Suspension k&c performance data generalization method and system

Abstract

The application discloses a suspension K&C performance data generalization method and system. The suspension parameter K&C performance data generalization method comprises the following steps: collecting suspension K&C test data, and constructing a "suspension parameter-K&C performance" original sample library; selecting key variables and setting physical constraint intervals according to the "suspension parameter-K&C performance" correlation logic; expanding the suspension parameter space by means of multidimensional interpolation, and generating large-scale virtual samples by predicting corresponding K&C performance data through a neural network model to realize database generalization; and sequentially performing geometric constraint checking, mechanical constraint checking and AI generalization consistency checking, and removing unreasonable samples to finally form a reliable generalization database. The application takes inherent suspension parameters as the core, and combines the cross-scene database generalization method of the interpolation method and the AI modeling method, so that a small sample can be used to pry a large database of multiple vehicle types.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of vehicle engineering and intelligent data fusion, in particular to a suspension K&C performance data generalization method and system combining multi-dimensional interpolation and AI mapping. BACKGROUND

[0002] In recent years, the automobile industry is in a critical stage of intelligentization and platformization transformation, and the suspension K&C (Kinematic, geometric kinematic characteristics and Compliance, elastic kinematic characteristics) performance database has gradually become a core factor affecting the development efficiency and performance limit of the chassis. However, the existing technology still has many deficiencies. First, the test cost is high and the reusability is low: the traditional K&C test highly depends on expensive equipment, the cost of complete test for a single vehicle model is high, and the results are only suitable for the current vehicle model, and there is lack of reusability between different vehicle models and suspension types. Second, the cross-vehicle generalization ability is insufficient: the existing database is mostly limited to single vehicle model, single suspension and single scene, lacks mapping mechanism of suspension inherent parameters and K&C performance, and cannot effectively deduce the characteristics of new vehicle model, resulting in prolonged research and development cycle. Third, the AI data support is limited: the sample size of the database is biased, and the data scene is single, which is easy to cause overfitting of AI diagnosis and optimization model, and then affect the diagnosis accuracy and optimization reliability. SUMMARY

[0003] The purpose of the present application is to overcome the deficiencies and shortcomings of the prior art, and to provide a suspension K&C performance data generalization method and system combining multi-dimensional interpolation and AI mapping, which is a cross-scene database generalization method for multi-vehicle suspension K&C performance, and is suitable for performance development and optimization of MacPherson, double wishbone and multi-link suspensions.

[0004] One purpose of the present application is to provide a suspension K&C performance data generalization method, comprising the steps of:

[0005] Collecting suspension K&C test data, removing outliers, extracting key feature parameters of the suspension by normalization and PCA-random forest method, and constructing a "suspension parameter-K&C performance" original sample library;

[0006] Selecting key variables and setting physical constraint intervals according to the "suspension parameter-K&C performance" correlation logic;

[0007] By expanding the suspension parameter space through multidimensional interpolation and predicting the corresponding K&C performance data through a neural network model, a large-scale virtual sample is generated to achieve database generalization. The prediction of the corresponding K&C performance data through the neural network model is based on the virtual suspension parameter samples generated by interpolation. A physical constraint-enhanced deep neural network is used for K&C performance prediction. The generated virtual suspension parameters are input into the trained model in batches, and the corresponding virtual K&C performance data is output. By integrating physical laws and data-driven modeling, the mapping from virtual parameters to reliable K&C performance is realized.

[0008] By sequentially passing geometric constraint verification, mechanical constraint verification, and AI generalization consistency verification, unreasonable samples are eliminated, ultimately forming a reliable generalization database.

[0009] The geometric constraint verification is based on suspension kinematics theory, constructs multi-parameter linkage geometric verification rules, eliminates samples that do not conform to kinematic laws, and ensures that the K&C performance index conforms to physical laws as key variables change.

[0010] Based on the aforementioned geometric constraint verification, and combining the principles of mechanics of materials and suspension structure mechanics, a rigidity verification system for mechanical properties is established to perform mechanical constraint verification and ensure the mechanical rationality of the extended data.

[0011] After completing the constraint verification, an AI generalization consistency check is performed. Multiple sets of measured data from vehicles on the same platform that did not participate in the training are selected as verification samples. The AI ​​prediction values ​​are compared with the measured values ​​to determine whether the data generalization is successful. If it is unsuccessful, the AI ​​model parameters are adjusted or interpolation samples are added.

[0012] The suspension K&C test data was obtained through multi-source collaborative data acquisition. The K&C performance data acquisition covered toe angle gradient, camber angle gradient, roll stiffness, vertical stiffness, and wheel center displacement-force curve. The suspension parameter acquisition included the three-dimensional coordinates of suspension hard points, spring stiffness, stabilizer bar diameter, and bushing stiffness. The acquisition covered at least different configuration models on the same platform, with each model obtaining multiple sets of valid working condition samples including no load, full load, and different roll angles. The total number of original samples met expectations.

[0013] Among them, a sliding window dynamic 3 is adopted. Methods to remove outliers from the data; using multiple consecutive samples as a sliding window, the mean of the data within the window is calculated in real time. with standard deviation Only retain those that meet the requirements. Data; among which This is a dynamic correction factor. This is data to be processed.

[0014] Based on the differences in suspension parameters and K&C indicators among different vehicle models, an industry benchmark mapping method was used to normalize the data after outlier removal to the [0,1] interval, eliminating the incomparability of cross-vehicle data. The normalized data is obtained by the following formula:

[0015] ;

[0016] in, The data to be processed, i.e., the original values. , These are the extreme values ​​for the corresponding indicators of vehicles in the same class. This is the industry average. These are the weighting coefficients. This is the normalized data.

[0017] The PCA-random forest method for extracting key suspension feature parameters includes calculating the covariance matrix and performing eigenvalue decomposition through principal component analysis, calculating feature importance scores through the random forest algorithm, and determining the final key suspension feature parameters based on the scores.

[0018] The selection of key variables and setting of physical constraint ranges based on the "suspension parameters – K&C performance" correlation logic includes:

[0019] Define the input-output mapping relationship, clarify the corresponding logic between suspension parameters and K&C performance, and ensure that the mapping relationship covers all dimensions of suspension parameters and K&C performance;

[0020] ;

[0021] in, The input parameter vector contains the three-dimensional coordinates of the hard points of the front and rear wheels and the spring stiffness. stabilizer bar diameter Bushing stiffness , The output K&C performance index vector includes the front head angle gradient. outward tilt gradient Lateral stiffness Vertical stiffness Wheel center displacement-force curve ; Subscript Represents the left rear wheel, subscript Represents the right rear wheel; subscript Represents the left front wheel, subscript Represents the right front wheel;

[0022] Establish three levels of constraints for input parameters and output performance metrics:

[0023] Limit range: The boundary that the set parameters must absolutely not exceed;

[0024] Safe range: Referencing the vehicle's suspension design specifications, set the parameters within the engineering feasible range;

[0025] Optimization range: Based on the parameter range corresponding to the optimal performance of the original samples, a performance-sensitive range is set;

[0026] All constraint intervals are connected by the following inequality, forming a closed-loop constraint system:

[0027] ;

[0028] Among them, Corresponding to the extreme values ​​at the boundaries of the safe interval, Corresponding to the optimization interval boundary extreme value, The corresponding extreme value at the boundary of the limiting interval, Indicates the first One input parameter.

[0029] The aforementioned expansion of the suspension parameter space using multidimensional interpolation involves employing the radial basis function (RBF) interpolation method to fill the high-dimensional space formed by the suspension parameters, generating new parameter combinations, and performing interpolation physical verification to form a virtual suspension parameter library, including:

[0030] Using key feature parameters as the dimension, an m-dimensional parameter space is constructed. The original sample parameters form a discrete point set {X1, X2, ..., X} in this m-dimensional parameter space. n}, where n is the number of original sample parameters; for the region in the m-dimensional parameter space not covered by the original sample parameters, new suspension parameter combinations are generated through RBF interpolation. :

[0031] ;

[0032] in, For radial basis functions, These are the interpolation weighting coefficients. For the first One original sample parameter, For new parameters Compared with the original sample parameters The Euclidean distance;

[0033] Physical verification of interpolation parameters is performed to ensure that the new parameters meet the three-level constraints of the limit range, safety range, and optimization range. 10% of the interpolation parameters are randomly selected and verified through multibody dynamics simulation to ensure that there is no motion interference in the hard point coordinates, the basic K&C performance trend is reasonable, the verification pass rate is ≥98%, and a virtual suspension parameter library is formed.

[0034] The construction and training of the physical constraint-enhanced deep neural network includes:

[0035] Based on the original measured samples, a training dataset with strong correlation between suspension parameters and K&C performance was constructed: during the model training phase, five-fold cross-validation was adopted, and the training set was randomly divided into 5 subsets. Four subsets were used for training and one subset for validation in turn. Finally, the average of the model parameters from the 5 training iterations was taken as the model parameters.

[0036] An improved residual attention neural network is employed to enhance the model's nonlinear fitting ability and strengthen its sensitivity to key parameters. The improved residual attention neural network includes:

[0037] Input layer: Its number of neurons is equal to the dimension of the feature parameters, and it receives the suspension parameter vector;

[0038] Hidden layer: Employs a 3-layer fully connected structure with residual connections, where each layer's output... ;in, For layer input, It is a nonlinear mapping function. The weights are used as the basis for calculation; Batch Normalization and Dropout are added after each layer of the residual connection to suppress overfitting.

[0039] The activation function chosen is Leaky-ReLU;

[0040] Insert an attention mechanism between the input layer and the first hidden layer:

[0041] Attention weights are calculated for parameters through fully connected layers. During training, the attention weights are automatically optimized as the loss function is backpropagated, ensuring that the model focuses on the parameters that have the greatest impact on K&C performance.

[0042] Output layer: The number of neurons is equal to the K&C performance dimension. A linear activation function is used to ensure that the output value retains the original physical unit and meets the requirements for direct use in engineering.

[0043] The physics-enhanced loss function employs a weighted multi-objective loss function, incorporating physical constraints into the data fitting process. :

[0044] ;

[0045] in, This is the weighted mean square error term. These are weighting coefficients, which vary depending on the importance of the indicator. Indicates the L2 regularization term; This represents a physical constraint penalty term, which penalizes predictions that violate the physical laws that roll stiffness increases with the diameter of the stabilizer bar and vertical stiffness increases with the stiffness of the spring, forcing the model to learn a mapping relationship that conforms to common engineering sense. , These represent the coefficients of the L2 regularization term and the physical constraint penalty term, respectively. They represent the first The measured true value of each K&C performance indicator and the model's predicted value for that K&C performance indicator;

[0046] The trained model is validated using a validation set and tested using a test set to obtain a physically constrained enhanced deep neural network model.

[0047] The geometric constraint verification includes setting a threshold for the deviation of the gradient change rate between adjacent displacement points, where the toe angle gradient increases monotonically with the wheel center compression stroke, and removing samples that exceed the gradient change rate deviation threshold; and setting a threshold for the deviation of the camber angle gradient with the wheel center outward movement, and removing samples that exceed the change rate deviation threshold.

[0048] Preferably, the mechanical constraint verification includes setting a threshold for the rate of change of stiffness, which is monotonically increasing with the spring compression, and removing samples that are greater than the threshold for the rate of change of stiffness.

[0049] Preferably, the AI ​​generalization consistency verification compares the AI ​​predicted value with the measured value. If the relative error of the core key indicators is ≤5% and the curve fitting determination coefficient R² is ≥0.95, then the data generalization is successful. If not, the AI ​​model parameters are adjusted backtracking or interpolation samples are added to ensure that the overall generalization error is ≤5%.

[0050] Another object of the present invention is to provide a suspension K&C performance data generalization system for implementing the suspension K&C performance data generalization method, comprising:

[0051] The data acquisition and preprocessing module is used to collect suspension K&C test data, remove outliers, extract key suspension feature parameters through normalization and PCA-random forest method, and build a raw sample library of "suspension parameters-K&C performance".

[0052] The parameter space modeling module is used to select key variables and set physical constraint ranges based on the correlation logic of "suspension parameters – K&C performance";

[0053] The generalized data generation module is used to expand the suspension parameter space with multidimensional interpolation and predict the corresponding K&C performance data through a neural network model to generate large-scale virtual samples and achieve database generalization.

[0054] The consistency verification module is used to sequentially pass geometric constraint verification, mechanical constraint verification, and AI generalization consistency verification to eliminate unreasonable samples and finally form a reliable generalized database.

[0055] The suspension K&C performance data generalization method of this invention, which integrates multidimensional interpolation and AI mapping, takes the inherent parameters of the suspension as the core and combines interpolation methods with AI modeling cross-scenario database generalization methods. It can leverage a large database of multiple vehicle models with a small sample, effectively breaking through the barriers between vehicle models and scenarios. While ensuring data adaptability and accuracy, it significantly reduces the testing cost and provides low-cost, high-efficiency, and high-quality data support for the development of intelligent connected vehicle chassis. Attached Figure Description

[0056] Figure 1 This is a flowchart of the suspension K&C performance data generalization method of the present invention.

[0057] Figure 2 This is a schematic diagram of the suspension K&C performance data generalization system of the present invention. Detailed Implementation

[0058] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0059] The exemplary embodiment of this application presents a suspension K&C performance data generalization method that integrates multidimensional interpolation and AI mapping. By collecting suspension parameters and performance data, a parameter-performance mapping and physical constraint system is constructed. Interpolation and constraint-enhanced AI are used to expand the parameter space and generate virtual data. Through consistency verification, small-sample cross-vehicle generalization is achieved, providing low-cost and high-efficiency data support for intelligent chassis development.

[0060] The exemplary embodiment of this application presents a suspension K&C performance data generalization method that integrates multidimensional interpolation and AI mapping. Through small sample expansion, cross-vehicle mapping, physical consistency control, and intelligent diagnostics, it achieves efficient generalization and data reuse across vehicle models and suspension types, reducing testing costs and improving chassis development efficiency.

[0061] See Figure 1 As shown in the exemplary embodiment of this application, the suspension K&C performance data generalization method integrating multidimensional interpolation and AI modeling includes the following steps:

[0062] S1. Collect suspension K&C test data, remove outliers, extract key suspension feature parameters through normalization and PCA-random forest method, and construct an original sample library of "suspension parameters-K&C performance";

[0063] S2. Select key variables and set physical constraint ranges based on the correlation logic of "suspension parameters – K&C performance";

[0064] S3. Expand the suspension parameter space by using multidimensional interpolation and predict the corresponding K&C performance data through a neural network model to generate large-scale virtual samples and achieve database generalization.

[0065] S4. Through sequential geometric constraint verification, mechanical constraint verification, and AI generalization consistency verification, unreasonable samples are eliminated, and a reliable generalization database is finally formed.

[0066] In this embodiment of the application, step S1 involves constructing a raw sample library of "suspension parameters – K&C performance" through data acquisition and preprocessing. This is primarily achieved through multi-source data acquisition and intelligent processing to construct a high-quality raw sample library of suspension parameters – K&C performance, providing benchmark data support for the subsequent generalization process.

[0067] Among these, multi-source data collaborative acquisition involves simultaneously collecting two core types of data: suspension parameters and performance data. This constructs basic data linking parameters and performance. K&C performance data acquisition covers toe angle gradient (unit: deg / mm), camber angle gradient (unit: deg / mm), roll stiffness (unit: N·m / deg), vertical stiffness (unit: N / mm), and wheel center displacement-force curve (wheel center vertical displacement range: -50~+100mm, vertical force measurement accuracy: ±1N). Suspension parameter acquisition includes the three-dimensional coordinates of suspension hardpoints. (Accuracy: ±0.1mm), Spring Stiffness (N / mm), stabilizer bar diameter (Unit: mm) Bushing stiffness (Unit: N / mm) Ensure that the parameters cover the core factors affecting suspension performance.

[0068] The collected data should cover at least three different configuration models on the same platform (e.g., standard, sport, and comfort models). For each model, obtain 5-8 sets of valid operating condition samples (including no-load, full-load, and different tilt angles). The total number of original samples should be ≥20 sets to ensure that the correlation between parameters and performance can be learned.

[0069] Specifically, a sliding window dynamic 3σ method is used to accurately screen out abnormal data in K&C performance and suspension parameters. For a single set of K&C indicators or inherent parameters, a sliding window of 5 consecutive samples is used to calculate the mean of the data within the window in real time. with standard deviation Only retain data that meets the following conditions:

[0070] (Equation 1)

[0071] in, This is a dynamic correction coefficient, ranging from 0.02 to 0.05, which is adaptively adjusted according to the data dispersion. After removing outliers, the data retention rate is stably maintained at 92% to 95%, ensuring the original data quality baseline.

[0072] To address the magnitude differences in parameters and K&C indicators across different vehicle models, an industry benchmark mapping method is used to uniformly normalize them to the [0,1] interval, eliminating the incomparability of cross-vehicle data.

[0073] (Equation 2)

[0074] in, The original value, , These are the extreme values ​​for this indicator for vehicles in the same class. This is the industry average. The weighting coefficient ranges from 0.6 to 0.8; it takes into account both industry benchmarks and data distribution characteristics, with a normalization error of less than 2%.

[0075] Furthermore, for the core influencing parameters, a PCA-random forest fusion feature extraction strategy is employed, using a two-stage feature extraction strategy of dimensionality reduction and importance screening to accurately identify the core influencing parameters: the covariance matrix is ​​calculated through principal component analysis and eigenvalue decomposition is performed.

[0076] (Equation 3)

[0077] in Let covariance matrix be the variance matrix. The matrix consists of orthogonal eigenvectors. The eigenvalue diagonal matrix is ​​used to retain principal components with an eigenvalue contribution rate ≥ 85%. Then, the feature importance score is calculated using a random forest algorithm to finally determine the three-dimensional coordinates of the suspension hardpoints. Spring stiffness stabilizer bar diameter Bushing stiffness The core key feature parameters.

[0078] In this embodiment of the application, step S2 mainly involves parameter space modeling. By defining the mapping relationship between parameters and performance, as well as physical constraints, the engineering feasibility of generalized data is ensured, including:

[0079] Define the input-output mapping relationship and clarify the precise correspondence between suspension parameters and K&C performance: (Equation 4)

[0080] in, The input parameter vector contains the three-dimensional coordinates of the front and rear wheel hardpoints, spring stiffness, stabilizer bar diameter, and bushing stiffness. The output K&C performance index vector includes the front head angle gradient. outward tilt gradient Lateral stiffness Vertical stiffness Wheel center displacement-force curve ; Subscript Represents the left rear wheel, subscript Represents the right rear wheel; subscript Represents the left front wheel, subscript This represents the right front wheel.

[0081] To ensure the physical validity of the data, and in accordance with international and domestic standards and engineering practices, a three-level constraint system is established for input parameters and output indicators: Limit range: setting an absolute boundary that the parameters cannot be exceeded; Safety range: setting an engineering-feasible range for the parameters with reference to the suspension design specifications of mainstream vehicle models; Optimization range: setting a performance-sensitive range based on the parameter range corresponding to the optimal performance of the original sample; All constraint ranges are related through the following inequality to form a closed-loop constraint system.

[0082] (Equation 5)

[0083] Among them, Corresponding to the boundary of the safe zone, Corresponding to the optimized interval boundary, The corresponding limit interval boundary.

[0084] In this embodiment, step S3 is mainly used for generalized data generation. Through parameter interpolation expansion and AI performance prediction, a large-scale virtual sample is generated to achieve database generalization. First, virtual suspension parameter samples are generated through interpolation. Then, based on these interpolated virtual suspension parameter samples, a physical constraint-enhanced deep neural network is used for K&C performance prediction. The generated virtual suspension parameters are batch-input into the trained model, and the corresponding virtual K&C performance data is output. By integrating physical laws and data-driven modeling, a mapping from virtual parameters to reliable K&C performance is achieved. Specifically, the virtual suspension parameter samples generated through interpolation address the insufficient coverage of parameters in the original samples. Radial basis function (RBF) interpolation is used to fill the high-dimensional space composed of suspension parameters such as hardpoint coordinates and stiffness parameters, generating new parameter combinations. This includes the following steps:

[0085] Using the extracted core / key feature parameters as dimensions, an m-dimensional parameter space (m=8~12) is constructed. The original sample parameters form a discrete point set {X1,X2,...,X} in this space. n}, (X i For the first (where n is the number of original samples), and for the regions in the parameter space not covered by the original samples, new suspension intrinsic parameter combinations are generated through RBF interpolation. :

[0086] (Equation 6)

[0087] in, For radial basis functions, a Gaussian function is chosen. , This is a shape parameter, ranging from 0.8 to 1.2. Larger values ​​are used when parameter density is high to ensure uniform distribution of interpolation points. This represents the Euclidean distance between two points in a high-dimensional parameter space. The interpolation weighting coefficients are obtained by satisfying the parametric geometric correlation constraints. These are the parameters of the original sample. This represents the original number of samples. The Euclidean distance between the new parameters and the original sample parameters is controlled within 0.3 times the diagonal length of the parameter space. Based on this, over-extrapolation is avoided.

[0088] Simultaneously, physical verification of the interpolation parameters must be performed: the generated new parameters must meet the three-level constraints in step 2.2, and 10% of the interpolation parameters are randomly selected for verification through multibody dynamics simulation to ensure that there is no motion interference in the hard point coordinates; the basic K&C performance trend is reasonable; the verification pass rate is ≥98%, and the number of interpolation parameters generated is 10 to 20 times that of the original samples, forming a "virtual suspension parameter library".

[0089] Through the above steps, the virtual suspension parameter samples generated by interpolation were realized, laying the foundation for the next step of generalization processing.

[0090] In this embodiment, the physically constrained enhanced deep neural network is used for AI modeling generalization and K&C performance prediction. The generated virtual suspension parameters are input in batches into the trained model, and the corresponding virtual K&C performance data is output. By fusing physical laws with data-driven modeling, the mapping from virtual parameters to reliable K&C performance is achieved. The construction and training of the physically constrained enhanced deep neural network (model) includes the following steps:

[0091] To ensure the model's realism, a training dataset with strong correlations between suspension parameters and K&C performance is constructed, using original measured samples as the core. This ensures that the physics-constrained enhanced deep neural network model learns the physical characteristics of real suspensions. The input to the physics-constrained enhanced deep neural network model is the original suspension intrinsic parameter vector. The output of the physics-constrained enhanced deep neural network model is the corresponding measured K&C performance index. The training dataset is divided into three parts: 70% training set, 20% validation set, and 10% test set. All data comes from original experimental samples (without interpolation parameters) to ensure that the model learns "real physical laws". When training the model using the training dataset, a five-fold cross-validation is used during the training phase: the training set is randomly divided into 5 subsets, and 4 subsets are used for training and 1 subset for validation in turn. Finally, the average of the model parameters from the 5 training iterations is taken as the model parameters to avoid overfitting caused by data partitioning bias.

[0092] To address the varying sensitivities to different key parameters, an improved residual attention neural network is employed. This enhances the nonlinear fitting capability while simultaneously strengthening the sensitivity to key parameters. The improved residual attention neural network comprises the following:

[0093] Input layer: The number of neurons is equal to the core parameter dimension. It directly receives the suspension's inherent parameter vector, which is strictly matched with the extracted key feature parameter dimension.

[0094] Hidden layer: A 3-layer fully connected structure with residual connections (64×32×16 neurons).

[0095] Residual connections: The output of each layer is as follows:

[0096] (Equation 7)

[0097] in, For layer input, It is a non-linear mapping function, which alleviates the gradient vanishing problem in deep networks and ensures that the model can learn the subtle relationship between parameters and performance;

[0098] Regularization: Batch Normalization (to accelerate convergence) and Dropout (p=0.2, randomly deactivating 20% ​​of neurons) are added after each layer to suppress overfitting;

[0099] Activation function: Leaky-ReLU (negative slope 0.01) is selected to solve the problem of neuron inactivation in the negative region of traditional ReLU and enhance the ability to capture weak signals such as small stiffness changes and hard point fine-tuning.

[0100] Parameter-sensitive attention module: An attention mechanism is inserted between the input layer and the first hidden layer to form a parameter-sensitive attention module.

[0101] Attention weights for each parameter are calculated using a fully connected layer. The key parameters such as the Y-coordinate of the front wheel hard point (which affects the track width and roll) and the bushing stiffness (which affects compliance) are given greater weight, while the secondary parameters (such as the Z-coordinate of some non-sensitive hard points) are given less weight.

[0102] Dynamic weight updates: During training, attention weights are automatically optimized as the loss function is backpropagated, ensuring that the model focuses on the parameters that have the greatest impact on K&C performance.

[0103] Output layer: The number of neurons is equal to the K&C performance dimension (e.g., 5 core metrics correspond to 5 neurons). A linear activation function is used to ensure that the output value retains the original physical units (e.g., deg / mm, Nm / deg) to meet the needs of direct engineering use.

[0104] Specifically, physical constraint embedding and multi-strategy optimization are used to ensure the model's prediction accuracy and generalization ability: A physically enhanced loss function is employed, using a weighted multi-objective loss function that incorporates physical constraints into the data fitting process.

[0105] (Equation 8)

[0106] in: This is the weighted mean square error term, with core indicators (toe angle gradient, roll stiffness). Secondary indicators ; This is an L2 regularization term, used to suppress excessively large parameters; As a physical constraint penalty term, it penalizes prediction results that violate physical laws such as "tilt stiffness increases with the increase of stabilizer rod diameter" and "vertical stiffness increases with the increase of spring stiffness", forcing the model to learn a mapping relationship that conforms to common sense in engineering.

[0107] Specifically, to enable the model to quickly adapt to other vehicle models, the Adam optimizer is selected to adapt to the optimization of non-stationary objective functions; adaptive learning rate decay: when the validation set loss does not decrease for 15 consecutive rounds, the learning rate is automatically halved (down to a minimum of 1e-6), balancing convergence speed and optimization accuracy.

[0108] To enhance the generalization ability of the model, an early stopping strategy is adopted: when the validation set loss increases for 25 consecutive rounds, training is terminated and the current optimal model parameters are saved to avoid overfitting; transfer learning: for new models on the same platform, the training weights of existing models are loaded as initial parameters, and only the top-level network is fine-tuned (the learning rate is reduced to 1e-5) so that the model can quickly adapt to the parameter distribution characteristics of the new models.

[0109] Among them, the training convergence criteria are as follows: the core indicators are: the loss decrease rate of the validation set is <1e-5 and remains stable for 20 consecutive rounds to ensure that the model fully converges; the accuracy requirements are: the prediction error of the test set (the original measured samples that did not participate in the training) is ≤3%, of which the error of core indicators such as the toe angle gradient and the roll stiffness is ≤2%, which provides a highly reliable model foundation for the subsequent K&C prediction of virtual parameters.

[0110] After the model training is completed, the generated virtual suspension parameters (which have passed the physical constraint verification) are input into the trained model in batches, and the corresponding virtual K&C performance data is output, thereby realizing the generalization of data.

[0111] The model predicts all dimensions of K&C indicators: toe angle gradient (output per displacement point, step size 1mm), camber angle gradient, roll stiffness, vertical stiffness, and wheel center displacement-force curve (-50~+50mm full stroke); the prediction time for a single set of parameters is ≤0.05s, and it can efficiently generate virtual K&C data with more than 100 times the original sample size. All data are marked with the "AI prediction" attribute and prediction error range (calculated based on the accuracy of the test set), providing core extended samples for database generalization.

[0112] In this embodiment of the application, during the consistency verification in step S4, a triple verification mechanism is mainly used to ensure the physical rationality and engineering usability of the generalized data:

[0113] First, to ensure the reasonableness of the data, a geometric constraint verification is established. Based on the suspension kinematics theory, a multi-parameter linkage geometric verification rule is constructed to ensure that the K&C index changes with key variables in accordance with physical laws.

[0114] For example: the gradient of the lead-beam angle Follow-through compression stroke Monotonically increasing, with deviations in the rate of change of gradient between adjacent displacement points:

[0115] (Equation 9)

[0116] outward gradient Outward displacement of the wheel center Deviation in rate of change:

[0117] (Equation 10)

[0118] By verifying geometric constraints, samples that do not conform to the laws of kinematics are eliminated.

[0119] Secondly, based on geometric constraint verification, mechanical constraint verification is established. Combining the principles of mechanics of materials and suspension structure mechanics, a rigidity verification system for mechanical properties is established:

[0120] For example: vertical stiffness With spring compression Monotonically increasing rate of change of stiffness:

[0121] (Equation 11)

[0122] The mechanical constraints are verified to ensure the mechanical rationality of the extended data, with a pass rate of ≥97%.

[0123] In addition, after completing the above constraint checks, to ensure successful generalization of the remaining data, an AI generalization consistency check is established: 10 sets of measured data from the same platform vehicle models that were not involved in the training are selected as verification samples, and the AI ​​predictions are compared with the measured values: if the relative error of the core indicators is ≤5%; and the curve fitting determination coefficient R² is ≥0.95, then the data generalization is successful. If not, the AI ​​model parameters are adjusted retrospectively (e.g., adding hidden layer neurons) or interpolation samples are added to ultimately ensure that the overall generalization error is ≤5%.

[0124] The method provided in this invention sequentially performs data acquisition and preprocessing, parameter space modeling, generalized data generation, and consistency verification, forming a complete closed loop through process connections. In the data acquisition and preprocessing stage, suspension K&C test data is collected, and 3D modeling is used to generate the data. The method involves eliminating outliers, extracting key features through normalization and PCA, and constructing an original sample library of "suspension parameters – K&C performance". In the parameter space modeling stage, key variables are selected based on the parameter-performance correlation logic, and physical constraint intervals are set to ensure rationality and engineering feasibility. In the generalization data generation stage, the inherent parameter space of the suspension is expanded using multidimensional interpolation, and the corresponding K&C performance data is predicted through a neural network model. In the consistency verification stage, geometric constraint verification, mechanical constraint verification, and AI generalization consistency verification are performed sequentially to eliminate unreasonable samples, ultimately forming a reliable generalization database. This method is not limited by suspension type and is applicable to mainstream suspensions such as MacPherson strut, double wishbone, and multi-link suspensions, and can be extended to the suspension development of different vehicle models.

[0125] The above analysis shows that the suspension K&C performance database generalization method integrating multidimensional interpolation and AI mapping of the present invention can establish an efficient generalization mechanism from small samples to large databases, and quickly expand K&C performance data for multiple vehicle models and suspension types; it can better balance the database construction cost and data reliability, ensuring low prediction error and strong engineering usability; and it can be deeply coupled with intelligent chassis development, supporting performance diagnosis, parameter optimization and cross-vehicle design, and accelerating the R&D process.

[0126] See Figure 2 As shown, embodiments of the present invention also provide a suspension K&C performance data generalization system for implementing the suspension K&C performance data generalization method described in this application, including:

[0127] The data acquisition and preprocessing module is used to collect suspension K&C test data, remove outliers, extract key suspension feature parameters through normalization and PCA-random forest method, and build a raw sample library of "suspension parameters-K&C performance".

[0128] The parameter space modeling module is used to select key variables and set physical constraint ranges based on the correlation logic of "suspension parameters – K&C performance";

[0129] The generalized data generation module is used to expand the suspension parameter space with multidimensional interpolation and predict the corresponding K&C performance data through a neural network model to generate large-scale virtual samples and achieve database generalization.

[0130] The consistency verification module is used to sequentially pass geometric constraint verification, mechanical constraint verification, and AI generalization consistency verification to eliminate unreasonable samples and finally form a reliable generalized database.

[0131] The suspension K&C performance data generalization system that integrates multidimensional interpolation and AI mapping described above adopts a technical solution that corresponds to the solution of the suspension K&C performance database generalization method that integrates multidimensional interpolation and AI mapping described above. Therefore, it will not be repeated here. Please refer to all the relevant descriptions of the embodiments of the suspension K&C performance data generalization method that integrates multidimensional interpolation and AI mapping described above. All of them are applicable to this suspension K&C performance data generalization system that integrates multidimensional interpolation and AI mapping.

[0132] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or basic features of the present invention.

[0133] Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of the equivalents of the claims be included within the invention.

[0134] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for generalizing suspension K&C performance data, characterized in that, Includes the following steps: Suspension K&C test data were collected, outliers were removed, and key suspension feature parameters were extracted by normalization and PCA-random forest method to construct an original sample library of "suspension parameters-K&C performance". Key variables are selected and physical constraint ranges are set based on the correlation logic of "suspension parameters – K&C performance"; The physical constraint range is defined as follows: Establish three levels of constraints for input parameters and output performance metrics: Limit range: The boundary that the set parameters must absolutely not exceed; Safe range: Referencing the vehicle's suspension design specifications, set the parameters within the engineering feasible range; Optimization range: Based on the parameter range corresponding to the optimal performance of the original samples, a performance-sensitive range is set; All constraint intervals form a closed-loop constraint system; By expanding the suspension parameter space through multidimensional interpolation and predicting the corresponding K&C performance data through a neural network model, a large-scale virtual sample is generated to achieve database generalization. The prediction of the corresponding K&C performance data through the neural network model is based on the virtual suspension parameter samples generated by interpolation. A physical constraint-enhanced deep neural network is used for K&C performance prediction. The generated virtual suspension parameters are input into the trained model in batches, and the corresponding virtual K&C performance data is output. By integrating physical laws and data-driven modeling, the mapping from virtual parameters to reliable K&C performance is realized. By sequentially passing geometric constraint verification, mechanical constraint verification, and AI generalization consistency verification, unreasonable samples are eliminated, ultimately forming a reliable generalization database. The geometric constraint verification is based on suspension kinematics theory, constructs multi-parameter linkage geometric verification rules, eliminates samples that do not conform to kinematic laws, and ensures that the K&C performance index conforms to physical laws as key variables change. Based on the aforementioned geometric constraint verification, and combining the principles of mechanics of materials and suspension structure mechanics, a rigidity verification system for mechanical properties is established to perform mechanical constraint verification and ensure the mechanical rationality of the extended data. After completing the constraint verification, an AI generalization consistency check is performed. Multiple sets of measured data from vehicles on the same platform that did not participate in the training are selected as verification samples. The AI ​​prediction values ​​are compared with the measured values ​​to determine whether the data generalization is successful. If it is unsuccessful, the AI ​​model parameters are adjusted or interpolation samples are added.

2. The suspension K&C performance data generalization method according to claim 1, characterized in that, The suspension K&C test data was obtained through multi-source data collaborative acquisition. The K&C performance data acquisition covers toe angle gradient, camber angle gradient, roll stiffness, vertical stiffness, and wheel center displacement-force curve. The suspension parameter collection includes the three-dimensional coordinates of the suspension hard points, spring stiffness, stabilizer bar diameter, and bushing stiffness. The collection covers at least different configuration models on the same platform, with multiple sets of valid working condition samples for each model, including no-load, full-load, and different roll angles. The total number of original samples meets expectations.

3. The suspension K&C performance data generalization method according to claim 1, characterized in that, Using a sliding window dynamic 3 Methods to remove outliers from the data; using multiple consecutive samples as a sliding window, the mean of the data within the window is calculated in real time. with standard deviation Only retain those that meet the requirements. Data; among which This is a dynamic correction factor. This is data to be processed.

4. The suspension K&C performance data generalization method according to claim 1, characterized in that, Based on the differences in suspension parameters and K&C indicators among different vehicle models, the industry benchmark mapping method is used to normalize the data after outlier removal to the [0,1] interval, eliminating the incomparability of cross-vehicle data. The normalized data is obtained by the following formula: ； in, The data to be processed, i.e., the original values. , These are the extreme values ​​for the corresponding indicators of vehicles in the same class. This is the industry average. These are the weighting coefficients. This is the normalized data.

5. The suspension K&C performance data generalization method according to claim 1, characterized in that, The PCA-Random Forest method for extracting key suspension feature parameters includes calculating the covariance matrix and performing eigenvalue decomposition through principal component analysis, then calculating the feature importance score through the random forest algorithm, and determining the final key suspension feature parameters based on the score.

6. The suspension K&C performance data generalization method according to claim 1, characterized in that, The selection of key variables and setting of physical constraint ranges based on the "suspension parameters – K&C performance" correlation logic includes: Define the input-output mapping relationship, clarify the corresponding logic between suspension parameters and K&C performance, and ensure that the mapping relationship covers all dimensions of suspension parameters and K&C performance; All the constraint intervals are related by the following inequality, forming a closed-loop constraint system: ； Among them, Corresponding to the extreme values ​​at the boundaries of the safe interval, Corresponding to the optimization interval boundary extreme value, The corresponding extreme value at the boundary of the limiting interval, Indicates the first One input parameter.

7. The suspension K&C performance data generalization method according to claim 6, characterized in that, The aforementioned expansion of the suspension parameter space using multidimensional interpolation employs the radial basis function (RBF) interpolation method to fill the high-dimensional space formed by the suspension parameters, generating new parameter combinations, and performing interpolation physical verification to form a virtual suspension parameter library, including: Using key feature parameters as the dimension, an m-dimensional parameter space is constructed. The original sample parameters form a discrete point set {X1, X2, ..., X} in this m-dimensional parameter space. n }, where n is the number of original sample parameters; for the region in the m-dimensional parameter space not covered by the original sample parameters, new suspension parameter combinations are generated through RBF interpolation. : ； in, These are radial basis functions; These are the interpolation weighting coefficients. For the first One original sample parameter, For new parameters Compared with the original sample parameters The Euclidean distance; Physical verification of interpolation parameters is performed to ensure that the new parameters meet the three-level constraints of the limit range, safety range, and optimization range. 10% of the interpolation parameters are randomly selected and verified through multibody dynamics simulation to ensure that there is no motion interference in the hard point coordinates, the basic K&C performance trend is reasonable, the verification pass rate is ≥98%, and a virtual suspension parameter library is formed.

8. The suspension K&C performance data generalization method according to claim 7, characterized in that, The construction and training of the physical constraint-enhanced deep neural network includes: Based on the original measured samples, a training dataset with strong correlation between suspension parameters and K&C performance was constructed: during the model training phase, five-fold cross-validation was adopted, and the training set was randomly divided into 5 subsets. Four subsets were used for training and one subset for validation in turn. Finally, the average of the model parameters from the 5 training iterations was taken as the model parameters. An improved residual attention neural network is employed to enhance the model's nonlinear fitting ability and strengthen its sensitivity to key parameters. The improved residual attention neural network includes: Input layer: Its number of neurons is equal to the dimension of the feature parameters, and it receives the suspension parameter vector; Hidden layer: Employs a 3-layer fully connected structure with residual connections, where each layer's output... ;in, For layer input, It is a nonlinear mapping function. Weights are used; BatchNormalization and Dropout are added after each layer of residual connections to suppress overfitting; The activation function chosen is Leaky-ReLU; Insert an attention mechanism between the input layer and the first hidden layer: Attention weights are calculated for parameters through fully connected layers. During training, the attention weights are automatically optimized as the loss function is backpropagated, ensuring that the model focuses on the parameters that have the greatest impact on K&C performance. Output layer: The number of neurons is equal to the K&C performance dimension. A linear activation function is used to ensure that the output value retains the original physical unit and meets the requirements for direct use in engineering. The physics-enhanced loss function employs a weighted multi-objective loss function, incorporating physical constraints into the data fitting process. : ； in, This is the weighted mean square error term. These are weighting coefficients, which vary depending on the importance of the indicator. Indicates the L2 regularization term; This represents a physical constraint penalty term, which penalizes predictions that violate the physical laws that roll stiffness increases with the diameter of the stabilizer bar and vertical stiffness increases with the stiffness of the spring, forcing the model to learn a mapping relationship that conforms to common engineering sense. , These represent the coefficients of the L2 regularization term and the physical constraint penalty term, respectively. They represent the first The measured true value of each K&C performance indicator and the model's predicted value for that K&C performance indicator; The trained model is validated using a validation set and tested using a test set to obtain a physically constrained enhanced deep neural network model.

9. The suspension K&C performance data generalization method according to claim 1, characterized in that, The geometric constraint verification includes setting a threshold for the deviation of the gradient rate of change of the toe angle gradient with the wheel center compression stroke, and removing samples that exceed the threshold for the gradient rate of change deviation; and setting a threshold for the deviation of the camber angle gradient with the wheel center outward movement, and removing samples that exceed the threshold for the rate of change deviation. The mechanical constraint verification includes setting a threshold for the rate of change of stiffness, which is monotonically increased with the spring compression, and removing samples that exceed the threshold for the rate of change of stiffness. The AI ​​generalization consistency check compares the AI ​​predictions with the measured values. If the relative error of the core key indicators is ≤5% and the coefficient of determination R² for curve fitting is ≥0.95, then the data generalization is successful. If these conditions are not met, the AI ​​model parameters are adjusted backtracking or interpolation samples are added to ultimately ensure that the overall generalization error is ≤5%.

10. A suspension K&C performance data generalization system, used to implement the suspension K&C performance data generalization method according to any one of claims 1-9, characterized in that, include: The data acquisition and preprocessing module is used to collect suspension K&C test data, remove outliers, extract key suspension feature parameters through normalization and PCA-random forest method, and build a raw sample library of "suspension parameters-K&C performance". The parameter space modeling module is used to select key variables and set physical constraint intervals based on the correlation logic of "suspension parameters – K&C performance"; The generalized data generation module is used to expand the suspension parameter space with multidimensional interpolation and predict the corresponding K&C performance data through a neural network model to generate large-scale virtual samples and achieve database generalization. The consistency verification module is used to sequentially pass geometric constraint verification, mechanical constraint verification, and AI generalization consistency verification to eliminate unreasonable samples and finally form a reliable generalized database.

Images