for reconstructing the parameters of the spring damping system of an electric vehicle under data loss, such as Figure 1 As shown in the figure, it is the overall flow chart of the whole reconstruction method. The whole reconstruction method first collects the sensor data of the spring damping system of the electric vehicle, normalizes the data, and uses some known data to fill in the missing data; secondly, the dynamic constraints of the physical system are embedded in the neural network through differential equations to form a differential neural network to predict parameter values; finally, a binary classification discriminant method is established to evaluate whether the estimation results are consistent with the actual distribution of the data, and to obtain the optimal spring damping system parameter estimation values.

[0053] Next, the above method is specifically described, including the following steps:

[0054] Step S1, obtaining a road input signal, sprung displacement and velocity information with data loss, and unsprung displacement and velocity information with data loss through a sensor installed on a spring damping system of an electric vehicle, normalizing the three types of information by means, constructing a reconstruction filter to complete the generation of missing data using part of the known data, and realizing data filling;

[0055] The sensor data of the spring damping system collected in this step is the basis of the entire parameter reconstruction. The road input signal, the displacement and velocity information on the sprung with data loss, and the displacement and velocity information on the unsprung with data loss are all types of information that can reflect the actual operating status of the spring damping system, providing a basis for subsequent analysis and parameter estimation. Due to the limitations of the sensor's own accuracy, complex external electromagnetic interference, or occasional data transmission failures, the collected data often have problems such as inconsistent dimensions and large numerical ranges. Therefore, in order to avoid subsequent calculation problems caused by large differences in data scales, this step performs mean normalization on the three types of information, which helps to improve the stability and accuracy of model training and parameter estimation, and also facilitates the comprehensive processing of different data.

[0056] The specific steps to implement data filling are:

[0057] Step S11, when performing the data filling operation, using part of the known data to fill in the missing data is one of the innovations of the present application. For example, through the speed change trend of the vehicle and the deformation of the spring at some moments, the damping force data missing at the corresponding moment can be inferred through the dynamic relationship of the spring damping system. Therefore, the present application provides the following Figure 2 The spring damping system diagram is shown in Figure 1, and the road input signal is defined as Z r , the sprung displacement information is the vertical displacement z of the sprung mass s , the sprung velocity information is the velocity of the sprung mass The unsprung displacement is the vertical displacement of the unsprung mass z w , the unsprung velocity information is the velocity of the unsprung mass The road input signal, the vertical displacement of the sprung mass, the speed of the sprung mass, the vertical displacement of the unsprung mass, and the speed of the unsprung mass are normalized, that is,

[0058]

[0059] As a major innovation of this application, using part of the known data to complete the missing data becomes a key link, that is, step S12, defining the data sequence obtained by the sensor as X n (obs), the missing data sequence is X n (miss), design a three-layer neural network, where the input layer is Contains 6 neurons, the middle layer is 1 layer, and the output layer is Contains 4 neurons, namely

[0060]

[0061] In formula (2), γ is random noise, and G is a neural network. The specific calculation method of the neural network is as follows:

[0062] From the input layer to the hidden layer, h1 = σ(W1[z,x obs ]+b1),[z,x obs ] is the input concatenated vector, W1 is the weight of the first layer, b1 is the bias of the first layer, and σ is the activation function (such as ReLU);

[0063] From hidden layer to output layer W2 is the weight of the output layer, and b2 is the bias of the output layer.

[0064] Step S13, from the obtained data sequence X n Select the displacement and velocity data sequence in (obs) The completed data set is

[0065]

[0066] At this point, the completed data set is obtained, ensuring the integrity and continuity of the data, laying a solid foundation for subsequent precise analysis.

[0067] The next step S2 is to establish a differential equation of the spring damping system, and based on the completed real data generated in step S1, use the constructed differential neural network to predict the parameter values ​​of the spring damping system of the electric vehicle, and output the parameter prediction values;

[0068] As another innovation of the present application, step S2 uses differential equations to guide the neural network. Compared with traditional neural networks, which simply rely on a large amount of data for learning and fitting when performing parameter estimation, resulting in a lack of effective integration of the inherent laws of the physical system, the neural network of the present application continuously adjusts the weights to minimize the difference between the predicted results and the actual data. It is also necessary to ensure that the system state corresponding to the parameters and estimated results of its output satisfies the differential equations of the spring-damping system, and embeds the dynamic constraints of the physical system into the neural network through differential equations to form a differential neural network to predict parameter values. Through continuous training, the prediction results of the neural network are gradually approached to the theoretical dynamic response specified by the differential equations. This can effectively improve the accuracy of parameter estimation and make the estimation results more in line with actual physical conditions.

[0069] The specific steps are as follows: Step S21, establish the spring damping system differential equation, which is

[0070]

[0071] In formula (4) and formula (5), M is the wheel mass, is the acceleration of the sprung mass, k s is the spring stiffness, z s is the vertical displacement of the sprung mass, z w is the vertical displacement of the unsprung mass, c s is the damping coefficient of the suspension system, is the velocity of the sprung mass, is the velocity of the unsprung mass, m is the sprung mass, k t is the stiffness of the tire, z r Input for road surface;

[0072] Step S22, initializing the neural network; the neural network structure is a multi-layer feedforward neural network with an appropriate number of layers and neurons;

[0073] Step S23, for each training sample, calculate the loss function based on physical constraints to ensure that the network output conforms to the spring damping system differential equation established in step S21. The calculation formula of the loss function based on physical constraints is:

[0074]

[0075] In formula (6), is the loss function based on physical constraints, N is the total number of training samples;

[0076] Step S24, calculating the data-driven loss function according to the sprung displacement data in the completed data set, the calculation formula is:

[0077]

[0078] In formula (7), is the data-driven loss function, z w_pre is the predicted vertical displacement of the unsprung mass, z s_pre is the predicted vertical displacement of the sprung mass;

[0079] Regarding step S23 and step S24, it should be emphasized that the traditional neural network only includes a data-driven loss function when training, while the present application also includes a loss function based on physical constraints, so that the entire training from data input to output ensures that the real physical constraint requirements are met, and effectively improves the accuracy of parameter estimation; when calculating the data-driven loss function, there are multiple options. In order to ensure the accuracy of data measurement, the present application selects the more accurate displacement data among all parameters for calculation;

[0080] Continuing to step S25, based on the physical constraint-based loss function obtained in step S23 and the data-driven loss function obtained in step S24, a total loss function is calculated, and the calculation formula is:

[0081]

[0082] In formula (8), is the total loss function, λ1 is the weight loss hyperparameter;

[0083] Step S26, using a gradient descent optimization algorithm (such as Adam or SGD) to update the parameters of the neural network to minimize the total loss function. In each round of training, the neural network updates its parameters based on physical constraints and data feedback. The training process will continue for several epochs until the total loss function converges, at which time the differential neural network will output parameter predictions, which include spring stiffness estimates. Suspension system damping coefficient estimate Estimated tire stiffness And the predicted values ​​of displacement and velocity of the spring damping system of electric vehicles

[0084] In this step, we can see that the differential neural network is a three-layer neural network. The difference from step S12 is that the input here is the data set X completed in step S1. n (com), the output is the predicted values ​​of the aforementioned parameters, including the estimated value of the spring stiffness Suspension system damping coefficient estimate Estimated tire stiffness And the predicted values ​​of displacement and velocity of the spring damping system of electric vehicles 7 neurons included.

[0085] Since in the previous parameter estimation process, whether it is data completion or the use of differential equations to guide the neural network, there may be certain errors or uncertainties. As pointed out in the background technology, although the existing technology involves the inspiration of related model fusion, its fusion ability and data processing ability cannot be directly obtained online. Therefore, the present application continues to provide an effective method to determine whether the final estimation result is consistent with the true distribution of the data, and optimizes the estimation accuracy online through continuous iteration and training, which is the next step S3, constructing a binary classification neural network, and inputting the parameter prediction value output by step S2 and the real data after completion in step S1 into the binary classification neural network to judge the authenticity of the parameter prediction value, and optimizing the differential neural network according to the judgment result. If the simulated data is highly similar to the real data in these feature dimensions and is close to the pre-set similarity threshold, it is judged to be consistent with the real distribution, and the corresponding estimated parameters are retained; otherwise, it is judged to be inconsistent, triggering a new round of parameter adjustment and optimization. Through continuous judgment and optimization cycles, until the parameter prediction value output by the differential neural network meets the termination condition, the latest differential neural network parameter prediction value is output, which is the parameter estimation value that best fits the actual working condition and can make the spring damping system performance reach the optimal value;

[0086] The specific steps of binary classification evaluation and optimization are:

[0087] Step S31, establish a binary classification neural network, where the input is the parameter prediction value output by the differential neural network in step S2 and the data set completed in step S1, and the output is the probability value D, and D∈(0,1), then

[0088]

[0089] The binary classification neural network designed here is also a three-layer neural network. Because the parameter prediction value needs to be compared with the actual data, its input includes the data set X completed in step S1 n (com) and the displacement and velocity output in the differential neural network, a total of 8 neurons, 1 layer in the middle layer, and 1 neuron in the output layer, which is the probability value D;

[0090] Step S32, fix the differential neural network, optimize the binary classification neural network, the binary classification neural network is With the completed dataset X n (com) comparison, training, so that the binary classification neural network can accurately distinguish between real data and generated data;

[0091] Step S33, fix the binary classification neural network, optimize the differential neural network, and minimize the total loss function The new total loss function obtained by optimization is:

[0092]

[0093] Step S34, repeating steps S32-S33, the differential neural network is iteratively trained until the output of the binary classification neural network is close to the real data, and when the binary classification neural network cannot further distinguish the generated data from the real data, the training is stopped, and the system parameter estimation result of the differential neural network at this time is used as the optimal output;

[0094] Obviously, through the binary classification method, the results of multiple parameter estimations can be screened, those estimates that do not conform to the actual distribution of the data can be removed, and the optimal spring damping system parameter estimates can be selected, ensuring that the final parameter estimates not only meet the physical laws and data characteristics in theory, but also are closest to the actual situation in actual application scenarios and can most accurately reflect the state of the spring damping system, thereby improving the effectiveness and practicality of the entire parameter reconstruction method.

[0095] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those generally understood by those skilled in the art to which this application belongs. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with the meanings in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless defined as herein.

[0096] The meaning of "and / or" described in this application means that the situations where each exists alone or both exist at the same time are included.

[0097] The term “connection” as used in this application may mean a direct connection between components or an indirect connection between components via other components.

[0098] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A method for reconstructing parameters of a spring damping system of an electric vehicle under data loss, characterized in that: The following steps are involved: Step S1, obtaining a road input signal, sprung displacement and velocity information with data loss, and unsprung displacement and velocity information with data loss through a sensor installed on a spring damping system of an electric vehicle, normalizing the three types of information by means, constructing a reconstruction filter to complete the generation of missing data using part of the known data, and realizing data filling; Step S2, establishing a differential equation of the spring damping system, based on the generated and completed real data obtained in step S1, using the constructed differential neural network to predict the parameter values ​​of the spring damping system of the electric vehicle, and outputting the parameter prediction values; Step S3, construct a binary classification neural network, input the parameter prediction value output by step S2 and the real data completed by step S1 into the binary classification neural network to judge the authenticity of the parameter prediction value, and optimize the differential neural network according to the judgment result, until the parameter prediction value output by the differential neural network meets the termination condition, and output the latest differential neural network parameter prediction value.

2. The method for reconstructing parameters of spring damping system of electric vehicle under data loss according to claim 1, characterized in that: In step S1, the steps to achieve data filling are: Step S11, define the road surface input signal as Z r , the sprung displacement information is the vertical displacement z of the sprung mass s , the sprung velocity information is the velocity of the sprung mass The unsprung displacement is the vertical displacement of the unsprung mass z w , the unsprung velocity information is the velocity of the unsprung mass The road input signal, the vertical displacement of the sprung mass, the speed of the sprung mass, the vertical displacement of the unsprung mass, and the speed of the unsprung mass are normalized, that is, Step S12, define the data sequence obtained by the sensor as X n (obs), the missing data sequence is X n (miss), design a three-layer neural network, where the input layer is Contains 6 neurons, the middle layer is 1 layer, and the output layer is Contains 4 neurons, namely X n (miss)=G(X n (obs),c) (2) In formula (2), γ is random noise and G is the neural network; Step S13, from the obtained data sequence X n Select the displacement and velocity data sequence in (obs) The completed data set is 3. The method for reconstructing parameters of spring damping system of electric vehicle under data loss according to claim 2, characterized in that: In step S2, the steps of outputting parameter prediction values ​​are: Step S21, establish the spring damping system differential equation, specifically: In formula (4) and formula (5), M is the wheel mass, is the acceleration of the sprung mass, k s is the spring stiffness, z s is the vertical displacement of the sprung mass, z w is the vertical displacement of the unsprung mass, c s is the damping coefficient of the suspension system, is the velocity of the sprung mass, is the velocity of the unsprung mass, m is the sprung mass, k t is the stiffness of the tire, z r Input for road surface; Step S22, initializing the neural network; Step S23, based on step S21, calculate the loss function based on physical constraints for each training sample, and the calculation formula is: In formula (6), is the loss function based on physical constraints, N is the total number of training samples; Step S24, calculating the data-driven loss function according to the sprung displacement data in the completed data set, the calculation formula is: In formula (7), is the data-driven loss function, z w_pre is the predicted vertical displacement of the unsprung mass, z s_pre is the predicted vertical displacement of the sprung mass; Step S25, based on the physical constraint-based loss function obtained in step S23 and the data-driven loss function obtained in step S24, a total loss function is calculated, and the calculation formula is: In formula (8), is the total loss function, λ1 is the weight loss hyperparameter; Step S26, using a gradient descent optimization algorithm to update the parameters of the neural network to minimize the total loss function, training until the total loss function converges, and outputting parameter prediction values, the parameter prediction values ​​including the spring stiffness estimation value Suspension system damping coefficient estimate Estimated tire stiffness And the predicted values ​​of displacement and velocity of the spring damping system of electric vehicles 4. The method for reconstructing parameters of spring damping system of electric vehicle under data loss according to claim 3, characterized in that: In step S3, the steps of binary classification evaluation and optimization are: Step S31, establish a binary classification neural network, where the input is the parameter prediction value output by the differential neural network in step S2 and the data set completed in step S1, and the output is the probability value D, and D∈(0,1), then Step S32, fix the differential neural network, optimize the binary classification neural network, the binary classification neural network is and the completed dataset X n (com) comparison, training, so that the binary classification neural network can accurately distinguish between real data and generated data; Step S33, fix the binary classification neural network, optimize the differential neural network, and minimize the total loss function The new total loss function obtained by optimization is: Step S34, repeating steps S32-S33, the differential neural network is iteratively trained until the output of the binary classification neural network is close to the real data. When the binary classification neural network cannot further distinguish the generated data from the real data, the training is stopped and the system parameter estimation result of the differential neural network at this time is used as the optimal output.

Images