Intelligent tire force estimation method, device and system

Abstract

The application provides a kind of intelligent tire force estimation method, device and system, method includes: real-time acquisition tire triaxial acceleration signal and vehicle working condition information;Based on vehicle working condition information, the triaxial acceleration signal of tire is handled, and three-channel input sequence is obtained;Through the force estimation model of offline training, tire trirr force estimation value and corresponding uncertainty index are output;Based on vehicle working condition information, wear index is output;Based on target domain unlabeled sample and source domain labeled sample in the specified period of time recently, and wear index, the deviation degree of target domain and source domain is determined, and online gating coefficient is determined in combination with uncertainty index;Based on online gating coefficient, model is updated on line domain self-adaptation or using safety output strategy to output tire trirr force estimation value, can realize cross wear evolution migration under the condition of weak labeling or no labeling in target domain, realize online credible inference and tire force estimation of safety mechanism constraint.

Description

Technical Field

[0001] This application relates to the field of vehicle dynamics technology, and in particular to an intelligent tire force estimation method, device and system. Background Technology

[0002] Tire forces are crucial fundamental quantities for vehicle braking, steering, stability control, and adhesion utilization. Directly measuring the three-part force of a tire typically relies on a six-part force rim, wheel center force sensor, or test bench, but this method suffers from high cost, complex installation, and insufficient durability and mass production compatibility. Indirect estimation based on mass-producible sensor signals has become an important engineering approach. Existing indirect estimation methods include parameter identification methods based on vehicle and tire dynamics models and observers, filtering-based state estimation methods, and data-driven regression methods. Data-driven methods can utilize high-frequency signals such as intelligent tire acceleration to characterize the relationship between ground impact, structural vibration, and load changes; however, these methods often implicitly assume that training data and application data are distributed similarly. In real-world road and test scenarios, tire wear, tire pressure, tire temperature, load, and speed changes cause systematic drift in the amplitude, spectral energy distribution, and temporal structure of acceleration signals. This results in significantly increased errors in models trained on new tires or under single wear conditions at other wear stages, and even output oscillations. Furthermore, the cost of obtaining the three-part force labels for the target tire wear domain is high, and continuous labeling is difficult; if the model is updated online directly during vehicle operation, it may introduce control instability and safety risks. Summary of the Invention

[0003] The purpose of this application is to provide an intelligent tire force estimation method, device and system that can realize cross-wear evolution migration under weak or unlabeled conditions in the target domain, and realize tire force estimation with online reliable reasoning and safety mechanism constraints.

[0004] Firstly, this application provides an intelligent tire force estimation method, which includes: real-time acquisition of tire triaxial acceleration signals and vehicle operating condition information and time alignment; based on the vehicle operating condition information, sequentially performing grounding event localization, contact segment extraction, and linear interpolation resampling on the tire triaxial acceleration signals to obtain a three-channel input sequence; inputting the three-channel input sequence into an offline-trained force estimation model, outputting tire three-force estimates and corresponding uncertainty indices; and outputting wear indices based on vehicle operating condition information; determining the degree of deviation between the target domain and the source domain based on unlabeled samples and labeled samples in the target domain within the most recent specified time period, as well as the wear indices, and determining online gating coefficients in conjunction with the uncertainty indices; and based on the online gating coefficients, performing online domain adaptive updates on the model or adopting a safe output strategy to output tire three-force estimates.

[0005] Furthermore, the steps described above, which involve sequentially performing grounding event localization, contact segment extraction, and linear interpolation resampling on the tire triaxial acceleration signal based on vehicle operating condition information to obtain a three-channel input sequence, include: obtaining a jump signal by calculating the time derivative of radial acceleration, detecting transient peaks at the entry and exit points from grounding, and calculating the contact angle and contact center point by combining the angular velocity in the vehicle operating condition information; using the contact center point as a reference, extracting acceleration segments within a specified angle range before and after the contact center point from the tire triaxial acceleration signal as valid segments; resampling each valid segment using linear interpolation to a specified number of time steps to form a fixed-length three-channel input sequence; or further including: using the sequence after normalizing the maximum absolute value of the three-channel input sequence as the model input.

[0006] Furthermore, the aforementioned force estimation model includes: a regression network of a feature extraction network; the feature extraction network includes: a cascaded structure of a temporal convolutional network and a bidirectional long short-term memory network that integrates a multi-head self-attention mechanism; the step of inputting the three-channel input sequence into the offline-trained force estimation model and outputting the tire three-part force estimates includes: extracting features from the three-channel input sequence through the cascaded structure of the temporal convolutional network and the bidirectional long short-term memory network to obtain a temporal feature sequence; performing residual connection, random deactivation, and normalization processing on the temporal feature sequence through a multi-head self-attention mechanism to obtain enhanced temporal features; and performing regression processing on the enhanced temporal features through a regression network to obtain the tire three-part force estimates.

[0007] Furthermore, the aforementioned wear indicators include wear embedding vectors and / or wear status labels; the step of outputting wear indicators based on vehicle operating condition information includes: summarizing vehicle operating condition information over a specified time scale to determine operating condition statistical characteristics; the operating condition statistical characteristics include: the mean and variance of the contact angle over time, the radial impact peak distribution, the spectral energy ratio, and the rate of change of peak-to-peak value and envelope within the period; based on vehicle operating condition information, updating the operating condition statistical characteristics online through a sliding window and mapping them to wear embedding vectors; and determining wear status labels based on the wear embedding vectors.

[0008] Furthermore, the steps described above for determining the degree of deviation between the target domain and the source domain based on unlabeled samples of the target domain and labeled samples of the source domain within the most recent specified time period, as well as wear indicators, include: obtaining unlabeled samples of the target domain and labeled samples of the source domain within the most recent specified time period; outputting shared feature vectors corresponding to the labeled samples of the source domain and unlabeled samples of the target domain through a feature extraction network; selecting or updating student kernel parameters according to wear indicators and operating condition statistical features; calculating MMD and VDR between the source domain and the target domain based on the updated student kernel parameters and shared feature vectors, respectively; and obtaining the domain adaptive loss by weighting MMD and VDR using uncertainty weighting coefficients to determine the degree of deviation.

[0009] Furthermore, the steps described above for determining the degree of deviation between the target domain and the source domain, and for determining the online gating coefficients in conjunction with uncertainty indicators, include: determining the trend of domain adaptive loss based on domain adaptive loss over multiple time periods; obtaining the historical output stability of the model and on-board resource constraints; and comprehensively determining the online gating coefficients by integrating the regression confidence, the trend of domain adaptive loss, the historical output stability, and the on-board resource constraints in the uncertainty indicators.

[0010] Furthermore, the steps described above for performing online domain adaptive updates to the model based on online gating coefficients or for outputting tire three-part force estimates using a safe output strategy include: when the online gating coefficients exceed a preset threshold, performing online alignment updates of the source and target domains of the feature extraction network in the force estimation model with small steps to correct the tire three-part force estimates; when the online gating coefficients do not exceed the preset threshold, not performing online alignment updates, but using downgrading processing, or determining the tire three-part force estimates as the final estimates for output.

[0011] Furthermore, the above-mentioned steps for online alignment and updating of the source and target domains of the feature extraction network in the force estimation model with small step sizes include: performing one or more gradient descent operations based on domain adaptive loss with a preset low learning rate, and backpropagating to update the parameters in the feature extraction network. Furthermore, after the above steps of outputting wear indicators based on vehicle operating condition information, the method further includes: if the wear status label is inconsistent with the historical wear label status in the previous time window, determining whether the built-in confidence of the wear status label is higher than the threshold for a consecutive specified number of sliding windows, and if the confidence difference with the historical wear label exceeds a set value, triggering a state switch, and using the wear status label as the latest status label; the latest status label is used for the adaptive adjustment of subsequent domain alignment weights, kernel parameters and online update frequency, and serves as a reference input for safety constraints and degradation strategies; after the state switch is triggered, the model output and adaptive step size are subjected to exponential smoothing to suppress the output abrupt changes caused by the state switch.

[0012] Furthermore, the training process of the above force estimation model is as follows: A source domain labeled data training set is obtained; the source domain samples in the training set include three-channel input sequences and corresponding true values ​​of the three tire forces; the source domain samples are input into the feature extraction network and the regression network, and training is performed with the goal of minimizing the source domain regression loss. The source domain regression loss integrates SmoothL1, MSE, and Log-Cosh losses, and the weights of each loss term are adaptively balanced using a learnable variance parameter; a joint domain alignment module is trained using source domain samples and unlabeled samples from the target wear domain, and the domain adaptive loss is calculated; the source domain regression loss and the domain adaptive loss are adaptively balanced using entropy weight coefficients to obtain the total model loss, which is used to adjust the parameters of the feature extraction network to obtain the force estimation model.

[0013] Furthermore, the regression network also introduces physical constraint regularization: the vertical force is non-negative and consistent with the load variation; the amplitudes of the longitudinal and lateral forces are constrained by the adhesion coefficient and the friction ellipse; during training, a penalty term is added to samples that violate the constraints, or projection correction is performed during inference.

[0014] Secondly, this application also provides an intelligent tire force estimation device, comprising: a signal acquisition module for real-time acquisition of tire triaxial acceleration signals and vehicle operating condition information and time alignment; a signal processing module for sequentially performing grounding event localization, contact segment extraction, and linear interpolation resampling on the tire triaxial acceleration signals based on vehicle operating condition information to obtain a three-channel input sequence; a model prediction module for inputting the three-channel input sequence into an offline-trained force estimation model, outputting tire three-force estimates and corresponding uncertainty indices; and outputting wear indices based on vehicle operating condition information; a deviation evaluation module for determining the degree of deviation between the target domain and the source domain based on unlabeled samples and labeled samples in the target domain within the most recent specified time period, as well as the wear indices, and determining online gating coefficients in conjunction with the uncertainty indices; and a safe output module for performing online domain adaptive updates to the model based on the online gating coefficients or outputting tire three-force estimates using a safe output strategy.

[0015] Thirdly, this application also provides an intelligent tire force estimation system, the system comprising: a controller and vehicle sensors; the controller is used to perform the method described in the first aspect.

[0016] Fourthly, this application also provides a computer-readable storage medium storing computer-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the method described in the first aspect above.

[0017] The intelligent tire force estimation method, device, and system provided in this application utilize tire triaxial acceleration ground contact segments combined with operating condition information. Through an integrated framework of "ground contact event localization and fragmentation—multi-scale temporal feature encoding—multi-task force regression—mean and variance joint statistical alignment—online gating adaptive and safety constraint management," it achieves real-time and continuous estimation of tire longitudinal force Fx, lateral force Fy, and vertical force Fz. The dynamic statistical domain adaptive method uses Student kernel maximum mean difference and variance difference regularization to align the feature distributions of the source and target domains, and achieves an adaptive balance between regression and alignment through uncertainty weighting and entropy weight allocation. Simultaneously, wear embedding and domain state lag updates are introduced, along with confidence gating, physical consistency constraints, amplitude limiting, and degradation estimation, to achieve stable operation and cross-wear stage generalization under vehicle deployment conditions. It can achieve cross-wear evolution migration under weakly labeled or unlabeled target domain conditions, realizing online reliable reasoning and safety mechanism-constrained tire force estimation. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating an intelligent tire force estimation method provided in this application embodiment; Figure 2 A structural block diagram of an intelligent tire force estimation device provided in an embodiment of this application; Figure 3 This is a flowchart of another execution step corresponding to an intelligent tire force estimation system provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of this application will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] To address the issues of high cost and difficulty in continuous labeling of the three-part force labels in the target tire wear domain in existing technologies, there is an urgent need for a tire force estimation scheme that can achieve cross-wear evolution migration under conditions of weak or no labeling in the target domain, and has online reliable reasoning and safety constraint mechanisms.

[0022] Based on this, embodiments of this application provide an intelligent tire force estimation method, device, and system that can achieve cross-wear evolution migration under weakly labeled or unlabeled conditions in the target domain, and realize tire force estimation with online reliable reasoning and safety mechanism constraints.

[0023] To facilitate understanding of this embodiment, a detailed description of an intelligent tire force estimation method disclosed in this application embodiment will be provided first. See Figure 1 The flowchart shown in this application embodiment provides a smart tire force estimation method, which specifically includes the following steps: Step S101: Real-time acquisition of tire triaxial acceleration signals and vehicle operating condition information, and time alignment. A triaxial accelerometer is installed in the tire liner or at a suitable location on the tire structure to collect triaxial acceleration response signals during tire contact with the ground. The sensor employs a miniaturized package and a sampling frequency of 3 kHz to 5 kHz to cover the critical frequency bands for ground impact and structural vibration. The acquisition module includes a triaxial accelerometer, a wireless signal transmission module, a WiFi communication unit, and a power supply unit. The power supply line is led out through the valve stem and connected to an external power source on the wheel rim. The collected data is wirelessly transmitted via WiFi to a computing terminal or vehicle gateway for storage and analysis, thus avoiding the need for devices such as slip rings. In addition to tire acceleration, the system can integrate vehicle operating condition information such as wheel speed, vehicle speed, longitudinal and lateral acceleration, yaw rate, tire pressure, and tire temperature for operating condition normalization and domain adaptive assistance.

[0024] The acquired multi-source signals enter the fusion and preprocessing process: First, they are synchronized according to the timestamp, and signals with different sampling frequencies are aligned to a unified time axis; then, the acceleration signals are denoised, de-drifted and amplitude corrected, and bandpass filtering or wavelet denoising can be used to suppress low-frequency drift and wireless link jitter.

[0025] Step S102: Based on vehicle operating condition information, the tire triaxial acceleration signal is sequentially subjected to grounding event location, contact segment extraction, and linear interpolation resampling to obtain a three-channel input sequence; In specific implementation, the jump signal is obtained by taking the time derivative of the radial acceleration, and the transient peaks at the entry and exit from the ground are detected. Combined with the angular velocity in the vehicle operating condition information, the contact angle and the contact center point are calculated. Based on the contact center point, acceleration segments within a specified angle range before and after the contact center point are extracted from the tire triaxial acceleration signal as valid segments. Each valid segment is resampled to a specified number of time steps using linear interpolation to form a fixed-length three-channel input sequence. Alternatively, it may also include: the sequence after the three-channel input sequence is normalized by the maximum absolute value, which is used as the model input.

[0026] For example, in the tire contact process, the impact "jump" characteristic of radial acceleration is used to locate the start and end points of contact and estimate the contact angle: the jump signal is obtained by taking the time derivative of radial acceleration, the transient peaks at the entry and exit of contact are detected, and the contact angle and contact center point are calculated by combining the angular velocity.

[0027] Using the contact center as a reference, acceleration segments within a 40° range before and after the contact point are extracted as valid input. To ensure dimensionality consistency, each segment is resampled to 400 time steps using linear interpolation, forming a fixed-length three-channel input sequence X. To improve training stability and cross-domain generalization ability, maximum absolute value normalization is performed on each segment. x_norm(i) = x(i) / max(|x(1)|,..., |x(i)| ,..., |x(L)|); i=1,2,...,L; Where x(i) represents the three-channel input sequence of the i-th effective acceleration segment; |x(i)| represents the absolute value of the three-channel input sequence of the i-th effective acceleration segment, i.e., taking the absolute value of each value in the three-channel input sequence; x_norm(i) represents the vector after normalizing the i-th effective acceleration segment by the maximum absolute value, L = 400, normalization compresses the signal amplitude to [-1,1] while preserving shape information. Based on the above processing, sample sets of the source domain (new tire or low-wear tire) and the target domain (high-wear tire) are constructed to provide a unified input for subsequent regression and domain adaptation.

[0028] Step S103: Input the three-channel input sequence into the offline trained force estimation model, output the tire three-part force estimate and the corresponding uncertainty index; and output the wear index based on the vehicle working condition information. The force estimation model described above includes: a regression network of a feature extraction network; the feature extraction network includes: a cascaded structure of a temporal convolutional network and a bidirectional long short-term memory network that integrates a multi-head self-attention mechanism; the model prediction process is as follows: through the cascaded structure of the temporal convolutional network and the bidirectional long short-term memory network, features are extracted from the three-channel input sequence to obtain a temporal feature sequence; through the multi-head self-attention mechanism, residual connections, random deactivation, and normalization are performed on the temporal feature sequence to obtain enhanced temporal features; through the regression network, regression processing is performed on the enhanced temporal features to obtain the tire three-part force estimates.

[0029] In practice, a multi-scale temporal feature extraction network is constructed using a preprocessed fixed-length triaxial acceleration sequence as input to encode the dynamic characterization of tire grounding. The feature extraction network employs a cascaded structure of a Temporal Convolutional Network (TCN) and a Bidirectional Long-Short Term Memory (BiLSTM): the TCN is used to extract multi-scale convolutional features of local impacts and periodic structures, while the BiLSTM is used to model the dependencies between different grounding events. To highlight the contributions of key moments such as contact start and end, and peak impact, a lightweight multi-head self-attention (MHSA) system is introduced. Head Self The Attention mechanism improves convergence stability through residual connections and layer normalization. = LayerNorm(X + Dropout(A)) In the above formula, X represents the temporal feature sequence encoded by TCN-BiLSTM and input to the multi-head self-attention layer, and its dimension can be represented as T×d, where T is the number of time steps after resampling, and d is the feature dimension of each time step; A represents the attention weighted output calculated by the multi-head self-attention based on the query, key, and value, and has the same dimension as X; Dropout(A) represents random deactivation of the attention output to suppress overfitting; LayerNorm(·) represents the layer normalization operation, used to stabilize the feature scale under different wear states and different working conditions; This represents the enhanced temporal features after residual connection, random deactivation, and normalization, and serves as the shared input for subsequent multi-task regression heads and domain alignment modules.

[0030] In force regression, a source domain regression head is designed to perform multi-task regression outputs on Fx, Fy, and Fz, and a multi-loss adaptive fusion mechanism is introduced to adapt to different error scales. The source domain regression loss L_reg(θ) is defined as: L_reg(θ) = Σ i=1 ³ (L i (θ) / (2σ i ²) + log σ i ) Where L1, L2, and L3 are the Smooth L1 loss, Mean Squared Error (MSE) loss, and Log-Cosh loss, respectively; σ i ² = exp(2 log σ i ) represents the learnable variance parameter; θ represents the trainable parameters of the network. By learning logσ...i The model can automatically adjust the contribution weights of each error metric during training, enhancing the regression head's ability to comprehensively fit peak impacts, steady-state segments, and anomalous noise segments. Furthermore, the regression head can output prediction uncertainty or confidence indices for online gating and safety constraints.

[0031] Furthermore, the aforementioned wear indicators include wear embedding vectors and / or wear state labels. The steps for outputting wear indicators based on vehicle operating condition information include: summarizing vehicle operating condition information over a specified time scale to determine operating condition statistical characteristics; these characteristics include: the mean and variance of the contact angle over time, the radial impact peak distribution, the spectral energy ratio, and the rate of change of peak-to-peak value and envelope within a period; updating the operating condition statistical characteristics online using a sliding window based on the vehicle operating condition information and mapping them to wear embedding vectors; and determining wear state labels based on the wear embedding vectors. Wear state labels include light wear, moderate wear, and heavy wear.

[0032] In practical implementation, this application further summarizes the statistical characteristics of operating conditions and wear characteristics over a longer time scale to improve cross-condition robustness. Statistical characteristics may include: the mean and variance of the contact angle over time, the radial impact peak distribution, the spectral energy ratio, and the rate of change of peak-to-peak value and envelope within a period. The system can update these characteristics online through a sliding window and map them to a wear embedding vector w, which is used as an auxiliary input for domain alignment and regression head conditionalization, enabling the network to distinguish structural differences caused by wear under similar operating conditions.

[0033] Through the feature extraction and regression modeling described above, the model in this embodiment outputs the tire three-force estimates [Fx, Fy, Fz], wear embedding w, and corresponding confidence level c and uncertainty u in each inference cycle. The confidence level characterizes whether the current input falls near the training distribution, and the uncertainty can be used to determine whether online domain adaptation or degraded estimation is needed. This output can directly serve applications such as vehicle longitudinal / lateral control, tire load monitoring, adhesion coefficient identification, and wear life management.

[0034] To obtain online credibility metrics, this embodiment employs two methods: first, simultaneously outputting variance or confidence intervals using the regression head; second, introducing random deactivation or lightweight ensemble inference, repeating the inference multiple times (preferably 10 times) on the same input, statistically analyzing the output variance and mapping it to uncertainty u, with smaller variance indicating higher credibility. Simultaneously, information entropy can be calculated from the probability distribution of the domain discrimination branch output as an uncertainty metric; smaller entropy indicates more reliable domain judgment. The aforementioned credibility metrics are used for online gating and security strategies.

[0035] Step S104: Based on the unlabeled samples of the target domain and the labeled samples of the source domain within the most recent specified time period, as well as the wear index, determine the degree of deviation between the target domain and the source domain, and determine the online gating coefficient in combination with the uncertainty index; Obtain unlabeled samples from the target domain and labeled samples from the source domain within the most recent specified time period; output the shared feature vectors corresponding to the labeled samples from the source domain and the unlabeled samples from the target domain through a feature extraction network; select or update the student kernel parameters based on wear indicators and operating condition statistical features; calculate the MMD (Maximum Mean Discrepancy) and VDR (Variance-Difference Regularization) between the source domain and the target domain based on the updated student kernel parameters and the shared feature vectors; and obtain the domain adaptive loss L_DA by weighted summing of MMD and VDR through uncertainty weighting coefficients to determine the degree of deviation.

[0036] Based on the domain adaptive loss over multiple time periods, the trend of domain adaptive loss change is determined; the historical output stability of the model and on-board resource constraints are obtained; and the regression confidence, the trend of domain adaptive loss change, the historical output stability, and the on-board resource constraints in the uncertainty indicators are integrated to comprehensively determine the online gating coefficient.

[0037] Step S105: Based on the online gating coefficients, perform online domain adaptive updates on the model or use a safe output strategy to output the tire three-part force estimates.

[0038] When the online gating coefficient exceeds the preset threshold, the feature extraction network in the force estimation model is updated online with a small step size to align the source and target domains, so as to correct the tire three-part force estimate. When the online gating coefficient does not exceed the preset threshold, no online alignment update is performed, and downgrading is adopted, or the tire three-part force estimate is determined as the final estimate and output.

[0039] The step of updating the source and target domains of the feature extraction network in the force estimation model with small step sizes includes: performing one or more gradient descents based on domain adaptive loss with a preset low learning rate, and backpropagating to update the parameters in the feature extraction network. In other words, in this embodiment, the regression confidence level c, the trend of domain adaptive loss, historical output stability, and resource constraint information are fused into an online gating coefficient g, which is used to determine whether to perform online domain adaptive updates and the update intensity. When g is below a threshold (preferably 0.6), online updates are turned off, only regression results are output, and stronger smoothing and physical constraints are enabled; when g is above the threshold, alignment updates with small steps are allowed, but this is limited by onboard computing power and power consumption budget.

[0040] In a preferred embodiment, after the step of outputting wear indicators based on vehicle operating condition information, the method further includes: If the wear status label is inconsistent with the historical wear label status in the previous time window, the system determines whether the built-in confidence of the wear status label is higher than the threshold for a specified number of consecutive sliding windows, and whether the difference in confidence between the wear status label and the historical wear label exceeds a set value. In this case, a state switch is triggered, and the wear status label is used as the latest status label. The latest status label is used for the adaptive adjustment of subsequent domain alignment weights, kernel parameters, and online update frequency, and serves as a reference input for safety constraints and degradation strategies. After the state switch is triggered, the model output and adaptive step size are subjected to exponential smoothing to suppress the output abrupt changes caused by the state switch.

[0041] The above steps are actually a hysteresis and debouncing mechanism used in wear embedding and domain state updates to avoid online adaptive instability caused by frequent jumps in wear level or domain label. Specifically, a state switch is triggered only when the confidence level of a new domain state or wear level is higher than a threshold (preferably 0.7) for multiple consecutive windows and the confidence difference between it and the current state exceeds a set value. After the switch, the output and adaptive step size are exponentially smoothed to ensure continuous estimation results. The updated wear state results are used for subsequent adaptive adjustment of domain alignment weights, kernel parameters, and online update frequency, and serve as reference input for safety constraints and degradation strategies.

[0042] The process of step S104 above will be explained in detail below: To address the drift in acceleration feature distribution between the source and target domains caused by tire wear, this embodiment employs joint distribution alignment processing to simultaneously constrain the encoded features at both the mean and variance levels, thereby achieving knowledge transfer even without labeled force data in the target domain. This process includes: maximum mean difference alignment based on the student kernel, variance difference regularization alignment, and an uncertainty weighting mechanism for dynamically balancing the contributions of different loss terms.

[0043] In this application scenario, the joint domain alignment process is not an isolated mathematical calculation, but is performed after the aforementioned regression network has output shared features, tire force estimates, and wear embeddings. The specific execution steps are as follows: First, the shared features corresponding to labeled samples in the source domain and unlabeled samples in the target wear domain are received; second, the student kernel parameters are selected or updated based on the wear embeddings and operating condition statistics; third, the MMD and VDR of the source and target domains are calculated respectively; subsequently, the domain adaptive loss is allocated based on uncertainty weighting and entropy weighting, and this loss is fed back to the shared feature extraction network; finally, the network parameters are updated during the training phase, and small-step alignment updates are performed only when gating is allowed during the online phase. The purpose of this joint distribution alignment is to maintain statistical consistency of the aforementioned tire force regression features across different wear stages, thereby reducing estimation bias caused by tire wear, tire pressure and temperature changes, and differences in road surface excitation.

[0044] Regarding mean alignment, the maximum mean difference (MMD) is used to measure the mean embedding difference (MMD²(P,Q) between the source domain distribution P and the target domain distribution Q in the RKHS (Reproducing Kernel Hilbert Space):

[0045] Where k(·,·) is the kernel function. To enhance robustness to heavy-tailed distributions of mechanical signals, outliers, and abrupt changes in wear stages, a Student t-distribution kernel is introduced instead of the Gaussian kernel. Its kernel function form, k_student(x,y), can be expressed as: k_student(x,y) = (1 + ||xy||² / (α·σ²_mmd))^(-α) Where α is the degree of freedom parameter of the Student t-distribution, the default value is α = 3, and the value range is α ∈ [1, 10]; σ_mmd is the kernel bandwidth parameter, which is a hyperparameter and is initialized using the median heuristic. The Student kernel decays more slowly in regions with larger distances, which can reduce the excessive dominance of the alignment gradient by a small number of outliers.

[0046] Regarding variance alignment, VDR is introduced to match the variance structure VDR(P,Q) of the source and target domains in the RKHS: VDR(P,Q) = ||Var_{x~P}(k(x,·)) - Var_{y~Q}(k(y,·))|| Its variance can be calculated using the centered kernel matrix: VDR(P,Q)=|1 / (n_P-1)·Tr(H_P·K_P·H_P) - 1 / (n_Q-1)·Tr(H_Q·K_Q·H_Q)| In the above formula, P represents the feature distribution of the source domain, and Q represents the feature distribution of the target wear domain; x and y represent the feature vectors sampled from P and Q, respectively; k(x,·) and k(y,·) represent the kernel feature functions after the samples are mapped by the student kernel; Var(·) represents the variance of the kernel features on the corresponding sample set; n_P and n_Q are the number of source domain samples and the number of target domain samples participating in the current batch or memory calculation, respectively; K_P and K_Q are the kernel matrices of the source domain and the target domain samples, respectively; H_P and H_Q are the centering matrices; I is the identity matrix; 1 is a column vector of all 1s; Tr(·) is the matrix trace operation; the smaller VDR(P,Q), the closer the source domain and the target domain are in terms of feature dispersion and fluctuation structure.

[0047] Where H_P = I - (1 / n_P)·1·1 Let be the centered matrix, and Tr(·) be the trace operation. MMD aligns the distribution center, and VDR aligns the discrete structure; their complementarity significantly improves the alignment effect across wear states. To dynamically balance the two types of alignment terms, an uncertainty weighting is introduced: λ1= 1 / (2·exp(2 log σ_mmd)), λ2= 1 / (2·exp(2 log σ_vdr)) Where log σ_mmd and log σ_vdr are learnable parameters. The domain adaptive loss can be defined as L_DA = λ1·MMD² + λ2·VDR. Furthermore, to adaptively balance "task regression" and "domain alignment" during training, entropy weights λ_reg are used for allocation: L_total = λ_reg·L_reg + (1-λ_reg)·L_DA λ_reg = exp(-τ·L_DA) / (exp(-τ·L_DA) + exp(-τ·L_reg)) Where τ is the temperature parameter, a dimensionless positive number that controls the sharpness of the softmax weight distribution, typically ranging from [0.1, 10], with a default value of τ = 1. L_total represents the total loss function for network training. This strategy prioritizes aligning the distribution in the early stages of training and gradually enhances regression accuracy in the later stages.

[0048] The online inference phase integrates regression output, unsupervised alignment incremental updates, and safety constraints, and is subject to gating mechanisms. The system receives tire acceleration and vehicle state signals at fixed intervals, performs grounding event localization, fragment extraction, and normalization, and then inputs the results into the regression network to output tire force. When the gating coefficient g meets the threshold and computational resources allow, the system performs small-step updates to the alignment module based on recent target domain samples, achieving online migration across wear stages. When the gating coefficient does not meet the conditions, only the regression results are output, and amplitude limiting, smoothing, and degradation strategies are implemented.

[0049] To ensure that tire force estimation possesses both generalization capability across wear states and meets on-vehicle real-time requirements, this embodiment employs a two-stage mechanism of "offline pre-training + online fine-tuning" at the deployment level. In the offline stage, the regression network and alignment module are initially trained using source domain labeled data. In the online stage, without introducing target domain force labels, the recent target domain feature distribution is corrected with small steps using MMD+VDR alignment loss. The system can maintain a target domain feature memory or prototype vector, forming statistical references for different wear stages, thereby achieving continuous adaptation rather than discrete jumps as wear gradually evolves.

[0050] The online execution process includes: acquiring the tire triaxial acceleration sequence and auxiliary working condition information; estimating the contact angle and extracting contact segments; resampling and normalizing to form a fixed-length input; outputting the tire three-dimensional forces and uncertainties through a feature extractor and regression head; calculating the L_DA (Domain-Adaptation Loss) between recent target domain samples and source domain references; calculating λ_reg and updating the total loss; performing alignment updates and refreshing the target domain memory under gating conditions; and finally, performing exponential smoothing and physical consistency correction on the output three-dimensional forces and publishing them to the chassis controller or data recording system.

[0051] The relationship between the above steps is as follows: Step S102 provides a three-channel input sequence that has been synchronized, truncated, and normalized. Step S103 outputs the tire force components, confidence level, and uncertainty based on this sequence. Step S104 calculates the statistical offset using recent target domain features and source domain reference features and provides the basis for whether adaptive alignment is needed. Step S105 determines whether to perform online updates and how to publish safe outputs based on confidence level, alignment loss changes, physical constraints, and vehicle resource status. The updated or frozen regression output is input into the safety constraint management process to complete amplitude limiting, smoothing, physical consistency correction, and degraded publication. Therefore, the fourth module is the vehicle-mounted closed-loop management link for the aforementioned signal acquisition, force regression, and domain alignment results. Its function is to prevent domain adaptive updates from directly disturbing the vehicle control input and to ensure the continuity and safety boundary of tire force estimation under abnormal operating conditions.

[0052] Step S105 does not redefine the tire force, but uses the features obtained in step S102 and the force confidence in step S103 to determine whether the estimation model has been affected by wear domain drift in step S104. Step S105 follows the judgment result of step S104, publishes the corrected estimate when adaptation is allowed, and stops updating and adopts a safe output strategy when adaptation is not allowed or anomalies are detected.

[0053] In terms of parameterized wear adaptation, the system can schedule the online update step size of the student t-distribution degree-of-freedom parameter α based on the wear embedding vector w or the wear level: when the wear is higher and the domain difference is greater, the alignment weight is increased or α is decreased to improve the tolerance for heavy-tailed distributions; after the domain difference converges, λ_reg is gradually increased and the online update frequency is reduced. The system can also establish domain prototype sets for different wear stages and use nearest prototype matching and interpolation to achieve smooth migration under continuous wear conditions.

[0054] To further enhance the engineering value, this embodiment extends the wear life management function on the basis of tire force estimation: it uses the wear embedding vector w over a long time scale and force statistics (such as the mean of Fz and the peak impact distribution) to construct wear evolution indicators, and can use a lightweight time series prediction model to predict the wear trend and force estimation deviation in the future, thereby providing forward-looking information for tire maintenance, calibration strategies and chassis control margin.

[0055] Online adaptive updates employ stabilization strategies to avoid abrupt changes in the control system: gradients are pruned and the magnitude of single parameter variations is limited; an exponential moving average is used to form a teacher network for the updated network parameters, which can generate high-confidence pseudo-labels in some weakly labeled scenarios to enhance regression consistency. Furthermore, self-supervised contrastive pre-training (such as segment contrast and frequency domain consistency enhancement) can be performed on the unlabeled target domain, first learning domain-invariant representations and then fine-tuning alignment, thereby improving convergence speed and robustness under extreme wear or sparse sample conditions.

[0056] In addition, to ensure the availability and security of the estimation results, this embodiment constructs a physical constraint and operational boundary mechanism: under different vehicle speeds and adhesion conditions, upper limits for amplitude and rate of change are set for Fx, Fy, and Fz; larger amplitude fluctuations are allowed in low-speed conditions to preserve details, while constraints are tightened in high-speed and high lateral acceleration conditions to suppress noise amplification; projection correction or amplitude limiting is performed on outputs that obviously violate the friction ellipse, have negative Fz, or abruptly exceed the limit.

[0057] Stability protection employs an "estimate consistency" boundary criterion: the system can run a backup estimate based on vehicle sensors and a simplified tire model / observer in parallel, monitoring the difference and trend between the intelligent tire estimate and the backup estimate in real time. When the difference exceeds a threshold related to vehicle speed and load, or when the domain discrimination entropy increases significantly, it is determined to be an out-of-distribution input or sensor anomaly, triggering a degradation strategy. This mechanism can suppress the spread of misestimation under conditions such as wet skies, strong impacts, and sensor loosening.

[0058] The intervention intensity reduction adopts continuous decay and safety mode switching: when the safety margin decreases, the online alignment update intensity and output update rate continuously decay according to the margin, while tightening the output limit; when the safety margin is lower than the critical value, it switches to safety mode, stops online updates and uses backup estimates or the most recent reliable output to prioritize meeting the stability requirements of vehicle control and safety assessment.

[0059] The system incorporates sensor anomaly degradation logic. When the wireless link is interrupted or tire sensor data is unavailable, the intelligent tire signal is marked as missing, and tire force estimation is downgraded to a vehicle-side sensor + model observer approach. When acceleration signal noise increases significantly or drifts, the smoothing intensity is increased and the online update frequency is reduced; if necessary, the wear embedding vector w is frozen to its most recent stable value. Physical constraints and output limiting are still applied during degradation to ensure safety boundaries are met even with incomplete information.

[0060] Through the aforementioned offline training, joint domain alignment, online trusted reasoning, and security constraint fusion mechanism, this embodiment ultimately generates a continuous, smooth, and robust tire three-part force estimation output across wear states. This enables the intelligent tire system to maintain stable and deployable force estimation performance under different wear levels, loads, and speed variations, providing reliable input for chassis control and tire health management.

[0061] Furthermore, the training process of the above force estimation model is as follows: A source domain labeled data training set is obtained; the source domain samples in the training set include three-channel input sequences and corresponding true values ​​of the three tire forces; the source domain samples are input into the feature extraction network and the regression network, and training is performed with the goal of minimizing the source domain regression loss. The source domain regression loss integrates SmoothL1, MSE, and Log-Cosh losses, and the weights of each loss term are adaptively balanced using a learnable variance parameter; a joint domain alignment module is trained using source domain samples and unlabeled samples from the target wear domain, and the domain adaptive loss is calculated; the source domain regression loss and the domain adaptive loss are adaptively balanced using entropy weight coefficients to obtain the total model loss, which is used to adjust the parameters of the feature extraction network to obtain the force estimation model.

[0062] Furthermore, to improve engineering interpretability and physical consistency, the regression network described above also incorporates physical constraint regularization: the vertical force is non-negative and consistent with the load variation; the amplitudes of the longitudinal and lateral forces are subject to... The adhesion coefficient is constrained by the friction ellipse (|Fx| ≤ μFz, |Fy| ≤ μFz, or Fx² + Fy² ≤ (μFz)²), where μ is the tire-road adhesion coefficient (peak friction coefficient), which is dimensionless and represents the maximum available friction level under the current road and tire conditions. During training, a penalty term is added to samples that violate the constraints, or a projection correction is performed during inference, thereby reducing the risk of non-physical output in cross-domain cases.

[0063] The beneficial effects of this embodiment are as follows: 1. Achieve high-precision estimation of tire three-part forces in response to wear state changes. This embodiment utilizes temporal modeling and attention-enhanced representation of triaxial acceleration ground segments, and achieves cross-wear domain transfer through joint alignment of student kernel MMD and VDR. This allows the model to maintain high estimation accuracy and stability without retraining for each wear level, significantly mitigating performance degradation caused by wear-induced distribution drift.

[0064] 2. Reduce calibration and data acquisition costs, and achieve adaptive calibration for weakly labeled / unlabeled tires. This embodiment utilizes uncertainty weighting, multi-loss regression, and entropy weight allocation mechanisms to enable alignment learning and online fine-tuning even when the target wear domain lacks three-component force labels. This reduces reliance on six-component force rim and bench tests, lowers engineering calibration workload, and improves transferability across vehicle models and tire batches.

[0065] 3. Improve the feasibility and operational safety of vehicle deployment. This embodiment adopts a low-intrusion wireless data acquisition and valve power supply scheme, avoiding complex devices such as slip rings, which facilitates vehicle installation and maintenance; at the same time, through confidence gating, physical constraints, output limiting and degradation estimation mechanisms, it can still ensure output continuity and safety boundaries under distributed external input or sensor abnormal conditions, meeting the engineering application requirements of chassis control and safety assessment.

[0066] Based on the above method embodiments, this application also provides an intelligent tire force estimation device, see [link to relevant documentation]. Figure 2As shown, the device includes: a signal acquisition module 21, used to acquire tire triaxial acceleration signals and vehicle operating condition information in real time and perform time alignment; a signal processing module 22, used to perform grounding event localization, contact segment extraction and linear interpolation resampling on the tire triaxial acceleration signals based on vehicle operating condition information to obtain a three-channel input sequence; a model prediction module 23, used to input the three-channel input sequence into an offline trained force estimation model, output tire three-force estimates and corresponding uncertainty indices; and output wear indices based on vehicle operating condition information; a deviation evaluation module 24, used to determine the degree of deviation between the target domain and the source domain based on unlabeled samples and labeled samples in the target domain within the most recent specified time period, as well as wear indices, and determine online gating coefficients in combination with uncertainty indices; and a safe output module 25, used to perform online domain adaptive updates of the model based on online gating coefficients or to output tire three-force estimates using a safe output strategy.

[0067] Furthermore, the aforementioned signal processing module 22 is used to obtain the jump signal by taking the time derivative of the radial acceleration, detect the transient peaks at the entry and exit points from the ground, and calculate the contact angle and contact center point by combining the angular velocity in the vehicle operating condition information; using the contact center point as a reference, extract acceleration segments within a specified angle range before and after the contact center point from the tire triaxial acceleration signal as valid segments; resample each valid segment to a specified number of time steps using linear interpolation to form a fixed-length three-channel input sequence; or it may also include: using the sequence after normalizing the maximum absolute value of the three-channel input sequence as the model input.

[0068] Furthermore, the aforementioned force estimation model includes: a regression network of the feature extraction network; the feature extraction network includes: a cascaded structure of a temporal convolutional network and a bidirectional long short-term memory network that integrates a multi-head self-attention mechanism; a model prediction module 23, used to extract features from the three-channel input sequence through the cascaded structure of the temporal convolutional network and the bidirectional long short-term memory network to obtain a temporal feature sequence; to perform residual connection, random deactivation and normalization processing on the temporal feature sequence through a multi-head self-attention mechanism to obtain enhanced temporal features; and to perform regression processing on the enhanced temporal features through a regression network to obtain the tire three-part force estimate.

[0069] Furthermore, the aforementioned wear indicators include wear embedding vectors and / or wear state labels; the model prediction module 23 is used to summarize vehicle operating condition information at a specified time scale and determine operating condition statistical features; the operating condition statistical features include: the mean and variance of the contact angle over time, the radial impact peak distribution, the spectral energy ratio, and the peak-to-peak value and envelope change rate within the period; based on the vehicle operating condition information, the operating condition statistical features are updated online through a sliding window and mapped to wear embedding vectors; based on the wear embedding vectors, wear state labels are determined.

[0070] Furthermore, the aforementioned deviation assessment module 24 is used to obtain unlabeled samples of the target domain and labeled samples of the source domain within the most recent specified time period; output the shared feature vectors corresponding to the labeled samples of the source domain and the unlabeled samples of the target domain through the feature extraction network; select or update the student kernel parameters according to the wear index and working condition statistical features; calculate the MMD and VDR between the source domain and the target domain based on the updated student kernel parameters and the shared feature vectors; and obtain the domain adaptive loss by weighting the MMD and VDR through the uncertainty weighting coefficient to determine the degree of deviation.

[0071] Furthermore, the aforementioned deviation evaluation module 24 is used to determine the trend of domain adaptive loss based on domain adaptive loss over multiple time periods; obtain the historical output stability of the model and vehicle resource constraints; and integrate the regression confidence, the trend of domain adaptive loss, the historical output stability, and vehicle resource constraints in the uncertainty indicators to comprehensively determine the online gating coefficient.

[0072] Furthermore, the aforementioned safety output module 25 is used to perform online alignment updates of the source and target domains of the feature extraction network in the force estimation model with small steps when the online gating coefficient exceeds a preset threshold, so as to correct and output the tire three-part force estimate; when the online gating coefficient does not exceed the preset threshold, no online alignment update is performed, and downgrading is adopted, or the tire three-part force estimate is determined as the final estimate and output.

[0073] Furthermore, the aforementioned secure output module 25 is used to perform one or more gradient descent operations based on domain adaptive loss at a preset low learning rate, and to backpropagate to update the parameters in the feature extraction network. Furthermore, the aforementioned safety output module 25 is used to, after the step of outputting wear indicators based on vehicle operating condition information, determine whether the confidence level of the wear status label is higher than a threshold for a consecutive specified number of sliding windows and whether the confidence level difference with the historical wear label exceeds a set value, if the wear status label is inconsistent with the historical wear label status in the previous time window. Then, it triggers a state switch and uses the wear status label as the latest status label. The latest status label is used for the adaptive adjustment of subsequent domain alignment weights, kernel parameters, and online update frequency, and serves as a reference input for safety constraints and degradation strategies. After the state switch is triggered, exponential smoothing is applied to the model output and adaptive step size to suppress output abrupt changes caused by the state switch.

[0074] Furthermore, the aforementioned device also includes: a model training module, used to perform the following training process for the force estimation model: acquiring a source domain labeled data training set; the source domain samples in the training set include a three-channel input sequence and the corresponding true values ​​of the three tire forces; inputting the source domain samples into the feature extraction network and the regression network, training with the goal of minimizing the source domain regression loss, the source domain regression loss integrating Smooth L1, MSE, and Log-Cosh losses, and adaptively balancing the weights of each loss term through a learnable variance parameter; training a joint domain alignment module using source domain samples and unlabeled samples from the target wear domain, and calculating the domain adaptive loss; adaptively balancing the source domain regression loss and the domain adaptive loss through entropy weight coefficients to obtain the total model loss, thereby adjusting the parameters of the feature extraction network to obtain the force estimation model.

[0075] Furthermore, the regression network also introduces physical constraint regularization: the vertical force is non-negative and consistent with the load variation; the amplitudes of the longitudinal and lateral forces are constrained by the adhesion coefficient and the friction ellipse; during training, a penalty term is added to samples that violate the constraints, or projection correction is performed during inference.

[0076] The device provided in this application embodiment has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts of the device embodiment not mentioned can be referred to the corresponding content in the aforementioned method embodiment.

[0077] Based on the above method embodiments, this application also provides an intelligent tire force estimation system, which includes: a controller and vehicle sensors; the controller is used to execute the method described in the foregoing method embodiments. See also Figure 3 As shown, the controller is used to perform the following steps: The signal acquisition and fusion step first converts the tire triaxial acceleration and vehicle operating condition information into a unified three-channel input sequence. The time-series feature extraction and tire force regression step extracts ground dynamic features from this input sequence and outputs estimated values ​​of Fx, Fy, and Fz. The joint domain alignment and dynamic weight allocation step does not replace the force regression results, but uses shared features in the regression network and wear domain sample statistics to judge and correct cross-wear distribution offsets, so that the output of the second step remains reliable under the target wear domain. The online adaptive inference and safety management step then determines whether to fine-tune online, whether to freeze the model, and how to limit, smooth, and degrade the tire force estimation based on the domain offset degree obtained in the third step, the confidence level of the output of the second step, and the physical safety boundary.

[0078] Intelligent tire signal acquisition and fusion: A triaxial accelerometer is installed in the central area of ​​the tire liner and a wireless acquisition link is established. The sensor's range, size, and sampling frequency meet the requirements for ground impact acquisition. The power supply line is led out from the valve stem and connected to an external power supply on the wheel rim. Acceleration data is transmitted to the terminal via WiFi. Auxiliary signals such as vehicle speed, wheel speed, IMU, tire pressure / temperature, etc. are acquired synchronously and time-aligned. The contact start and end points are located based on radial jump and the contact angle is estimated. A 40° segment is extracted before and after the contact center. The segment is linearly interpolated and resampled to 400 steps and normalized to the maximum absolute value to form a fixed-length three-channel input sequence.

[0079] Temporal feature extraction and tire force regression: A fixed-length acceleration sequence is input into the TCN-BiLSTM-MHSA feature extraction network, and stable features are obtained using residual and layer normalization structures; the regression head uses shared encoding and multi-task output to simultaneously estimate Fx, Fy, and Fz, and is trained using Smooth L1, MSE, and Log-Cosh multi-loss adaptive fusion; optional physical constraint regularization such as friction ellipse and non-negativity of Fz is introduced, or the output is limited and projected during inference, while an uncertainty index is output for subsequent gating.

[0080] Joint domain alignment and dynamic weight allocation: Using labeled samples from the source domain (new tires / low wear) and unlabeled samples from the target domain (high wear) as input, the student kernel MMD is calculated in the feature space to achieve mean alignment, and the VDR is calculated to achieve variance structure alignment. Uncertainty weighted dynamic balancing of MMD and VDR contributions is adopted to obtain the domain adaptive loss L_DA. Furthermore, entropy weight allocation is used to balance L_reg and L_DA, so that the training phase adaptively switches between "alignment priority" and "regression priority" to improve the generalization performance across wear domains.

[0081] Online adaptive inference and safety management: When deployed in vehicles, tire force estimates are output with a 100 ms cycle, and the confidence gating coefficient g determines whether to perform online alignment fine-tuning. When g meets the threshold and resources allow, a small amount of alignment updates are performed using recent target domain samples, and the target domain memory is refreshed. When g is low or anomalies are detected, online updates are stopped, and stronger smoothing, physical constraints, output limiting, and backup estimation degradation are enabled. Optionally, self-supervised pre-training and teacher-student stabilization updates are performed on the unlabeled target domain to improve convergence speed and robustness under extreme wear or sparse sample conditions, ensuring that the system continuously outputs stable and reliable tire force estimates under different wear levels and complex operating conditions.

[0082] The system provided in this application embodiment has the same implementation principle and technical effects as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the system embodiment can be referred to the corresponding content in the aforementioned method embodiment.

[0083] This application also provides a computer-readable storage medium storing computer-executable instructions. When the computer-executable instructions are called and executed by a processor, the computer-executable instructions cause the processor to implement the above-described method. For specific implementation details, please refer to the foregoing method embodiments, which will not be repeated here.

[0084] The computer program products of the methods, apparatus, and electronic devices provided in the embodiments of this application include a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementations, please refer to the method embodiments, which will not be repeated here.

[0085] Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application.

[0086] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0087] In the description of this application, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0088] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the technical scope disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for estimating the force of an intelligent tire, characterized in that, The method includes: Real-time acquisition of tire triaxial acceleration signals and vehicle operating condition information, and time alignment; Based on the vehicle operating condition information, the tire triaxial acceleration signal is sequentially subjected to grounding event localization, contact segment extraction and linear interpolation resampling to obtain a three-channel input sequence. The three-channel input sequence is input into the offline-trained force estimation model, which outputs the tire three-part force estimates and corresponding uncertainty indices; and based on the vehicle operating condition information, the wear index is output. Based on the unlabeled samples of the target domain and the labeled samples of the source domain within the most recent specified time period, as well as the wear index, the degree of deviation between the target domain and the source domain is determined, and the online gating coefficient is determined in combination with the uncertainty index. Based on the online gating coefficients, the model is updated in an online domain or a safe output strategy is used to output the tire force estimates.

2. The method according to claim 1, characterized in that, Based on the vehicle operating condition information, the steps of sequentially performing grounding event location, contact segment extraction, and linear interpolation resampling on the tire triaxial acceleration signal to obtain a three-channel input sequence include: The jump signal is obtained by taking the time derivative of the radial acceleration, and the transient peaks at the entry and exit of the ground are detected. Combined with the angular velocity in the vehicle operating condition information, the contact angle and contact center point are calculated. Using the contact center point as a reference, acceleration segments within a specified angular range before and after the contact center point are extracted from the tire triaxial acceleration signal as valid segments. For each of the effective segments, linear interpolation is used to resample to a specified number of time steps to form a fixed-length three-channel input sequence; Alternatively, it may include: using the sequence obtained by normalizing the three-channel input sequence to the maximum absolute value as the model input.

3. The method according to claim 1, characterized in that, The force estimation model includes: a regression network of a feature extraction network; the feature extraction network includes: a cascaded structure of a temporal convolutional network and a bidirectional long short-term memory network that integrates a multi-head self-attention mechanism; the step of inputting the three-channel input sequence into the offline-trained force estimation model and outputting the estimated values ​​of the three forces of the tire includes: The temporal feature sequence is obtained by extracting features from the three-channel input sequence through the cascaded structure of the temporal convolutional network and the bidirectional long short-term memory network. The time-series feature sequence is subjected to residual connection, random deactivation and normalization through the multi-head self-attention mechanism to obtain enhanced time-series features. The enhanced temporal features are regressed using the regression network to obtain the tire force estimates.

4. The method according to claim 1, characterized in that, The wear index includes a wear embedding vector and / or a wear status label; The steps for outputting wear indicators based on the vehicle operating condition information include: Vehicle operating condition information at a specified time scale is summarized to determine the statistical characteristics of the operating conditions; the statistical characteristics of the operating conditions include: the mean and variance of the contact angle over time, the radial impact peak distribution, the spectral energy ratio, and the peak-to-peak value and envelope change rate within the period. Based on the vehicle operating condition information, the operating condition statistical features are updated online through a sliding window and mapped to a wear embedding vector; Based on the wear embedding vector, the wear state label is determined.

5. The method according to claim 3, characterized in that, The step of determining the degree of deviation between the target domain and the source domain based on unlabeled samples of the target domain and labeled samples of the source domain within the most recent specified time period, as well as the wear index, includes: Retrieve unlabeled samples of the target domain and labeled samples of the source domain within the most recent specified time period; The feature extraction network outputs a shared feature vector corresponding to labeled samples in the source domain and unlabeled samples in the target domain. Select or update student kernel parameters based on the wear index and operating condition statistical characteristics; Based on the updated student kernel parameters and the shared feature vector, the MMD and VDR between the source domain and the target domain are calculated respectively; the MMD and VDR are weighted and summed using uncertainty weighting coefficients to obtain the domain adaptive loss, thereby determining the degree of deviation.

6. The method according to claim 5, characterized in that, The steps of determining the degree of deviation between the target domain and the source domain, and determining the online gating coefficient in conjunction with the uncertainty index, include: Based on the domain adaptive loss over multiple time periods, determine the trend of domain adaptive loss changes; Obtain the historical output stability of the model and the constraints of onboard resources; The online gating coefficient is determined by integrating the regression confidence level, the domain adaptive loss change trend, the historical output stability, and the vehicle resource constraints among the uncertainty indicators.

7. The method according to claim 1, characterized in that, Based on the online gating coefficients, the steps of performing online domain adaptive updates to the model or outputting tire force estimates using a safe output strategy include: When the online gating coefficient exceeds the preset threshold, the feature extraction network in the force estimation model is updated online with small steps to align the source and target domains, so as to correct and output the estimated value of the tire three forces. When the online gating coefficient does not exceed the preset threshold, no online alignment update is performed, and a downgrade process is adopted, or the estimated value of the tire three-part force is determined as the final estimated value and output.

8. The method according to claim 7, characterized in that, The step of performing online alignment updates of the source and target domains of the feature extraction network in the force estimation model with small step sizes includes: With a preset low learning rate, one or more gradient descent operations are performed based on domain adaptive loss, and backpropagation is used to update the parameters in the feature extraction network.

9. The method according to claim 4, characterized in that, After the step of outputting the wear index based on the vehicle operating condition information, the method further includes: If the wear status label is inconsistent with the historical wear label status in the previous time window, it is determined whether the built-in confidence of the wear status label is higher than the threshold for a consecutive specified number of sliding windows, and if the confidence difference with the historical wear label exceeds a set value, a state switch is triggered, and the wear status label is used as the latest status label; the latest status label is used for the adaptive adjustment of subsequent domain alignment weights, kernel parameters and online update frequency, and serves as a reference input for security constraints and degradation strategies; After the state transition is triggered, the model output and adaptive step size are subjected to exponential smoothing to suppress the output abrupt change caused by the state transition.

10. The method according to claim 1, characterized in that, The training process of the force estimation model is as follows: Obtain a source domain labeled data training set; the source domain samples in the training set include three-channel input sequences and corresponding true values ​​of tire three-dimensional forces; The source domain samples are input into the feature extraction network and the regression network, and are trained with the goal of minimizing the source domain regression loss. The source domain regression loss integrates Smooth L1, MSE and Log-Cosh loss, and the weights of each loss term are adaptively balanced by a learnable variance parameter. Using source domain samples and unlabeled samples from the target wear domain, a joint domain alignment module is trained, and the domain adaptive loss is calculated. By adaptively balancing the source domain regression loss and the domain adaptive loss using entropy weight coefficients, the total model loss is obtained. This loss is then used to adjust the parameters of the feature extraction network, resulting in a force estimation model.

11. The method according to claim 10, characterized in that, The regression network also introduces physical constraint regularization: the vertical force is non-negative and consistent with the load variation; the amplitudes of the longitudinal and lateral forces are constrained by the adhesion coefficient and the friction ellipse; during training, a penalty term is added to samples that violate the constraints, or projection correction is performed during inference.

12. An intelligent tire force estimation device, characterized in that, The device includes: The signal acquisition module is used to acquire tire triaxial acceleration signals and vehicle operating condition information in real time and perform time alignment. The signal processing module is used to perform grounding event localization, contact segment extraction and linear interpolation resampling on the tire triaxial acceleration signal based on the vehicle operating condition information to obtain a three-channel input sequence. The model prediction module is used to input the three-channel input sequence into the offline trained force estimation model, output the tire three-part force estimate and the corresponding uncertainty index; and output the wear index based on the vehicle operating condition information. The deviation assessment module is used to determine the degree of deviation between the target domain and the source domain based on the unlabeled samples of the target domain and the labeled samples of the source domain within the most recent specified time period, as well as the wear index, and to determine the online gating coefficient in combination with the uncertainty index. The safety output module is used to perform online domain adaptive updates to the model based on the online gating coefficients or to output tire force estimates using a safety output strategy.

13. An intelligent tire force estimation system, characterized in that, The system includes: a controller and vehicle sensors; the controller is configured to perform the method as described in any one of claims 1-11.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions that, when invoked and executed by a processor, cause the processor to perform the method described in any one of claims 1-11.

Images