Automobile suspension-body full attachment point load prediction method and system

Abstract

The application provides a vehicle suspension-body full attachment point load prediction method and system, the method comprising: collecting synchronous time sequence data and vehicle model parameters of the vehicle, constructing time sequence samples and performing mask self-supervised pre-training; after the excitation sequence is encoded through channel attention and multi-scale, the relative position Transformer conditioned by the vehicle model parameters is used to output the full attachment point load sequence, and a joint loss is formed by data fitting, six-degree-of-freedom balance of the vehicle body, spectral consistency and load change rate boundary. Only the excitation sequence and the vehicle model parameters are required during inference, and the full attachment point six-direction dynamic load spectrum with an uncertainty interval is output. The application can reduce the test and virtual iteration cost, improve the physical consistency of extreme working conditions and the cross-vehicle model generalization ability.

Description

Technical Field

[0001] This application belongs to the field of vehicle fatigue life assessment technology, and in particular relates to a method and system for predicting loads at all attachment points of an automobile suspension-body. Background Technology

[0002] The dynamic load spectrum of the vehicle suspension-body attachment points is an important boundary condition for fatigue life assessment of the body, suspension components, and connecting structures. The six-axis load at each attachment point typically includes force components in three directions along the vehicle coordinate system and moment components in three directions. Due to the large number of attachment points, limited installation space, and significant impact from reinforced road surfaces, directly placing strain gauges or multi-axis force sensors at all attachment points presents challenges such as difficult placement, large calibration workload, high cost of multi-channel synchronous acquisition, and insufficient signal stability.

[0003] Existing load decomposition methods typically input the wheel center six-component forces, suspension displacements, or accelerations collected from actual vehicles into a multibody dynamics model of the vehicle, and obtain the loads at component connection points through virtual iteration or dynamic solution. This method relies on a high-precision vehicle model and nonlinear parameters such as springs, shock absorbers, and bushings, resulting in long model tuning and iterative calculation cycles, and may also fail to converge under complex working conditions. Existing neural network methods can be used for time-series response or load spectrum prediction, but purely data-driven models lack mechanical boundary constraints, and may produce prediction results that do not satisfy the balance relationship of the resultant force and resultant moment of the vehicle body under insufficient training data coverage or extreme impact conditions. At the same time, models trained only for a single vehicle model are difficult to adapt to changes in vehicle mass, inertia, hard point locations, suspension stiffness, and damping parameters. Summary of the Invention

[0004] In view of this, this application aims to propose a method and system for predicting loads at all attachment points of an automotive suspension-body assembly, in order to solve at least one of the above-mentioned problems.

[0005] To achieve the above objectives, the technical solution of this application is implemented as follows: In a first aspect, this application provides a method for predicting loads at all attachment points of an automotive suspension system and body, including: Acquire vehicle synchronization timing data and vehicle model parameters. The synchronization timing data includes six-component forces at the four wheel centers, damping forces of each shock absorber, vehicle body inertial measurement signals, and six-axis loads at each suspension-body attachment point as tags. The synchronous time-series data and the vehicle parameters are preprocessed, training samples are constructed based on a sliding time window, and the time encoder is pre-trained for mask reconstruction using an unlabeled excitation sequence. A load prediction model is constructed. The excitation sequence consisting of the six component forces of the wheel center and the damping force of the shock absorber is subjected to channel weighting and multi-scale time-series encoding. The encoding features are conditionally modulated using modulation parameters generated by vehicle model parameters. The modulated encoding features are input into a Transformer encoder with relative position offset, and the six-axis load prediction sequence of all attachment points is output. The total loss is constructed based on the data fitting error of the predicted sequence and the label, the balance error of the six-degree-of-freedom resultant force and resultant moment established according to the motion state of the white body and the position of each attachment point, the spectrum error, and the load change rate exceeding the limit error due to the stiffness of the attachment point, and the load prediction model is trained. Obtain the wheel center six-component force, shock absorber damping force and vehicle model parameters of the vehicle to be predicted, input the trained load prediction model according to the time sequence window, and obtain and output the six-axis dynamic load sequence of all suspension-body attachment points of the vehicle to be predicted.

[0006] Secondly, based on the same inventive concept, this application also provides a vehicle suspension-body full attachment point load prediction system, comprising: The data acquisition module is configured to acquire the vehicle's synchronous timing data and vehicle model parameters. The synchronous timing data includes the six-component forces at the four wheel centers, the damping forces of each shock absorber, the vehicle body inertial measurement signals, and the six-axis loads at each suspension-body attachment point as tags. The self-supervised pre-training module is configured to perform data preprocessing on the synchronous time-series data and the vehicle parameters, construct training samples based on a sliding time window, and perform mask reconstruction pre-training on the time-series encoder using unlabeled excitation sequences. The model building module is configured to build a load prediction model. It performs channel weighting and multi-scale time-series encoding on the excitation sequence consisting of the six component forces of the wheel center and the damping force of the shock absorber. It uses the modulation parameters generated by the vehicle model parameters to conditionally modulate the encoded features. The modulated encoded features are input into the Transformer encoder with relative position offset, and the six-axis load prediction sequence of all attachment points is output. The model training module is configured to construct the total loss based on the data fitting error of the predicted sequence and labels, the balance error of the six-degree-of-freedom resultant force and resultant moment established according to the motion state of the white body and the position of each attachment point, the spectrum error, and the load change rate exceeding the limit error limited by the stiffness of the attachment point, and train the load prediction model. The prediction output module is configured to acquire the wheel center six-component force, shock absorber damping force and vehicle model parameters of the vehicle to be predicted, input the trained load prediction model according to the time sequence window, and obtain and output the six-axis dynamic load sequence of all suspension-body attachment points of the vehicle to be predicted.

[0007] Thirdly, based on the same inventive concept, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in the first aspect.

[0008] Fourthly, based on the same inventive concept, this application also provides a non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium stores computer instructions for causing the computer to perform the method as described in the first aspect.

[0009] Compared with the prior art, the method and system for predicting loads at all attachment points of an automotive suspension-body assembly described in this application have the following advantages: This application only requires the collection of easily obtainable wheel center six-component forces, shock absorber damping forces, and key vehicle parameters as inputs to achieve prediction of six-axis loads at all attachment points between the suspension and the vehicle body. This significantly reduces the cost of load acquisition, ensures strict self-consistency of physical constraints and dynamic boundaries, improves the prediction accuracy under extreme conditions, cross-vehicle generalization ability, and engineering application efficiency, and the prediction results can be directly used for fatigue and stress verification of the vehicle body. Attached Figure Description

[0010] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a method for predicting loads at all attachment points of an automobile suspension and body, as described in an embodiment of this application. Figure 2 This is a schematic diagram of a vehicle suspension-body full attachment point load prediction system according to an embodiment of this application; Figure 3 This is a schematic diagram of the hardware structure of the electronic device described in an embodiment of this application. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0012] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0013] The "full attachment points" mentioned in this application refer to all suspension-body connection points in the target suspension topology that transmit loads to the vehicle body, including but not limited to the connection points between the control arm and the vehicle body or subframe, the connection points between the upper shock absorber support and the vehicle body, the connection points between the torsion beam trailing arm and the vehicle body, and the connection points between the stabilizer bar support and the vehicle body. Each attachment point outputs three force components and three torque components. Within the same training task, all vehicle models use the same target attachment point definition, number, and output arrangement; for suspensions with different attachment point topologies or numbers, corresponding training tasks and output heads can be constructed separately.

[0014] The embodiments of this application are described in detail below with reference to the accompanying drawings.

[0015] Please see Figure 1 As shown, this embodiment provides a method for predicting loads at all attachment points of an automotive suspension-body system, specifically including the following steps: Step S10: Obtain the vehicle's synchronization timing data and model parameters. The synchronization timing data includes the six-component forces at the four wheel centers, the damping forces of each shock absorber, the vehicle body inertial measurement signals, and the six-axis loads at each suspension-body attachment point as tags.

[0016] Specifically, in this embodiment, a sedan with a front MacPherson strut and rear torsion beam suspension is used as an example. Other suspension type combinations can be handled in the same way. Using a model with a front MacPherson strut and rear torsion beam suspension, a real-vehicle road test is conducted under full-load conditions based on actual durability conditions (cobblestone roads, tortuous roads, washboard roads, pothole roads, Belgian roads, etc.). The sampling frequency is 100Hz, and all timing signals are uniformly collected using a CSM multi-channel synchronous data acquisition system. The channel synchronization time difference is <1ms, ensuring complete alignment of the wheel center six-component force, shock absorber force, IMU body motion signal, and suspension-body attachment point load timestamps. The system collects six component forces at the wheel centers of the four wheels of the vehicle, and outputs six channels of FX / FY / FZ / TX / TY / TZ per wheel, totaling 24 channels of original road surface excitation signals; it also collects the dynamic damping tension and pressure of the four hydraulic shock absorbers on the left front, right front, left rear, and right sides of the vehicle during the entire compression and recovery process, totaling 4 channels of original shock absorber damping force signals. Vehicle IMU signal acquisition: An inertial measurement unit (IMU) is placed near the vehicle's center of gravity, and the data acquisition system synchronously acquires the vehicle's three-axis translational acceleration output by the IMU. Triaxial angular velocity The three-axis angular accelerations of the vehicle body are obtained by performing time difference analysis on the angular velocity time sequence. The IMU outputs a total of 6 raw motion signals, which are strictly time-aligned with all other channels.

[0017] In this embodiment, the tag data is the six-axis load time-series data of all suspension-body attachment points, including: control arm-body / body side subframe mounting points, shock absorber-body mounting points, torsion beam-body connection points, stabilizer bar-body supports, and all other points where the suspension transmits forces to the body; each attachment point contains 3 force components and 3 torque components (six-axis load). The tag data is obtained by: after establishing a vehicle dynamics model, simulating the six-axis time-series load of all suspension-body attachment points based on the wheel center six-component forces and shock absorber forces; assuming the total number of attachment points is P, the total number of tag channels is... .

[0018] Record the key parameter vectors for each test vehicle model, including: body-in-white mass. Front and rear axle load distribution, wheelbase L, track width B, center of gravity height h, and moment of inertia of the body-in-white about the X / Y / Z axes. Front suspension coil spring stiffness Rear suspension torsion bar stiffness Damping coefficients of front and rear shock absorbers The position vectors of each suspension-body attachment point relative to the body's center of gravity ( ), the radial / axial stiffness of each attachment point bushing, totaling 20~30 vehicle parameters, denoted as vectors. , which serves as the input for subsequent model conditionalization.

[0019] It should be noted that, , These are parameters for the body-in-white, excluding moving suspension components such as control arms, subframes, and wheels. It is directly used for calculating the force balance equations of the vehicle body isolator.

[0020] Step S20: Perform data preprocessing on the synchronous time series data and vehicle parameters, construct training samples based on the sliding time window, and use the unlabeled excitation sequence to perform mask reconstruction pre-training on the time series encoder.

[0021] Specifically, in this embodiment, through the above step S10, a total of 28 input variable timing data are obtained, including 24 channels of six-component forces at the four wheel centers and 4 channels of forces at the four shock absorbers; the IMU outputs 6 channels of motion timing data, including the vehicle's three-axis acceleration and three-axis angular velocity; and the data is labeled as six-axis loads at all suspension-vehicle attachment points. Road tag timing data; sampling frequency 100Hz, one frame of data is output every 0.01s.

[0022] This embodiment uses a normalization tool (such as MinMaxScaler) to uniformly normalize the 28 input excitation signals to a normalized state. Interval: ; In the formula, This represents the original input excitation signal at time step t and channel d. This represents the global minimum value of the d-th input signal; This represents the global maximum value of the d-th input signal; This represents the normalized input excitation signal, with a value range of... .

[0023] The IMU signal is not input into the network; it is only used to calculate the physical constraint loss during the training phase. Therefore, the three-axis acceleration and three-axis angular velocity signals of the vehicle body acquired by the IMU are not normalized; only time window sliding is performed subsequently. Each road target label is fitted with an independent normalized scaler and normalized to a specific value. The interval should be defined to avoid interference between signals of different dimensions and amplitudes. The vehicle parameter vectors are standardized using Z-Score: ; In the formula, Represents the original vehicle model parameter vector. dimension; This represents the mean of the parameter in this dimension across the dataset; This represents the standard deviation of the parameter in this dimension on the dataset; This represents the vehicle parameters after Z-score standardization.

[0024] This embodiment uses a length of A sliding time window is used to divide continuous time-series data into supervised learning samples. The input of each sample is 28 input excitation signals from the previous 60 time steps. The labels are the load sequences of all suspension-body attachment points corresponding to 60 time steps. Each sample synchronously carries the IMU vehicle body motion timing within the corresponding window, as well as the parameter vector of that vehicle model as conditional input.

[0025] The above data collection and processing can be repeated for no fewer than ten similar vehicle models with the same target attachment point definition. As a data configuration example, approximately 70% of the samples retain complete attachment point load labels for supervised training, while approximately 30% of the samples retain only 28 input stimuli for self-supervised pre-training. The obtained labeled dataset is split into a training set (80%), a validation set (10%), and a test set (10%). When splitting, it is preferable to group by vehicle or complete road segment to avoid highly overlapping windows of the same road segment entering both the training and test sets simultaneously.

[0026] This embodiment utilizes readily available, unlabeled data on the six component forces at the wheel center and the damper force to perform masked time series modeling before supervised training. By introducing Masked Time Series Modeling (MTSM) in the self-supervised pre-training stage, and leveraging a large amount of readily available, unlabeled data on the six component forces at the wheel center and the damper force, the model learns the inherent dynamic laws of the road excitation signal, reducing its reliance on labeled data for docking point loads. A batch of 28-channel activation sequences of length 60 was randomly sampled from the unlabeled dataset. B is the batch size; generate a binary mask matrix. In a random masking process, for 30% to 50% of the time steps (with priority given to continuous masking and loss of analog signal segments), the features at the masked locations are set to zero. ; In the formula, This represents the Hadamard product (element-by-element multiplication). Indicates the outer product operation; This represents the masking matrix, where 0 indicates that the time step is masked and 1 indicates that it is preserved. Represents a 28-dimensional column vector of all ones; This indicates the input timing after the mask.

[0027] The masked sequence The input is fed into an encoder (structure: multi-scale encoder + Transformer encoder) to obtain coded features. ; Through linear reconstruction head Map the encoded features back to the 28-dimensional original signal space. The reconstruction MSE loss is calculated only at the masked locations: ; In the formula, Indicates the sample index within the batch; Indicates the time step index; Indicates the mask matrix in the sample Time step The value is taken at the location; This indicates the total number of masked positions within the entire batch; The reconstructed first... The first sample Incentives for Bus Route 28; This represents the original, true excitation signal.

[0028] In this embodiment, the AdamW optimizer is used to train for 50-100 epochs, and the pre-trained encoder weights are saved as initialization parameters for subsequent supervised training.

[0029] Step S30: Construct a load prediction model. The excitation sequence consisting of the six-component force of the wheel center and the damping force of the shock absorber is subjected to channel weighting and multi-scale time-series coding. The coding features are conditionally modulated using modulation parameters generated from vehicle model parameters. The modulated coding features are input into a Transformer encoder with relative position offset, and the six-axis load prediction sequence of all attachment points is output.

[0030] Specifically, in this embodiment, the dimension of the input batch time-series data tensor is [missing information]. Where B represents the batch sample size, 60 is the sliding window time sequence length, and 28 corresponds to the synchronously acquired six-component force of the four wheel centers and the tensile and compressive excitation characteristics of the four shock absorbers; vehicle model parameter vector C represents the vehicle model parameter dimension. The model consists of the following modules in sequence: 1) Multi-scale temporal convolutional coding module: Dimension The original multivariate temporal input is fed into parallel multi-scale one-dimensional convolution branches. Let the kernel size of the k-th branch be... ( (taking 3, 5, and 7 respectively), the number of output channels is Then the calculation for this branch is: ; In the formula, This represents a one-dimensional convolution, with the kernel length being... Zero padding ensures that the timing length remains unchanged; Represents the activation function of the rectified linear unit; This represents the output feature of the k-th convolutional branch.

[0031] Multi-scale convolutional branches are used to extract local temporal features of road impact signals at different time scales (short kernels capture high-frequency abrupt changes, and long kernels capture mid-to-low-frequency vibration trends). The outputs of each branch are concatenated along the feature dimension: ; In the formula, This indicates a feature dimension concatenation operation; This represents the combined features after multi-branch concatenation. It is then compressed to 128 dimensions via a linear layer. ; In the formula, This represents a fully connected linear layer with an input dimension of 192 and an output dimension of 128. This represents the output feature of convolutional coding.

[0032] 2) Channel Attention Module (ECA): Efficient Channel Attention (ECA) is applied to the 28 input channels. First, the input... Perform global average pooling along the time dimension to obtain channel statistics: ; In the formula, This represents the average value of the d-th input channel over all 60 time steps; This represents a vector composed of statistics from all channels.

[0033] Then to Core size is One-dimensional convolution learns local correlations between channels: ; In the formula, This represents a one-dimensional convolution with a kernel size of 5; This represents the convolution output, which indicates the importance score of each channel.

[0034] The channel weight vector is obtained after Sigmoid activation. Multiply by the original input channel by channel: ; In the formula, This represents the Sigmoid activation function, which outputs weights between 0 and 1. Represents a 60-dimensional column vector of all ones; This represents the input timing after channel attention weighting.

[0035] This module implements adaptive weighting of the contributions of different excitation channels (such as wheel center vertical force, shock absorber force, etc.). The channel attention module can be placed before or after multi-scale convolutional coding; in this embodiment, it is placed before.

[0036] 3) Relative position encoding module: Will Features are fed into the relative position encoding module. Unlike standard sine / cosine absolute position encoding, this invention employs relative position bias (RPB). In the multi-head self-attention attention score calculation, a learnable relative position bias term is added to each pair of query position i and key position j. : ; ; In the formula, Represents the query matrix; Represents the key matrix; Represents a value matrix; Indicates the single-head attention dimension; Scaling; The matrix represents the relative position bias matrix; i represents the query sequence time step; j represents the key sequence time step; w represents the learnable bias parameter vector. : Truncation function, restricting the relative distance to This embodiment .

[0037] This encoding method makes the attention weights explicitly dependent on the relative intervals between time steps, making it more suitable for capturing the local temporal dependencies of road surface stimuli, and it also supports sequence length extrapolation. The output tensor dimension remains the same. .

[0038] 4) Vehicle-specific FiLM modulation module: The standardized vehicle parameter vector The condition vector is mapped through two layers of MLP: ; In the formula, This represents a linear layer, with input vehicle model parameter dimension C and output 64 dimensions. This represents the hidden layer vector of the vehicle model parameters after passing through the MLP.

[0039] Then, channel-wise scaling parameters are generated through two linear layers. and bias parameters (All dimensions are 128): ; In the formula, This represents the scaling factor of FiLM; This represents the offset coefficient of FiLM.

[0040] Features after position encoding Perform Feature-wise Linear Modulation (FiLM): ; In the formula, Representation layer normalization; This indicates element-wise multiplication; t represents the time step. This indicates the characteristics after FiLM modulation is completed.

[0041] This module enables the model to explicitly adjust the feature distribution based on vehicle model parameters (body-in-white mass, inertia, stiffness, hard point coordinates of attachment points, etc.), achieving conditional prediction across vehicle models. New vehicle models only require parameter vectors for low- or zero-sample adaptation. The output tensor dimension is... .

[0042] 5) Transformer encoder module: FiLM modulated Send in A layer-stacked Transformer encoder. The single-layer encoder uses a Pre-LN structure (LayerNorm is placed before the sub-layers), configured... Multi-head self-attention (including the aforementioned relative position bias), feedforward network FFN (128→512→128, GELU activation), Dropout rate The calculation for the l-th layer encoder is as follows: ; ; In the formula, Indicates the first Layer output features; This indicates a multi-head self-attention module; Representation layer normalization; This represents a feedforward neural network; This represents the sum of the output residuals from multi-head attention; This represents the final output of the l-th layer encoder.

[0043] A multi-head self-attention mechanism is used to capture the long-range temporal dependence and multivariate nonlinear coupling relationship between road excitation, damping force, and suspension-body attachment point load over a 60-step time series. Encoder output. Dropout is always enabled during the training phase; during the inference phase, Monte Carlo Dropout (MCDropout) uncertainty estimation is achieved through multiple random forward propagations.

[0044] 6) Full sequence output header module: Encoder output All 60 time-step features are fed into a shared linear output layer. Mapped to the load output space: ; In the formula, This represents the output layer weight matrix; This represents the output layer bias vector; This indicates the predicted output at time t; P represents the total number of suspension-body attachment points; the output is the six-axis dynamic load sequence for all attachment points.

[0045] The full-sequence output paradigm allows the model to obtain the load spectrum of the entire window in a single forward propagation, avoiding the repeated calculations of single-step sliding inference, improving inference efficiency by about an order of magnitude, while the output sequence naturally possesses temporal consistency.

[0046] It should be noted that the IMU vehicle motion signal is not input into the network model, but only participates in the physical constraint loss calculation during the loss function calculation stage; the network input only retains the six-component force of the wheel center, the 28-way excitation of the shock absorber force, and the vehicle parameter vector.

[0047] Step S40: Construct the total loss based on the data fitting error of the predicted sequence and label, the balance error of the resultant force and resultant moment of the six degrees of freedom established according to the motion state of the white body and the position of each attachment point, the spectrum error, and the load change rate exceeding the limit error due to the stiffness of the attachment point, and train the load prediction model.

[0048] Specifically, in this embodiment, the total loss is defined as: ; In the formula, This represents the data fitting term, specifically the mean square error between the predicted load and the true value. This represents the physical constraint loss in the six-degree-of-freedom force and torque balance with the vehicle body as the isolator. This represents the frequency domain amplitude spectrum consistency constraint loss; This represents the physical boundary constraint loss due to the rate of change of load. This represents the weighted hyperparameters of each physical regularization term; typical values. This can be optimized through validation set tuning.

[0049] Among them, the data fitting term This represents the mean square error between the predicted load sequence and the true value sequence, calculated step-by-step and channel-by-channel, and then averaged. ; In the formula, B represents the batch sample size; Indicates the timing window length; Indicates the total number of output payload channels; Indicates the first sample, Time step, number Predicted load on the channel; This represents the truth label payload at the corresponding position.

[0050] ① Force balance constraint terms (Body body isolation body, IMU measured body motion state): Using the body-in-white body as the force isolation body, cut off all suspension-body connection points; the external load of the isolation body is the force and torque of each suspension-body attachment point on the body as predicted by the model.

[0051] There are a total of P suspension-body attachment points; each attachment point exerts a 3-dimensional force on the body. 3D torque , The position vector of the attachment point p relative to the center of mass of the white body is denoted as... (Vehicle model hard point parameters are given). Based on the model-predicted loads, calculate the net external force and net external moment applied to the body-in-white at all attachment points: ; In the formula, P represents the total number of suspension-body attachment points; This represents the force vector acting on the body-in-white at the p-th attachment point, which is analytically obtained from the model prediction output; This represents the torque vector acting on the body-in-white at the p-th attachment point, which is analytically obtained from the model prediction output; This represents the position vector of the attachment point p relative to the center of mass of the white body; Represents the cross product operation of three-dimensional vectors; This represents the resultant force exerted by all suspension attachment points on the body-in-white. It represents the resultant torque of all suspension attachment points on the body-in-white (including the torque of the attachment point itself + the additional torque generated by the lever arm cross product).

[0052] IMU directly measured the motion state of the body-in-white: This represents the three-axis translational acceleration vector of the body-in-white's center of mass (IMU accelerometer output). This represents the three-axis angular velocity vector of the body-in-white (IMU gyroscope output). This represents the three-axis angular acceleration vector of the white body, obtained by time difference of the angular velocity time sequence. Indicates the mass of the body-in-white; The tensor represents the diagonal rotational inertia of the white body about its own center of mass.

[0053] According to the Euler-D'Alembert principle for rigid bodies, the resultant force and resultant torque applied to the vehicle body at the suspension attachment point should balance the inertial force and inertial torque generated by the motion of the body-in-white: ; ; In the formula, This represents the inertial force in the translational theory of the white body; This represents the theoretical moment of inertia during the rotation of the body-in-white. This represents the gyro-coupled torque term; it can be approximately ignored in small tilt conditions.

[0054] ② Force balance constraint loss: Measures the L2 error between the resultant force / resultant moment calculated from the predicted load and the theoretical inertial load of the vehicle body. ; In the formula, Indicates the sample index within the batch; Indicates the first The sample is obtained by calculating the resultant force on the body-in-white by predicting the load at the attachment point using the model; Indicates the first The sample is obtained by calculating the resultant torque on the body-in-white by predicting the load at the attachment point using the model; Indicates the first The theoretical inertial force of the body-in-white calculated from the IMU signal of the sample window; Indicates the first The theoretical inertial torque of the body-in-white is calculated from the IMU signal of the sample window; This represents the squared 2-norm of a vector.

[0055] ③ Frequency domain consistency constraint Fatigue life is highly sensitive to the frequency components of the load spectrum; therefore, consistency between the predicted and true values ​​must be constrained in the frequency domain. Fast Fourier Transform (FFT) is performed on both the predicted load sequence and the true value sequence. ; ; In the formula, Represents the imaginary unit; Indicates the load channel index; Indicates the time step in the time domain; Indicates the FFT frequency point index; Indicates the predicted load number The complex FFT result of the channel; This represents the complex FFT result of the d-th channel of the true load.

[0056] Take amplitude spectrum and Calculate the weighted MSE in the frequency domain: ; In the formula, Representing frequency point The weighting coefficients; a higher weight is set for the 5-20Hz frequency band of the road excitation main frequency, typically taking... The weights for the remaining frequency bands are set to 1. This indicates that taking the modulus of a complex number yields the spectral amplitude.

[0057] ④ Rate of change boundary constraint term The rate of change of load on the suspension-body attachment point bushing is constrained by the physical upper limits of bushing stiffness, damping, and structural inertia. The first-order difference (rate of change of load) is calculated for the predicted load sequence: ; In the formula, Indicates the first sample, aisle, The first-order difference of the load at any given time; characterizing the instantaneous rate of change of the load. Physical upper limit. The stiffness of the bushing at the corresponding attachment point Maximum relative speed with suspension Estimate: ; In the formula, This indicates the bushing stiffness in the direction corresponding to the attachment point; This indicates the maximum relative speed between the suspension and the vehicle body in this direction, which is a reference value based on engineering experience, as shown in Table 1; Indicates the sampling time interval.

[0058] Table 1. Engineering Experience Reference Values ​​for Maximum Relative Motion Velocity at Attachment Points Apply a Softplus differentiable penalty to the rate of change exceeding the physical upper limit: ; ; In the formula, Represents a smooth, differentiable approximation of the ReLU function; when A penalty for loss is incurred; if the limit is not exceeded, the loss approaches zero. This represents the absolute value of the load differential.

[0059] This embodiment employs the following optimizations and training: The total loss scalar is fed into the AdamW optimizer (weight decay). , , + Cosine annealing learning rate scheduler (initial learning rate) Minimum learning rate ,cycle ): ; In the formula, Indicates the current epoch number; Indicates the first Learning rate per epoch; This represents the number of epochs in the cosine annealing cycle.

[0060] Calculate the loss for the trainable parameters of each layer of the network. The gradient is then fed back to all network layers with learnable weights to complete the iterative parameter update: ; In the formula, This represents the gradient of the total loss with respect to the network parameters; This represents all trainable parameters of the model.

[0061] A validation set loss early stopping mechanism (patiently for 10 epochs) is used to preserve the globally optimal network weights. The initial weights for supervised training are loaded from the encoder weights obtained from the self-supervised pre-training in step three. The output head is randomly initialized, and a progressive unfreezing strategy is used to mean that only the output head and the FiLM layer are trained for the first 5 epochs, and then all parameters are unfrozen and fine-tuned together.

[0062] For the inference phase: IMU signals are only used for calculating physical constraint loss during the training phase; when the model infers and predicts the load, it does not need to input IMU signals, only the wheel center six-component force, the shock absorber force, and the vehicle model parameter vector are needed to complete the prediction.

[0063] This application constrains the resultant force and resultant torque of the vehicle body isolation body with the inertial force and inertial torque determined by the motion state of the IMU, so that the prediction results satisfy the rigid body dynamics boundary of the vehicle outside the data fitting, which is beneficial to suppressing non-physical output under extreme conditions.

[0064] Step S50: Obtain the wheel center six-component force, shock absorber damping force and vehicle model parameters of the vehicle to be predicted, input the trained load prediction model according to the time sequence window, and obtain and output the six-axis dynamic load sequence of all suspension-body attachment points of the vehicle to be predicted.

[0065] Specifically, in this embodiment, after completing multiple rounds of iterative training, the globally optimal network weights are saved according to the validation set loss early stopping mechanism; during the actual vehicle road test, 28 excitation timing signals, including the six-component forces at the four wheel centers and the tension and compression of the four shock absorbers, are simultaneously collected. Using preprocessing rules completely consistent with the training set, the 28 real-time collected signals are normalized to... The current vehicle model parameter vector is entered and standardized within the range; no IMU signal input is required during the inference phase.

[0066] A time-series buffer queue of length 60 is set up, and normalized 28-channel activation features are stored at each sampling time step. While the buffer queue is not full for 60 time steps, data is continuously collected, but prediction results are not output. Once the queue is full of 60 sets of continuous time-series data, a sequence with dimension [missing information] is formed. The single inference input sample; every 60 new time steps of data are collected, the 60-step suspension-body full attachment point load prediction sequence is output at once to realize block inference; if higher time resolution is required, overlapping blocks (such as step size 30) can be used and the weighted average of the overlapping areas can be taken. Uncertainty Quantization (MCDropout): Keeps the model's Dropout layer enabled during inference, for the same input Execute T times (typical) Random forward propagation yields T sets of predicted payload sequences. According to the Monte Carlo integral principle, the predicted mean and variance are: ; In the formula, T represents the number of Monte Carlo forward propagations; Let represent the predicted payload obtained from the s-th random forward propagation; This represents the predicted mean, which is used as the final output load of the model. The variance of the model's inherent observation noise is obtained from the residual statistics of the training set. Represents a unit diagonal matrix; This represents the total predicted variance; This represents the predicted standard deviation.

[0067] Further calculations can be performed using the 95% confidence interval: ; The quantiles of the 95% confidence interval of a normal distribution; This represents the 95% confidence upper and lower limits of the load channel at each time step and each attachment point.

[0068] This mechanism provides a confidence level reference for engineering decisions, and low-confidence fragments can guide targeted supplementary sampling tests or virtual iterative verification.

[0069] Reuse training phase: targeting Independent MinMax scalers fitted to the road target loads are used to perform inverse normalization on the mean prediction vector to restore the six-axis dynamic load time sequence of the suspension-body full attachment points with real engineering dimensions (force / kN, torque / N·m); the uncertainty interval is synchronously inverse normalized.

[0070] The output will synchronously store the restored original dimensional load data and corresponding uncertainty intervals by timestamp, and export them as time-series tables or time-domain waveform curves. The output suspension-body full attachment point load spectrum can be directly imported into fatigue life calculation tools (such as nCode and Fe-Safe) for durability verification of the vehicle body and suspension components, and cumulative damage assessment of the entire vehicle. Uncertainty intervals can be used to identify low-confidence prediction segments, guiding targeted supplementary sampling tests or virtual iterative verification.

[0071] This application only requires the collection of easily obtainable wheel center six-component forces, shock absorber damping forces, and key vehicle parameters as inputs to achieve prediction of six-axis loads at all attachment points between the suspension and the vehicle body. This significantly reduces the cost of load acquisition, ensures strict self-consistency of physical constraints and dynamic boundaries, improves the prediction accuracy under extreme conditions, cross-vehicle generalization ability, and engineering application efficiency, and the prediction results can be directly used for fatigue and stress verification of the vehicle body.

[0072] It should be noted that the above description describes some embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0073] Based on the same inventive concept, and corresponding to any of the above embodiments, the embodiments of this application also provide a multi-attachment point load prediction system for automobile suspension.

[0074] like Figure 2 As shown, the vehicle suspension multi-attachment point load prediction system includes: The data acquisition module 11 is configured to acquire the vehicle's synchronous timing data and vehicle model parameters. The synchronous timing data includes the six-component forces at the four wheel centers, the damping forces of each shock absorber, the vehicle body inertial measurement signals, and the six-axis loads at each suspension-body attachment point as tags. The self-supervised pre-training module 12 is configured to perform data preprocessing on synchronous time-series data and vehicle parameters, construct training samples based on a sliding time window, and perform mask reconstruction pre-training on the time-series encoder using unlabeled stimulus sequences. Model building module 13 is configured to build a load prediction model. It performs channel weighting and multi-scale time-series encoding on the excitation sequence consisting of the six-component force of the wheel center and the damping force of the shock absorber. It uses the modulation parameters generated by the vehicle model parameters to conditionally modulate the encoded features. The modulated encoded features are input into a Transformer encoder with relative position offset, and the output is a six-axis load prediction sequence of all attachment points. The model training module 14 is configured to construct the total loss based on the data fitting error of the predicted sequence and label, the balance error of the six-degree-of-freedom resultant force and resultant moment established according to the motion state of the white body and the position of each attachment point, the spectrum error, and the load change rate exceeding the limit error due to the stiffness of the attachment point, and to train the load prediction model. The prediction output module 15 is configured to acquire the wheel center six-component force, shock absorber damping force and vehicle model parameters of the vehicle to be predicted, input the trained load prediction model according to the time sequence window, and obtain and output the six-axis dynamic load sequence of all suspension-body attachment points of the vehicle to be predicted.

[0075] For ease of description, the above system is described by dividing it into various modules based on their functions. Of course, in implementing the embodiments of this application, the functions of each module can be implemented in one or more software and / or hardware.

[0076] The system described in the above embodiments is used to implement the corresponding method in any of the foregoing embodiments and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0077] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, embodiments of this application also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the methods described in any of the above embodiments.

[0078] Figure 3 This embodiment illustrates a more specific hardware structure of an electronic device, which may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0079] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0080] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0081] The input / output interface 1030 is used to connect input / output modules to realize information input and output. The input / output modules can be configured as components in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touch screens, microphones, various sensors, etc., and output devices may include displays, speakers, vibrators, indicator lights, etc.

[0082] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0083] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0084] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0085] The electronic devices described above are used to implement the corresponding methods in any of the foregoing embodiments and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0086] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides a non-transitory computer-readable storage medium that stores computer instructions for causing the computer to perform the methods described in any of the above embodiments.

[0087] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0088] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to perform the methods described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0089] Although this application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0090] The embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the claims of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of this application.

Claims

1. A method of predicting a load at a full attachment point of a vehicle suspension-body, characterized by, include: Acquire vehicle synchronization timing data and vehicle model parameters. The synchronization timing data includes six-component forces at the four wheel centers, damping forces of each shock absorber, vehicle body inertial measurement signals, and six-axis loads at each suspension-body attachment point as tags. The synchronous time-series data and the vehicle parameters are preprocessed, training samples are constructed based on a sliding time window, and the time encoder is pre-trained for mask reconstruction using an unlabeled excitation sequence. A load prediction model is constructed. The excitation sequence consisting of the six component forces of the wheel center and the damping force of the shock absorber is subjected to channel weighting and multi-scale time-series encoding. The encoding features are conditionally modulated using modulation parameters generated by vehicle model parameters. The modulated encoding features are input into a Transformer encoder with relative position offset, and the six-axis load prediction sequence of all attachment points is output. The total loss is constructed based on the data fitting error of the predicted sequence and the label, the balance error of the six-degree-of-freedom resultant force and resultant moment established according to the motion state of the white body and the position of each attachment point, the spectrum error, and the load change rate exceeding the limit error due to the stiffness of the attachment point, and the load prediction model is trained. Obtain the wheel center six-component force, shock absorber damping force and vehicle model parameters of the vehicle to be predicted, input the trained load prediction model according to the time sequence window, and obtain and output the six-axis dynamic load sequence of all suspension-body attachment points of the vehicle to be predicted.

2. The method of claim 1, wherein, The mask reconstruction pre-training includes: By randomly masking consecutive time steps in the unlabeled excitation sequence, the masked excitation sequence is input into a time encoder consisting of a multi-scale time coding module and a Transformer encoder. The coded features are mapped to the excitation signal space by the reconstruction head, and the time encoder is updated according to the error between the reconstructed value and the original value of the masked time step.

3. The method according to claim 1, characterized in that: The load prediction model includes a channel attention module, a multi-scale temporal convolutional coding module, a relative position coding module, a vehicle type conditional modulation module, a Transformer encoder, and a full-sequence output head connected in sequence. The multi-scale temporal convolutional coding module extracts local features of the activation sequence in parallel using one-dimensional convolutional kernels of different lengths and then concatenates and compresses them. The vehicle model conditional modulation module maps vehicle model parameters to scaling and offset parameters for each feature channel in order to linearly modulate the encoded features. The full sequence output header maps the encoded features of each time step within the time window into a multi-dimensional payload vector.

4. The method of claim 1, wherein, The methods for obtaining the balance error of the resultant force and resultant moment of the six degrees of freedom include: The forces and torques exerted on the body-in-white by the preset attachment points are analyzed from the predicted sequence. The predicted resultant force is obtained by summing the forces acting at all attachment points, and the predicted resultant moment is obtained by summing the cross products of the torques acting at each attachment point with the corresponding position vectors and forces. The inertial force is obtained from the mass and translational acceleration of the body-in-white, and the inertial torque is obtained from the moment of inertia, angular velocity and angular acceleration of the body-in-white. The balance error is determined based on the difference between the predicted resultant force and the inertial force, and the difference between the predicted resultant torque and the inertial torque.

5. The method according to claim 1, characterized in that: The spectral error is obtained by performing fast Fourier transforms on the predicted sequence and the label respectively, and calculating the weighted error between the corresponding amplitude spectra, and assigning a higher weight to the preset road excitation main frequency band than other frequency bands.

6. The method according to claim 1, characterized in that: The load change rate exceeding the limit error is obtained by calculating the first-order time difference of the prediction sequence, determining the change rate boundary based on the attachment point bushing stiffness in the corresponding direction, the maximum relative motion speed between the suspension and the vehicle body, and the sampling time interval, and applying a differentiable penalty to the time difference exceeding the change rate boundary.

7. The method according to claim 1, characterized in that: When training the load prediction model, the pre-trained time encoder parameters are used as initialization parameters. First, the full sequence output head and vehicle type conditional modulation module are trained, and then all trainable parameters are unfrozen for joint fine-tuning. The model parameters are updated using an adaptive optimizer with weight decay and a cosine annealing learning rate scheduler, and early stopping and optimal model parameters are saved based on the validation set loss.

8. The method according to claim 1, characterized in that, The output of the six-axis dynamic load sequence includes: During the inference phase, the random deactivation layer in the load prediction model is kept enabled. Multiple random forward propagations are performed on the same excitation sequence and vehicle parameters. The prediction mean and prediction variance of the load channel at each time step and each attachment point are calculated based on the multiple prediction results. The prediction mean is used as the load prediction result and the confidence interval is determined based on the prediction variance. The predicted mean and the confidence interval determined based on the predicted variance are inversely normalized, and the six-dimensional dynamic load sequence with engineering dimensions and its confidence interval are output according to the timestamp.

9. A load prediction system for all attachment points of an automotive suspension-body assembly, characterized in that, include: The data acquisition module is configured to acquire the vehicle's synchronous timing data and vehicle model parameters. The synchronous timing data includes the six-component forces at the four wheel centers, the damping forces of each shock absorber, the vehicle body inertial measurement signals, and the six-axis loads at each suspension-body attachment point as tags. The self-supervised pre-training module is configured to perform data preprocessing on the synchronous time-series data and the vehicle parameters, construct training samples based on a sliding time window, and perform mask reconstruction pre-training on the time-series encoder using unlabeled excitation sequences. The model building module is configured to build a load prediction model. It performs channel weighting and multi-scale time-series encoding on the excitation sequence consisting of the six component forces of the wheel center and the damping force of the shock absorber. It uses the modulation parameters generated by the vehicle model parameters to conditionally modulate the encoded features. The modulated encoded features are input into the Transformer encoder with relative position offset, and the six-axis load prediction sequence of all attachment points is output. The model training module is configured to construct the total loss based on the data fitting error of the predicted sequence and labels, the balance error of the six-degree-of-freedom resultant force and resultant moment established according to the motion state of the white body and the position of each attachment point, the spectrum error, and the load change rate exceeding the limit error limited by the stiffness of the attachment point, and train the load prediction model. The prediction output module is configured to acquire the wheel center six-component force, shock absorber damping force and vehicle model parameters of the vehicle to be predicted, input the trained load prediction model according to the time sequence window, and obtain and output the six-axis dynamic load sequence of all suspension-body attachment points of the vehicle to be predicted.

10. The system according to claim 9, characterized in that: The data acquisition module includes six force sensors respectively installed on the four wheels, force sensors respectively used to detect the dynamic tension and compression of the four shock absorbers, an inertial measurement unit installed near the center of gravity of the body-in-white, and a multi-channel synchronous acquisition unit connected to each sensor and the inertial measurement unit.

Images