Heavy commercial vehicle chassis virtual load spectrum efficient twin system

Abstract

The invention discloses a heavy commercial vehicle chassis virtual load spectrum efficient twin system, and relates to the technical field of automobile chassis. The method specifically comprises the following steps: S1, a load spectrum twinning method based on a virtual iteration method; comprising the following steps: S11, constructing a whole vehicle multi-body dynamic model; s2, an AM-P-sLSTM model is constructed; and S3, a load spectrum twinning method based on the AM-P-eLSTM model is carried out. According to the method, refined construction of the whole vehicle MBD model is achieved, the high-precision rigid-flexible coupling vehicle frame multi-body dynamic model with the flexible body as the main body is established mainly through the vehicle mass center, the rotational inertia and actual measurement data of chassis suspension connecting pieces, and a basis is provided for subsequent virtual iteration and load decomposition; by iterating the road surfaces with different characteristics, the time domain and relative damage error of the iteration signal and the target signal under each road surface are small, the iteration precision is high, and the requirements of subsequent virtual load spectrum twinning and fatigue analysis are met.

Description

Technical Field

[0001] The present invention relates to the technical field of automobile chassis, and in particular to a heavy-duty commercial vehicle chassis virtual load spectrum efficient twin system. Background Art

[0002] The accurate acquisition of load spectrum is the core foundation for carrying out fatigue durability analysis of components, and its data quality directly affects the accuracy and efficiency of fatigue life prediction. Due to its large size and high load capacity, heavy-duty trucks are faced with significant multi-axial alternating loads under complex working conditions. A large number of studies have shown that chassis frame structures (especially welds and weld areas) are prone to fatigue cracks under long-term dynamic loads. This failure mode will directly threaten the reliability and driving safety of the vehicle. Therefore, establishing an efficient and accurate load spectrum acquisition method has important engineering value for the durability analysis and optimization of heavy-duty truck chassis structures. The traditional load spectrum acquisition method mainly adopts a technical route that combines physical experiments with simulation analysis: first, the structural stress concentration area is identified through CAE simulation, and then acceleration sensors and strain gauge groups are deployed at the corresponding positions of the actual vehicle to collect road load data.

[0003] Speed ​​sensors and strain gauges are used to collect road load data. While this method can obtain valid data, it suffers from significant drawbacks such as long testing cycles and high testing costs. According to literature, approximately 35% of testing time in traditional methods is consumed by sensor calibration and data verification. To overcome the limitations of traditional methods, twinning methods based on virtual load spectra have become a common practice in durability engineering in recent years. Existing technologies primarily employ a hybrid modeling approach that couples road testing with multi-body dynamics (MBD). Specifically, actual vehicle testing is first used to obtain acceleration signals from key chassis support points and dynamic stress data from the suspension system. Subsequently, a full-vehicle MBD model incorporating nonlinear suspension characteristics is established. Finally, through inverse dynamics simulation, the measured data is input as boundary conditions into the model, resulting in the generation of a high-fidelity virtual load spectra. This virtual load spectra twinning approach has been proven to significantly reduce the number of physical tests and lower testing costs.

[0004] However, in the inverse solution of the load spectrum based on multi-body dynamics (MBD), if acceleration and stress excitation are directly applied at the wheel center node, numerical singularities will occur in the system dynamic equations, resulting in non-physical oscillations in the time-domain curves of the loads at key connection points, seriously affecting the credibility of the load spectrum. To effectively suppress numerical singularities, two types of regularized solution strategies, the constraint body method and the virtual iteration method (VI), are mainly used in current engineering practice. The constraint body method introduces virtual rigid units to constrain redundant degrees of freedom and establishes an indeterminate system solution equation. Although this method can guarantee the existence of the solution, it will change the original dynamic characteristics of the system and cause distortion of the load transfer path. The VI method has always dominated the field of vehicle load spectrum twinning due to its high accuracy, and a large number of studies have been proposed. For example: Zhou Donglong and Chang Jianlong used the VI method to obtain the equivalent excitation of the key parts of the rear axle, and applied it to the vehicle MBD model to obtain the load spectrum of the key points of the driving rear axle. Finally, the key load spectrum was combined with the finite element analysis results and material fatigue characteristic curve of the rear axle for the life fatigue assessment of the rear axle; Wang Tao et al. used VIM to obtain the cab loading spectrum, and combined it with the strain life (eN) method to calculate the weak position of the cab durability; Li, Donghai et al. established a cab-frame rigid-flex coupling multi-body dynamics model and obtained the fatigue analysis load spectrum through Femfat-Lab virtual iteration. With mass and fatigue damage value as optimization targets, and bending-torsion stiffness and first-order bending-torsion mode as constraints, a particle swarm algorithm was used to perform multi-objective optimization design of cab durability. In summary, VI has the advantage of high efficiency compared to the traditional load spectrum twin method. However, VI is essentially a model-based load identification method, and its disadvantages are summarized as follows: (1) VI involves many differential equations to solve, which is inefficient; (2) as the degree of freedom of the dynamic model increases, there is a dimensional disaster analysis; (3) the transfer function between different vehicle models needs to be recalculated, is not portable, and has poor flexibility.

[0005] Due to the flaws and shortcomings of the VI method, deep learning models have been widely adopted in the field of virtual load spectrum twinning in recent years. This is because they can directly generate the mapping relationship between load and response from data without relying on precise physical models or simplified assumptions, thus circumventing the problems of insufficient model accuracy and difficulty in nonlinear coupling modeling in traditional methods. For example, Chen Weihuan et al. used a spatial attention one-dimensional convolutional neural network deep model (SA-1DCNN) to twin the load spectrum of a vehicle frame and conducted extensive comparisons with multiple different twinning methods and models, demonstrating the engineering practicality and efficiency of this technology. However, the spatial attention mechanism of the SA-1DCNN model proposed by the research institute has not been verified to be sensitive to the local features of multi-source time series signals (such as strain and acceleration), and may ignore long-range dependencies or dynamic coupling effects. In addition, the SAGRU neural network has also been used for efficient twinning of virtual load spectra.

[0006] Among many state-of-the-art deep learning models, the LSTM model, through its unique structure of cell state and gating mechanisms (forget gate, input gate, output gate), can autonomously determine whether information is retained or forgotten, thus demonstrating significant advantages in time series data processing. Numerous studies have demonstrated that the LSTM model maintains stable performance and nonlinear fitting and prediction results even with complex time series data. For example, Wen Changkai et al. introduced a multi-head attention mechanism and ConvLSTM network based on the LSTM model and used the proposed MHA-ConvLSTM model to identify high-intensity load segments in tractor load spectra. However, this study did not fully explore a broad comparison of the MHA-ConvLSTM model, and its lack of robustness to input data noise may limit its stability and accuracy in practical applications. Huang Wenbo et al. used the LSTM model to predict load and strain changes during flight based on aircraft flight parameters. However, this study did not fully verify the generalization ability of the LSTM model and may have over-relied on simulation data and idealized calibration conditions. At the same time, it did not compare and demonstrate with other models, resulting in uncertainty in prediction accuracy and robustness in actual engineering applications. Wei. Xiaolu et al. combined LSTM with the transfer learning model and proposed the TR-LSTM model for predicting high-cycle fatigue SN curves of steel. This study showed that the LSTM model has superiority in the fields of time domain data and fatigue durability technology.

[0007] In summary, through the coordinated design of gating mechanisms and cell states, LSTM achieves accurate modeling of long-term dependencies in time series data, robust noise handling, and rapid response to dynamic changes, making it a promising candidate for application in virtual load spectrum digital twin engineering. However, its memory storage capacity and decision-making limitations limit its in-depth application in this field. Therefore, in the rapidly changing automotive market, a more efficient and accurate approach to vehicle chassis accessory load spectrum twinning is urgently needed. Summary of the Invention

[0008] The purpose of the present invention is to solve the shortcomings of the prior art and to propose an efficient twin system of virtual load spectrum for heavy commercial vehicle chassis.

[0009] In order to achieve the above object, the present invention adopts the following technical solutions:

[0010] An efficient twinning system for virtual load spectrum of heavy-duty commercial vehicle chassis, including:

[0011] S1: Load spectrum twinning method based on virtual iteration method; including: S11: Construction of vehicle multi-body dynamics model, S12: Virtual iteration method and error correction;

[0012] S2: Constructing the AM-P-sLSTM model; including: S21: Constructing the attention mechanism model, S22: Constructing the sLSTM model, S23: Introducing the patch and channel independence mechanism, S24: The overall structure of the AM-P-sLSTM model;

[0013] S3: Load spectrum twin method based on AM-P-eLSTM model; including: S31: training data preparation, S32: model training, S33: model verification, S34: model application.

[0014] Preferably: the S11 includes: S111: constructing a rigid-flexible coupling MBD model based on the modal neutral file of the twin flexible frame using the modal synthesis method; establishing a multi-scale connection system for the connection parts; S112: the established finite element model covers subsystems such as the body-in-white, cargo box, chassis and suspension system; obtaining information related to the mass of the heavy-duty truck through a center of mass and inertia test bench based on the above-mentioned finite element model construction; measuring the characteristic parameters of the elastic elements through bench tests to effectively ensure the accuracy of the dynamic model.

[0015] Furthermore: the S12 includes: S121: performing system identification before iteration; S122: taking the equivalent acceleration spectrum obtained by the damage equivalence principle as the target, performing virtual iteration on each characteristic road surface, so that the dynamic model response spectrum is close to the target spectrum, and obtaining the external point load data required for the virtual fatigue analysis of the cab.

[0016] As a preferred embodiment of the present invention, the step S121 includes: obtaining the frequency domain representations X(ω) and Y(ω) by Fourier transform based on the time domain input and output observation data x(t) and y(t) of the dynamic model, and calculating the system frequency response function H according to the formula mn , establish the input-output frequency domain transfer relationship; the formula is as follows:

[0017] H mn =y(ω) / x(ω)

[0018] Apply a broadband low-amplitude white noise excitation signal in the system linear working area, and construct the multi-input and multi-output transfer function matrix H by the formula mn , and solve its generalized inverse matrix H -1 mn , which provides a mathematical basis for inverse displacement drive problems; the formula is as follows:

[0019]

[0020] In the above two formulas, n is the number of driving channels; m is the number of response channels; S122 includes: based on the target signal Y0(t), obtaining the initial driving spectrum x0(t) by the formula; the formula is as follows:

[0021]

[0022] The dynamic model is driven by the initial driving spectrum x0(t) to obtain the first iterative response spectrum Y1(ω), and the error spectrum ε1 with the target signal is calculated synchronously by the formula; the formula is as follows:

[0023] ε1=Y1(t)-Y0(t)

[0024] Inverse the error spectrum ε1, reasonably set the weighting coefficient a and obtain the next driving spectrum x i+1 (t), repeat the above process until the response spectrum reaches the accuracy requirement; the formula is as follows:

[0025]

[0026] In the process of evaluating iterative accuracy, the final quantitative evaluation is performed based on the comparison in the time domain and frequency domain using the relative damage ratio and root mean square error percentage.

[0027] Based on the above solution: S21 includes: S211: attention score S i ; S212: Calculate attention weight a i ; S213: Calculate the final attention output vector x t .

[0028] Preferably: S of S211 i The calculation formula is as follows:

[0029]

[0030] Let x = (x1, x2, ..., x n ) as the input vector and q as the query vector; use the dot product function to calculate the correlation between each input vector and the query vector;

[0031] The S212 is specifically as follows: normalize the result of the formula using Softmax to obtain the attention weight a i ; The formula is as follows:

[0032]

[0033] The S213 is specifically as follows: according to the attention distribution weight a i , perform weighted average on the output data to get the final output; the formula is as follows:

[0034]

[0035] Furthermore, the cell state update of the LSTM model at time t follows the following mathematical rules:

[0036] c t =f t ⊙c t-1 +i t ⊙Z t

[0037]

[0038] In the above LSTM infrastructure, the input feature x at the current moment is t By parameter matrix (W z ,W i ,W f ,W o ) is connected to the candidate state and the gating system, where the cyclic parameter matrix (R z ,R i ,R f ,R o ) is responsible for outputting the previous hidden layer h t-1 Integrate timing information with each gate; the network component also contains four sets of bias parameters (b z ,b i ,b f ,b o ) is used to adjust the activation threshold.

[0039] Based on the above scheme: the sLSTM introduces three key differences;

[0040] Forget gate: sLSTM introduces an exponential activation function based on the traditional sigmoid function of the forget gate. The formula is as follows:

[0041]

[0042] Input gate: sLSTM uses an exponential function as the input gate, replacing the sigmoid function used in traditional LSTM units:

[0043]

[0044] Normalized state and hidden state: sLSTM introduces a normalized state n t , and modified the calculation of the hidden state:

[0045]

[0046] Among them, the normalized state update expression is:

[0047] n t =f t ⊙n t-1 +i t .

[0048] As a preferred embodiment of the present invention, the S31 uses the Z-Score normalization method to perform dimensionless processing on the original data to address the dimensional heterogeneity problem of displacement and acceleration; the calculation formula is as follows:

[0049] X=x-μ / σ

[0050] Where: x is the original data; μ is the mean; σ is the standard deviation.

[0051] Preferably, the S33 uses three indicators, MAE, MSE, and RMS, to evaluate the generalization ability of the neural network model. The specific expressions are as follows:

[0052]

[0053] In the above formula, y n Indicates the actual value, represents the predicted value, and N represents the total number of samples.

[0054] The beneficial effects of the present invention are:

[0055] 1. This invention achieves a refined construction of the vehicle MBD model. This primarily utilizes measured data on the vehicle's center of mass, moment of inertia, and chassis suspension connectors to establish a highly accurate rigid-flexible coupling multibody dynamics model of the vehicle frame, with the flexible body as the primary component. This model provides a foundation for subsequent virtual iteration and load decomposition. By iterating across different characteristic road surfaces, the time domain and relative damage errors between the iteration signal and the target signal for each road surface are minimal, resulting in high iteration accuracy, meeting the requirements for subsequent virtual load spectrum twinning and fatigue analysis.

[0056] 2. Based on the LSTM model, this paper introduces a weight mechanism, an exponential gating strategy, a patch, and an independent channel mechanism, creatively proposing an efficient and accurate AM-P-sLSTM model. This model demonstrates strong superiority when compared with six advanced prediction models (sLSTM, LSTM, iTransformer, Rliner, SAGRU, and SA-1DCNN) on five of the latest multivariate datasets (weather, electricity, solar energy, ETTm1, and PEM03), especially for non-stationary sequences, achieving an improvement of approximately 29% over the baseline model.

[0057] 3. The comparative study of physical vehicle testing and model training verification shows that all models and the VI method described in this study can successfully obtain the chassis virtual load spectrum. Through multi-dimensional comparison, it is found that the VI method has better twin robustness. However, the AM-P-sLSTM model proposed in this paper has better efficiency in virtual load spectrum twinning while ensuring high accuracy. Compared with the VI method, the efficiency is improved by 35.9% and 45.9% respectively at different cutoff frequencies. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 A local diagram for meshing the finite element modeling of the present invention;

[0059] Figure 2 A three-dimensional schematic diagram of a flexible vehicle frame model of the present invention;

[0060] Figure 3 Schematic diagram of the MBD model of the vehicle of the present invention;

[0061] Figure 4 This is the AM model structure diagram of the present invention;

[0062] Figure 5 This is a schematic diagram of the sLSTM system architecture of the present invention;

[0063] Figure 6 This is the AM-P-sLSTM system architecture diagram of the present invention;

[0064] Figure 7 This is the load spectrum twin flow chart based on the AM-P-sLSTM model of the present invention;

[0065] Figure 8 This is a comparison table of the prediction error results of each model of the present invention in terms of MSE and MAE;

[0066] Figure 9 The load spectrum measured data of the present invention;

[0067] Figure 10 This is a comparison diagram of vertical acceleration data of the twisted C-road frame of the present invention;

[0068] Figure 11 This is a comparison diagram of the displacement data of the twisted C road of the present invention;

[0069] Figure 12 This is a comparison chart of acceleration data on a cobblestone road according to the present invention;

[0070] Figure 13 This is a comparison diagram of the displacement data of the cobblestone road of the present invention;

[0071] Figure 14 This is a comparison table of RMS test results of various models of the present invention;

[0072] Figure 15 This is a comparison result diagram of the torsional C-road suspension force of the present invention;

[0073] Figure 16 This is a comparison diagram of the cobblestone road suspension force of the present invention;

[0074] Figure 17 This is a comparison table of pseudo-damage results of the right front shock absorber mounting point of the present invention;

[0075] Figure 18 This is the stress cloud map of the AM-P-sLSTM model of the present invention;

[0076] Figure 19 The stress cloud diagram of the VI method of the present invention;

[0077] Figure 20 This is a time cost comparison table of various models of the present invention;

[0078] Figure 21 This is a time cost comparison chart of each model of the present invention. DETAILED DESCRIPTION

[0079] The technical solution of the present invention will be further described in detail below in conjunction with specific implementation methods.

[0080] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0081] Example 1:

[0082] An efficient twin system for virtual load spectrum of heavy commercial vehicle chassis, such as Figures 1 to 7 As shown, including:

[0083] S1: Load spectrum twin method based on virtual iteration method.

[0084] S11: Construct a multi-body dynamics model of the entire vehicle; the construction of a high-precision multi-body dynamics (MBD) model of the entire vehicle is the key technical foundation for ensuring the accuracy of virtual iteration.

[0085] Traditional modeling methods use a rigid body assumption based on the mass characteristics of the center of mass point. Its inherent defect is that it cannot characterize the impact of chassis flexible deformation on the system dynamic behavior. This application breaks through the limitations of rigid body modeling by constructing an intrinsic dynamic model that integrates a fully detailed cab subsystem, interior and exterior trim assembly, and a flexible frame.

[0086] S111: Based on the modal neutral file of the twin flexible frame using the modal synthesis method, a rigid-flexible coupling MBD model is constructed. The frame body is discretized using 5mm quadrilateral shell elements. A local mesh encryption strategy (minimum size 2mm) is implemented for areas with complex geometric features. Quality indicators such as the Jacobian determinant (>0.6), aspect ratio (<5), and warpage (<5°) are strictly checked to ensure that the finite element discretization error is controllable.

[0087] S112: Establish a multi-scale connection system for the connection parts; refer to Figure 1 The multi-scale connection system includes: 1) the weld point is modeled by the third-order Spot unit, and the outer core-inner core-beam unit is hierarchically characterized to characterize the material gradient characteristics of the weld zone; 2) the weld connection is realized by the common node Shell unit, and the geometric coordination is ensured by normal vector calibration; 3) the bolt joint is simulated by the Beam-RBE2 hybrid unit to simulate the preload effect, and the Washer unit is set around the bolt hole to eliminate the boundary singularity; 4) the structural adhesive bonding adopts the Hexa-RBE3 composite unit to transmit the multi-directional coupling load.

[0088] S113: The established finite element model covers subsystems such as the body-in-white, cargo box, chassis, and suspension system; a total of 1,571,842 elements and 1,281,103 nodes are generated, with an overall mass of 9,223 kg.

[0089] S114: Based on the above-mentioned finite element model, the mass information of the heavy-duty truck was obtained through the center of mass and inertia test bench. The total mass of the bench test was 9190kg, with an error of about 13% compared with the finite element. The errors of other key parameters such as the moment of inertia and center of mass position were all controlled within 15%, indicating that the finite element model has a high degree of consistency with the actual prototype vehicle parameters.

[0090] In addition, considering the influence of the flexible deformation of the frame on the dynamic response characteristics of the system, the front half of the frame where the cab suspension is installed is cut off to establish a finite element model, and the twin flexible body modal neutral file is used for dynamic modeling; the flexible body frame is as follows Figure 2 shown.

[0091] S115: The characteristic parameters of elastic elements are measured through bench tests to effectively ensure the accuracy of the dynamic model.

[0092] Based on the above content, a high-precision rigid-flexible coupling multi-body dynamics model including a flexible cab, a flexible frame, and a vehicle suspension is established in Adams software. Figure 3 shown.

[0093] S12: Virtual Iteration Method and Error Correction; As a semi-analytical load decomposition method, the core mechanism of the virtual iteration method is to establish a displacement-driven closed loop between the measured load and the nonlinear dynamic model.

[0094] This method solves the displacement excitation function that satisfies the measured external point load constraints by inverse engineering, significantly reducing the computational complexity while ensuring the load reconstruction accuracy, thereby achieving coordinated optimization of accuracy and computational efficiency.

[0095] S121: Perform system identification before iteration. First, based on the time domain input and output observation data x(t) and y(t) of the dynamic model, obtain the frequency domain representation X(ω) and Y(ω) through Fourier transform, and calculate the system frequency response function H according to formula (1): mn , establish the input-output frequency domain transfer relationship.

[0096] Then, a broadband low-amplitude white noise excitation signal is applied in the linear working area of ​​the system, and the multi-input multi-output transfer function matrix H is constructed by formula (2): mn , and solve its generalized inverse matrix H -1 mn , providing a mathematical basis for inverse displacement drive problems.

[0097] H mn =y(ω) / x(ω) (1)

[0098]

[0099] Where: n is the number of driving channels; m is the number of response channels.

[0100] S122: Using the equivalent acceleration spectrum obtained from the damage equivalence principle as the target, perform virtual iterations on each characteristic road surface to make the dynamic model response spectrum close to the target spectrum, thereby obtaining the external point load data required for virtual fatigue analysis of the cab.

[0101] First, based on the target signal Y0(t), the initial driving spectrum x0(t) is obtained by equation (3).

[0102]

[0103] Subsequently, the dynamic model is driven by the initial driving spectrum x0(t) to obtain the first iterative response spectrum Y1(ω), which is then calculated synchronously with the target signal error spectrum ε1 through equation (4).

[0104] ε1=Y1(t)-Y0(t) (4)

[0105] Finally, the error spectrum ε1 is inverted, the weighting coefficient a is reasonably set, and the next driving spectrum x is obtained. i+1 (t), and repeat the above process until the response spectrum reaches the accuracy requirement.

[0106]

[0107] In the process of evaluating iterative accuracy, based on the comparison in the time domain and frequency domain, a final quantitative evaluation needs to be performed using the relative damage ratio and root mean square error percentage.

[0108] The final results show that the error between the measured signal and the iterative signal in the iterative results of different characteristic pavements meets the requirements. At the same time, the changes in the time domain and amplitude of the monitoring signal are consistent, indicating that the established MBD model has good accuracy and meets the requirements of subsequent load twinning and fatigue analysis.

[0109] S2: Build the AM-P-sLSTM model.

[0110] S21: Construct an attention mechanism model. The attention mechanism (AM) model aims to simulate the human brain's attention allocation mechanism and establish a framework for enhancing and suppressing dynamic features. In recent years, it has been increasingly applied to deep learning. Its essence is to focus on more important features during model training and assign greater weight to them. Less important information is assigned a smaller weight to reduce sensitivity to secondary information, thereby improving the model's feature representation capabilities within the constraints of limited computing resources.

[0111] The structure of the AM model is as follows Figure 4 As shown. The main calculation process of the attention value is divided into the following three steps:

[0112] S211: Attention score S i .

[0113]

[0114] Let x = (x1, x2, ..., x n) as the input vector and q as the query vector; the dot product function is used to calculate the correlation between each input vector and the query vector.

[0115] S212: Calculate attention weight a i .

[0116] Use Softmax to normalize the result of the formula and get the attention weight a i ;

[0117]

[0118] S213: Calculate the final attention output vector x t ; According to the attention distribution weight a i , perform weighted average on the output data to get the final output. The calculation formula is:

[0119]

[0120] S22: Build an sLSTM model.

[0121] S221: LSTM model; The long short-term memory network (LSTM) is designed to solve the gradient vanishing problem in traditional recurrent neural networks (RNNs) and can effectively capture long-range dependencies in time series data. Its core structure implements information regulation through three gating mechanisms: a forget gate responsible for filtering historical information, an input gate that controls the entry of new information, and an output gate that determines the current output.

[0122] The unit state update of the network at time t follows the following mathematical rules:

[0123] c t =f t ⊙c t-1 +i t ⊙Z t (9)

[0124]

[0125]

[0126] In the above LSTM infrastructure, the input feature x at the current moment is t By parameter matrix (W z ,W i ,W f ,W o ) is connected to the candidate state and the gating system, where the cyclic parameter matrix (R z ,R i ,R f ,R o ) is responsible for outputting the previous hidden layer ht-1 Integrate timing information with each gate; the network component also contains four sets of bias parameters (b z ,b i ,b f ,b o ) is used to adjust the activation threshold. Candidate state generation module The hidden layer activation module ψ usually adopts a hyperbolic tangent function with nonlinear saturation characteristics. This function effectively suppresses the numerical divergence of the hidden layer state by constraining the output value range to the (-1,1) interval.

[0127] Its mathematical form can be expressed as:

[0128]

[0129] In the standard LSTM network, the activation function of all gates is the sigmoid function, and the function formula is as follows:

[0130]

[0131] S222: sLSTM model; Extensive literature demonstrates that exponential gating, because its derivative is the function itself, allows its gradient to flow more easily during backpropagation, further helping to alleviate the vanishing gradient problem often encountered with sigmoid function-based gating. Therefore, the new sLSTM model enhances the LSTM architecture by introducing exponential gating and a novel memory hybrid mechanism.

[0132] Compared with the traditional LSTM architecture, sLSTM introduces several key differences: Forget gate: sLSTM introduces an exponential activation function based on the traditional sigmoid function of the forget gate, which is expressed as:

[0133]

[0134] Input gate: sLSTM uses an exponential function as the input gate, replacing the sigmoid function used in traditional LSTM units:

[0135]

[0136] Normalized state and hidden state: sLSTM introduces a normalized state n t , and modified the calculation of the hidden state:

[0137]

[0138] Among them, the normalized state update expression is:

[0139] n t =f t ⊙nt-1 +i t (twenty one)

[0140] These modifications to the exponential gating in the forget gate and input gate provide greater flexibility in controlling the flow of information, while the normalizer state helps stabilize the hidden state computation over long sequences. Figure 5 shown.

[0141] S23: Introducing patch and channel independence mechanisms.

[0142] After adjusting the activation function of LSTM, it still lacks the ability to capture long-term dependencies; introducing patch and channel-independent mechanisms has been shown to be able to better improve the ability to input long sequences or large-size data and prevent overfitting problems.

[0143] The patch mechanism enables the model to capture longer historical sequences, thereby significantly improving the prediction performance. Its working principle is: the input time series is divided into independent patches, the length of each patch is set to P, and the non-overlap length between two consecutive patches is S. The patching process will generate a patch sequence Where N represents the number of patches. And before patching, the last value The original sequence is padded with S repeated digits. This reduces the length of the input time series from L to L / S. Channel independence, on the other hand, means that each input token contains information from only a single channel. This approach has been proven effective in convolutional neural networks (CNNs) and linear models, but has not yet been applied to this technical field.

[0144] S24: Overall structure of the AM-P-sLSTM model; In summary, this paper innovatively constructs a new AM-P-sLSTM model. The overall structure of the model is as follows Figure 6 shown.

[0145] S3: Load spectrum twin method based on AM-P-eLSTM model.

[0146] The load spectrum twin method based on the AM-P-eLSTM model is as follows Figure 7 The specific steps are as follows:

[0147] S31: Training data preparation; Apply vertical displacement excitation to the wheel center of the vehicle MBD simulation model, and then solve the MBD model's wire sensor displacement and axle head acceleration simulation response. 2 ), the Z-Score standardization method was used to perform dimensionless processing on the original data.

[0148] The Z-Score normalization method is used to eliminate scale differences and optimize the training convergence and generalization performance of the deep learning model. The calculation formula is as follows:

[0149] X=x-μ / σ (22)

[0150] Where: x is the original data; μ is the mean; σ is the standard deviation.

[0151] Then, the normalized load spectrum response signal and the wheel center vertical displacement excitation signal in the simulation are used as the training set, verification set and test set, accounting for 70%, 20% and 10% respectively.

[0152] S32: Model training; the response signal obtained after processing in step S31 is used as the input of the neural network, and the corresponding simulated wheel center vertical excitation signal is used as the output of the neural network. Figure 5 The structure shown achieves long-term dependency modeling of time series data through exponential gating mechanism and gradient optimization: first, the time series data is preprocessed into input sequences and divided into batches and the corresponding weights are applied, and then forward propagation is performed; each time step updates the cell state and hidden state through the forget gate, input gate and output gate in turn to generate a prediction result; the loss is defined based on the root mean square error calculated based on the predicted value and the true label, and the gradient is propagated back layer by layer from the terminal time step using the time backpropagation algorithm, the partial derivatives are calculated for the weight matrix and bias parameters, and the gradient range is constrained by gradient clipping; the network parameters are updated using an optimizer (such as Adam) combined with a learning rate strategy; and finally, the model is converged on the validation set through multiple rounds of iterations.

[0153] S33: Model verification: Input the input data of the test set into the trained neural network, and compare the output of the neural network with the true value.

[0154] Based on the test and subsequent verification requirements, MAE, MSE, and RMS are used to evaluate the generalization ability of the neural network model. The specific expressions are as follows:

[0155]

[0156]

[0157] In the above formula, y n Indicates the actual value, represents the predicted value, and N represents the total number of samples.

[0158] S34: Model Application: Based on the measured cable sensor displacement and axle head vertical acceleration signals, standardized preprocessing and then input into a fully trained deep neural network model, the vertical displacement excitation spectra of the 12 wheel hubs are inversely solved. This excitation spectrum is coupled with the measured six-degree-of-freedom load spectrum of the wheel hub to drive a high-precision vehicle multi-body dynamics (MBD) model, enabling load spectrum reconstruction and structural response monitoring at key frame nodes. A closed-loop verification framework for simulation and measured data is constructed to verify the goodness of fit of the time-domain signals of axle head acceleration and cable displacement. If the time-domain RMS relative error is less than 20%, the load inversion accuracy is determined to be satisfactory. Otherwise, the neural network weights are iteratively optimized until the engineering accuracy requirements are met.

[0159] Example 2:

[0160] Based on the first embodiment, the heavy-duty commercial vehicle chassis virtual load spectrum efficient twin system was tested and the results were analyzed. Figures 8 to 21 The specific steps are as follows:

[0161] S1: Comprehensive performance comparison test of AM-P-sLSTM model.

[0162] To further evaluate the superiority of the AM-P-sLSTM algorithm proposed in this paper, this paper conducted experimental comparisons with multiple prediction models (sLSTM, LSTM, iTransformer, Rliner, SAGRU, SA-1DCNN) on five latest multivariate datasets (weather, electricity, solar energy, ETTm1 and PEM03).

[0163] The evaluation indicators are shown in Equations (23) and (24). The lookback window of the PEMS dataset is set to L = 96, and the prediction length T∈{96,192,336,720}; the prediction length of other datasets is T∈{12,24,48,96}, and the lookback window is set to L = 336.

[0164] The final test results are as follows Figure 8 As shown in the table, the indicator champion is indicated by red bold font, and the runner-up is indicated by underline and bold font.

[0165] Figure 8The experimental results in the table show that the models show significant performance differences in cross-domain time series prediction tasks: the traditional recursive model (sLSTM / LSTM) is limited by the gradient decay problem and has larger errors than other models in long-term prediction. For example, the MAE in ETTm1 / 720 steps is 0.749 and 0.801 respectively; while the temporal attention model (SAGRU) performs well in short-cycle patterns (Electricity / 96 steps MSE is 0.152 and 0.172 respectively), it over-relies on the stationary assumption, which leads to a lag in the detection of sudden events; the convolutional architecture (SA-1DCNN) has weak long-term correlation modeling capabilities due to the limited model field of view (ETTm1 / 720 steps MAE = 0.613), and the linear baseline model ( Rliner) performs poorly in complex scenarios due to ignoring nonlinear dynamic characteristics (ETTm1 / 336-step MAE = 0.389). In comparison, AM-P-sLSTM shows significant advantages in long-term prediction through its patch mechanism, independent channel twins, and weight-guided attention mechanism. Its Weather / 720-step MAE (0.346) is 12.3% lower than that of the suboptimal model, and it maintains the lowest error standard deviation (MAE ≤ 0.051) in cross-domain tasks. In particular, it improves by 29% compared with the baseline model for non-stationary sequences (such as PEMS03 / 48-step MSE = 0.171), verifying the effectiveness of progressive memory decay control and multi-granularity feature fusion strategy in suppressing error accumulation and enhancing model generalization ability.

[0166] S2: Physical vehicle data collection test.

[0167] Data collection tests were conducted on a physical vehicle. The onboard sensors included six-component wheel center force sensors, axle head acceleration sensors (used to measure vertical acceleration at the left and right ends of the front and rear axles), and cable displacement sensors positioned at the shock absorbers. The collected signals included the six-component wheel center force of each of the 12 wheels, displacement signals from the six cable sensors in the front, rear, left, and right directions, and axle head acceleration signals. The six-component wheel center force of the 12 wheels served as the excitation signal for load twinning, while the cable sensor displacements and axle head acceleration signals were used to inversely determine the wheel center vertical displacement excitation, thereby evaluating the accuracy of load spectrum twinning.

[0168] The road load spectrum measurement work is carried out in the automobile durability test field. The signal sampling frequency is 1000Hz. The test is repeated three times. One set of test data is used for load spectrum twinning, and the other test data is used as a backup.

[0169] Before the load spectrum test, the physical prototype vehicle status must be checked to ensure that the physical prototype vehicle is consistent with the design status. Then, according to the corresponding test specifications, the load spectrum is collected on different reinforced road surfaces at a specified speed under full load.

[0170] In terms of raw data processing, in order to eliminate drift and glitches in the measured signal caused by factors such as temperature changes in the test equipment and electromagnetic interference, the raw signal needs to be pre-processed by deburring, de-drifting and filtering, and the filtering frequency is 256Hz.

[0171] The test road types include washboard road, cobblestone road, curb impact road, twisted road, billy stone road and brake road, etc.; the wheel center six-component load spectrum actual vehicle collection data of some road conditions are as follows Figure 9 The load spectrum is shown in the measured data.

[0172] S3: Comparison and analysis of load spectrum twin test results.

[0173] In order to verify the effectiveness and superiority of the neural network model proposed in this application, the sLSTM, LSTM and other advanced neural network structures described in step S2 of the synchronous embodiment 1 and the VI method described in the background technology are used for load spectrum twinning and comparative analysis is performed.

[0174] The comparison data includes the vertical acceleration and displacement of the axle, and the strain at key positions of the frame. The comparison content includes RMSE and pseudo-damage value. This article mainly describes the comparison working conditions of typical cobblestone roads and twisted road C conditions.

[0175] S31: Comparison results of vehicle frame vertical acceleration and shock absorber displacement data.

[0176] Use respectively Figure 9 The various models shown simulate the load conditions of the MBD model on a cobblestone road and a twisting road; the frame vertical acceleration and shock absorber displacement data are compared with the data collected from the physical vehicle.

[0177] Figure 10 and Figure 12 The comparison results of the vertical acceleration of the frame under two types of road conditions are shown. Figure 11 and Figure 13 Shows the comparison results of shock absorber displacement data under two types of working conditions.

[0178] The comparison results show that the simulation results of all simulation methods are highly consistent with the physical measured test results in the time domain.

[0179] The RMSE value and the relative error between simulation and test are as follows Figure 14 The RMS test results comparison table of each model is shown in the figure. The RMS calculation results also show that the relative errors of the RMS values ​​obtained by the simulation and VI method at the monitoring location are within 20%. Figure 14The calculation results show that the simulation results of VI, AM-P-sLSTM and other models are highly consistent with the test results. In addition, the error of AM-P-sLSTM is smaller than that of sLSTM, and the VI method can achieve better error control stability.

[0180] S32: Comparison results of vehicle load spectrum and pseudo damage value.

[0181] Pseudo damage is a simplified indicator used to evaluate the cumulative fatigue damage of structures or materials under dynamic loads. It is commonly used in engineering vibration testing and durability analysis. Pseudo damage is usually not accurate damage, but it can approximately characterize the relative strength of the damage effect between load spectra.

[0182] The calculation formula of the pseudo damage value is as follows:

[0183]

[0184] Among them, S i represents the amplitude of the i-th stress, n i Indicates the corresponding stress amplitude S i The number of cycles, m and c represent the slope and intercept of the SN curve, respectively. In this study, 7 and 4.8×10 14 .

[0185] In summary, this paper collects load spectrum data for the suspension force at the right front shock absorber installation point, and the collection and comparison results are as follows: Figure 15 and Figure 16 shown.

[0186] based on Figure 15 and Figure 16 The data comparison results further Figure 8 The pseudo damage values ​​of the model and VI method shown are compared with the actual collected stress. The comparison results are shown in Figure 17 The comparison table of pseudo damage results of the right front shock absorber mounting point is shown in FIG.

[0187] according to Figure 17 The comparison results of pseudo-damage value calculation data show that the pseudo-damage prediction errors of all models are controlled within 10%, among which the VI model has the best overall performance (1% error on cobblestone roads and 3% error on long-wave roads). The AM-P-sLSTM model has an error of 3% under both road conditions, which is very close to the VI model and shows similar high-precision characteristics. The errors of other models such as iTransformer and SA-1DCNN are between 2% and 8%, and the overall stability is slightly worse, but they all meet the engineering error requirements.

[0188] S33: Comparison and analysis of finite element stress of vehicle frame assembly.

[0189] The principle of stress solution for vehicle frame simulation is to first discretize the vehicle frame geometric model into a finite element mesh through simulation software, assign material properties, impose constraints and load spectrum; based on the principle of continuum mechanics, the complex structure is decomposed into multiple units, and a linear equation group of node displacements is established through equilibrium equations and constitutive relations. The displacement field is solved using numerical methods and then converted into unit stress and strain distribution.

[0190] The final calculation results of the maximum stress cloud diagram of the frame based on the load spectrum finite element simulation are as follows Figure 18 and Figure 19 As shown in the figure, it can be seen that the maximum stress results of the frame assembly calculated by AM-P-sLSTM and VI's load spectrum twin method are basically consistent.

[0191] To further validate the load spectrum twinning method, a data acquisition prototype vehicle was placed in a durability test field for a full-cycle road test. The location of the maximum stress essentially coincided with the road test fracture result. The combined results of finite element simulation and actual road testing further demonstrated that the AM-P-sLSTM model for load spectrum twinning is feasible and effective.

[0192] S34: Comparative analysis of solution time of different load twinning methods.

[0193] In order to study the differences in solution efficiency among AM-P-eLSTM, VI method and other models, the modal cutoff frequencies of the flexible body of the frame component in MBD are set to 1000 Hz and 3000 Hz, respectively, for load spectrum twinning.

[0194] Depend on Figure 9 The total time comparison of the twin load spectrum in the test road is shown as follows Figure 20 and Figure 21 shown.

[0195] The time cost of deep learning methods primarily comes from generating the training dataset, which typically involves running the MBD model, processing the data, and adjusting the parameters and training the deep learning model. The time cost of VI methods primarily comes from calculating the transfer function and iteratively solving the vertical wheel center displacement, which primarily involves running the MBD model and adjusting the error.

[0196] from Figure 20 and Figure 21 It can be seen that when the cutoff frequency of the flexible body is 1000Hz and 3000Hz, the efficiency of the AM-P-sLSTM model method is 35.9% and 45.9% higher than that of the VI method, respectively. As the complexity of the MBD model increases, the efficiency advantage of the AM-P-sLSTM deep learning method becomes more obvious. The main reason for the relatively low efficiency of the VI method is that VI involves more differential equation solution loops.

[0197] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. An efficient twinning system for virtual load spectrum of heavy commercial vehicle chassis, characterized by: include: S1: Load spectrum twinning method based on virtual iteration method; including: S11: Construction of vehicle multi-body dynamics model, S12: Virtual iteration method and error correction; S2: Constructing the AM-P-sLSTM model; including: S21: Constructing the attention mechanism model, S22: Constructing the sLSTM model, S23: Introducing the patch and channel independence mechanism, S24: The overall structure of the AM-P-sLSTM model; S3: Load spectrum twin method based on AM-P-eLSTM model; including: S31: training data preparation, S32: model training, S33: model verification, S34: model application.

2. The heavy-duty commercial vehicle chassis virtual load spectrum efficient twin system according to claim 1 is characterized by: The S11 includes: S111: constructing a rigid-flexible coupling MBD model based on the modal neutral file of the twin flexible frame using the modal synthesis method; establishing a multi-scale connection system for the connection parts; S112: the established finite element model covers subsystems such as the body-in-white, cargo box, chassis and suspension system; obtaining information related to the mass of the heavy-duty truck through a center of mass and inertia test bench based on the above-mentioned finite element model construction; measuring the characteristic parameters of the elastic elements through bench tests to effectively ensure the accuracy of the dynamic model.

3. The heavy-duty commercial vehicle chassis virtual load spectrum efficient twin system according to claim 1 is characterized by: The S12 includes: S121: performing system identification before iteration; S122: taking the equivalent acceleration spectrum obtained by the damage equivalence principle as the target, performing virtual iteration on each characteristic road surface, so that the dynamic model response spectrum is close to the target spectrum, and obtaining the external point load data required for the virtual fatigue analysis of the cab.

4. The heavy-duty commercial vehicle chassis virtual load spectrum efficient twin system according to claim 3 is characterized by: The step S121 includes: obtaining the frequency domain representations X(ω) and Y(ω) based on the time domain input and output observation data x(t) and y(t) of the dynamic model through Fourier transform, and calculating the system frequency response function H according to the formula mn , establish the input-output frequency domain transfer relationship; the formula is as follows: H mn =y(ω) / x(ω) Apply a broadband low-amplitude white noise excitation signal in the system linear working area, and construct the multi-input and multi-output transfer function matrix H by the formula mn , and solve its generalized inverse matrix H -1 mn , which provides a mathematical basis for inverse displacement drive problems; the formula is as follows: In the above two formulas, n is the number of driving channels; m is the number of response channels; S122 includes: based on the target signal Y0(t), obtaining the initial driving spectrum x0(t) by the formula; the formula is as follows: The dynamic model is driven by the initial driving spectrum x0(t) to obtain the first iterative response spectrum Y1(ω), and the error spectrum ε1 with the target signal is calculated synchronously by the formula; the formula is as follows: ε1=Y1(t)-Y0(t) Inverse the error spectrum ε1, reasonably set the weighting coefficient a and obtain the next driving spectrum x i+1 (t), repeat the above process until the response spectrum reaches the accuracy requirement; the formula is as follows: In the process of evaluating iterative accuracy, the final quantitative evaluation is performed based on the comparison in the time domain and frequency domain using the relative damage ratio and root mean square error percentage.

5. The heavy-duty commercial vehicle chassis virtual load spectrum efficient twin system according to claim 1 is characterized by: The S21 includes: S211: attention score S i ; S212: Calculate attention weight a i ; S213: Calculate the final attention output vector x t .

6. The heavy-duty commercial vehicle chassis virtual load spectrum efficient twin system according to claim 5, characterized in that: The S211 S i The calculation formula is as follows: Let x = (x1, x2, ..., x n ) as the input vector and q as the query vector; use the dot product function to calculate the correlation between each input vector and the query vector; The S212 is specifically as follows: normalize the result of the formula using Softmax to obtain the attention weight a i ; The formula is as follows: The S213 is specifically as follows: according to the attention distribution weight a i , perform weighted average on the output data to get the final output; the formula is as follows:

7. The heavy-duty commercial vehicle chassis virtual load spectrum efficient twin system according to claim 1 is characterized by: The cell state update of the LSTM model at time t follows the following mathematical rules: c t =f t ⊙c t-1 +i t ⊙Z t In the above LSTM infrastructure, the input feature x at the current moment is t By parameter matrix (W z ,W i ,W f ,W o ) is connected to the candidate state and the gating system, where the cyclic parameter matrix (R z ,R i ,R f ,R o ) is responsible for outputting the previous hidden layer h t-1 Integrate timing information with each gate; the network component also contains four sets of bias parameters (b z ,b i ,b f ,b o ) is used to adjust the activation threshold.

8. The heavy-duty commercial vehicle chassis virtual load spectrum efficient twin system according to claim 1 is characterized by: The sLSTM introduces three key differences; Forget gate: sLSTM introduces an exponential activation function based on the traditional sigmoid function of the forget gate. The formula is as follows: Input gate: sLSTM uses an exponential function as the input gate, replacing the sigmoid function used in traditional LSTM units: Normalized state and hidden state: sLSTM introduces a normalized state n t , and modified the calculation of the hidden state: Among them, the normalized state update expression is: n t =f t ⊙n t-1 +i t 。 9. The heavy-duty commercial vehicle chassis virtual load spectrum efficient twin system according to claim 1, characterized in that: The S31 solves the dimensional heterogeneity problem of displacement and acceleration by using the Z-Score normalization method to perform dimensionless processing on the original data. The calculation formula is as follows: X=x-μ / σ Where: x is the original data; μ is the mean; σ is the standard deviation.

10. The heavy-duty commercial vehicle chassis virtual load spectrum efficient twin system according to claim 1, characterized in that: The S33 uses MAE, MSE, and RMS indicators to evaluate the generalization ability of the neural network model. The specific expressions are as follows: In the above formula, y n Indicates the actual value, represents the predicted value, and N represents the total number of samples.

Images