Tire pattern depth estimation method based on physical information fusion network

Abstract

The invention relates to a tread pattern depth estimation method based on a physical information fusion network, and belongs to the technical field of intelligent traffic systems and vehicle active safety. The method comprises the following steps: firstly, acquiring three-axis acceleration data under different working conditions through a tire built-in sensor; zero phase shift filtering, phase alignment and resampling preprocessing are carried out on the signals to form standard waveform data; then constructing a spatial-temporal feature extraction network, extracting time sequence features through a convolutional layer and a circulating layer, and fusing normalized load and speed working condition information; a physical constraint module based on a tire rigidity mechanism is introduced, the pattern depth and the vertical deformation amount are predicted by using learnable parameters, and a theoretical load is reversely deduced to establish a mechanical equilibrium constraint; and finally, network parameters and physical parameters are synchronously optimized by adopting a mixed loss function, and wear state grades and confidence coefficients are output through a probability statistical method. According to the method, the estimation precision and the physical interpretability of the model under the complex variable load working condition are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation systems and vehicle active safety technology, specifically relating to a tire tread depth estimation method based on a physical information fusion network. Background Technology

[0002] As the only component of a car in contact with the road surface, the health of tires directly determines the vehicle's power, braking performance, and driving safety. Tread depth is the most crucial quantitative indicator of tire wear. As vehicle mileage increases, tread depth gradually decreases, directly leading to a decline in tire water drainage capacity and reduced wet grip, greatly increasing the risk of hydroplaning and tire blowouts. Therefore, achieving accurate and real-time monitoring of tire tread depth is of significant practical importance for ensuring driving safety, optimizing fleet maintenance strategies, and reducing operating costs.

[0003] Traditional tire wear detection relies primarily on regular manual maintenance, using mechanical depth gauges to measure tread groove depth. This method is not only inefficient and time-consuming, but also has significant blind spots, failing to reflect sudden wear conditions during driving. With the development of sensor technology, non-contact detection devices based on laser triangulation and machine vision have emerged. However, these external detection devices are typically fixed at parking lot entrances or specific road sections, significantly affected by changes in ambient light, tire surface mud and water coverage, and cleanliness. Furthermore, their deployment costs are high, making large-scale deployment to in-vehicle terminals for mobile monitoring difficult.

[0004] In recent years, indirect monitoring methods based on smart tire technology have gradually become a research hotspot. This type of method collects dynamic vibration signals by integrating micro-sensors such as accelerometers and strain gauges into the tire liner, and uses signal processing and machine learning techniques to invert the wear state. However, existing data-driven methods still face many severe challenges in practical engineering applications: (1) Lack of physical interpretability. Currently, most mainstream convolutional neural networks or recurrent neural networks adopt a "black box" mapping mode, focusing only on the statistical correlation between data input and output, ignoring the physical properties of the tire itself, such as rubber stiffness and contact mechanics. When facing working conditions not included in the training set, this pure data-driven model often lacks physical constraints and is prone to outputting prediction results that violate physical laws, leading to doubts about the credibility of the system. (2) Generalization problem under complex variable load conditions. The load variation range of heavy commercial transport vehicles is large, and the huge difference in physical dimensions causes the data distribution of vibration signals to drift drastically. Existing pure data models lack the ability to adaptively perceive physical working conditions and cannot dynamically adjust weights to adapt to changes in physical boundary conditions during the feature extraction stage, resulting in a cliff-like drop in the estimation accuracy of the model under extreme working conditions.

[0005] In summary, there is an urgent need for a tire tread depth estimation method that can effectively integrate physical mechanism constraints and has adaptive perception capabilities under complex variable load conditions, in order to break through the bottlenecks of existing technologies. Summary of the Invention

[0006] The purpose of this invention is to provide a tire tread depth estimation method based on physical information fusion network, which can overcome the shortcomings of the prior art such as poor physical interpretability and weak generalization ability under heavy load conditions.

[0007] To achieve the above objectives, the technical solution of the present invention is: a tire tread depth estimation method based on a physical information fusion network, comprising:

[0008] Step S1: Using an intelligent tire system with a triaxial acceleration sensor attached to the inner liner of the tire, triaxial acceleration data of tires with different wear levels under different load and speed conditions are obtained.

[0009] Step S2: Perform zero-phase-shift filtering to denoise the collected acceleration data, use the feature point matching method to align the phases of signals from different channels, and divide the continuous signal into single-circle samples according to the tire rotation cycle characteristics. Then, resample the samples to unify each circle into standard waveform data of a fixed length.

[0010] Step S3: Construct a spatiotemporal feature extraction neural network and use convolutional and recurrent layers to extract the temporal features of standard waveform data; at the same time, introduce a physical condition fusion mechanism to expand the normalized load and speed data into vectors, and concatenate them with waveform features in the feature extraction stage to obtain deep features containing condition information.

[0011] Step S4: Construct a physical constraint module based on tire stiffness mechanism, set learnable physical parameters, use deep features to predict tire tread depth and vertical deformation, and use the predicted values ​​to calculate the theoretical load based on the mechanical equilibrium formula.

[0012] Step S5: Train the network by establishing a hybrid loss function that includes data error and physical consistency error, update the network weights and physical parameters synchronously, and use probabilistic statistical methods to process the continuous prediction values ​​of the network, outputting the tire wear status level and confidence level.

[0013] Furthermore, in step S1, the data acquisition includes: selecting an all-steel radial heavy-duty commercial vehicle tire, attaching and installing a triaxial acceleration sensor at the center of the tire crown in the inner liner, and collecting the acceleration vibration signals of the tire along the tangential and radial directions in real time through the sensor, while simultaneously recording the real-time vertical load L and driving speed V of the tire, thus constructing an orthogonal experimental condition covering the wear state of the entire life cycle and multi-level load and speed combinations.

[0014] Furthermore, in step S2, the signal preprocessing includes: constructing a fourth-order Butterworth low-pass filter and using a bidirectional filtering strategy to perform zero-phase-shift denoising on the signal; using the ground impact extreme point of the radial acceleration signal as the absolute time anchor point, calculating the local mean within the dynamic search window of the tangential acceleration signal. with standard deviation Construct an adaptive feature search threshold:

[0015]

[0016]

[0017] Where k is the statistical sensitivity coefficient; the search satisfies greater than The peak and its immediate neighbor satisfy less than Calculate the time difference between the center and the anchor point of the trough. And perform reverse translation alignment:

[0018]

[0019] in, This represents the value of the aligned tangential strain signal at time t. This represents the original tangential strain signal. This represents the calculated time lag.

[0020] Furthermore, in step S2, the signal preprocessing also includes: identifying adjacent grounding impact anchor points. and Define the slice boundary as The single-cycle sample is extracted and resampled into a fixed N-point angle domain waveform through linear interpolation, and finally Z-Score normalization is performed.

[0021] Furthermore, in step S3, the construction of the spatiotemporal feature extraction neural network includes: using a feature extraction module composed of three stacked temporal convolutional residual blocks, each layer adopting a causal dilated convolutional structure with dilation rates of d=1, 2, and 4 respectively, and configuring a Chomp1d causal pruning layer and a SiLU activation function after the convolutional layer.

[0022] Furthermore, in step S3, the physical condition fusion mechanism includes: setting a load normalization benchmark. With speed normalization benchmark Dimensional normalization is performed on the real-time load L and velocity V:

[0023]

[0024]

[0025] scalar and Copy and expand into a time series vector in the time dimension. and Before entering the bidirectional gated recurrent unit layer, the expanded temporal vector is combined with the waveform feature tensor output by the convolutional layer. The components are concatenated along the channel dimension to form a hybrid feature tensor. :

[0026]

[0027] in, This represents the splicing operator along the feature channel dimension;

[0028] Subsequently, a bidirectional gated recurrent unit (Bi-GRU) is used to extract global temporal features, and weights are calculated through a self-attention mechanism to focus on the ground imprint region and output a deep feature vector.

[0029] Furthermore, in step S4, the construction of the physical constraint module includes: setting a learnable tire stiffness coefficient. With material structure factor And it is indexed to ensure physical positive definiteness:

[0030]

[0031] in, These are the unconstrained weight parameters that the network actually updates during backpropagation; the network output layer branches predict the pattern depth respectively. With vertical deformation ,in The activation function in Softplus is constrained to a positive value:

[0032]

[0033] in, For smoothing coefficients, This is the linear output of the fully connected layer.

[0034] Furthermore, in step S4, the mechanical equilibrium formula includes: defining the instantaneous stiffness of the tread based on the improved brush model. The total vertical stiffness of the tire is defined based on a series spring model. By combining Hooke's Law to construct mechanical equilibrium equations, the theoretical load can be derived. :

[0035] .

[0036] Furthermore, in step S5, the hybrid loss function The mathematical expression is:

[0037]

[0038] in, For physical weighting coefficients, The error is data-driven and is calculated using mean square error. To account for physical consistency errors and avoid gradient explosion caused by large load values, a normalized ratio is used for construction:

[0039]

[0040] in, To infer the load from the theory, For the actual data acquisition load, This is a numerical stability constant; during training, the gradient descent algorithm is used to synchronously update the network weights and physical parameters. and .

[0041] Furthermore, in step S5, the probability statistics method includes: pre-setting a set of standard wear state anchor points covering the entire life cycle of the tire. ; Calculate the continuous predicted value z of the network output belonging to each anchor point The Gaussian likelihood probability is calculated and transformed into a posterior probability distribution using a Gaussian kernel function. :

[0042]

[0043] Selection probability The largest anchor point is output as the final wear state, and the information entropy H of this distribution is calculated:

[0044]

[0045] When the information entropy H is greater than the preset warning threshold, a low confidence warning signal is output.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] (1) This invention constructs a spatiotemporal feature extraction network for physical perception and utilizes signal phase locking and physical condition covariate embedding mechanism to effectively solve the problem of physical phase lag of multi-source signals. At the same time, it endows the model with adaptive perception capability for wide load and wide speed domain conditions, significantly reduces the impact of data distribution drift on prediction accuracy, and ensures the robustness and generalization performance of the model under extreme conditions.

[0048] (2) This invention transforms the “black box” prediction of the neural network into a forward inference that conforms to the laws of tire mechanics by using a physical constraint layer with embedded stiffness mechanism and Bayesian probabilistic decision-making mechanism, effectively avoiding prediction errors that violate physical common sense and giving the model physical interpretability; at the same time, it uses probabilistic statistical methods to map a single regression value into a state distribution with confidence, providing a decision basis with both high precision and high reliability for the vehicle safety management system, and greatly improving the safety and accuracy of the monitoring system. Attached Figure Description

[0049] Appendix Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0050] Appendix Figure 2 This is a schematic diagram of multidimensional signal phase alignment according to an embodiment of the present invention;

[0051] Appendix Figure 3 This is a network architecture diagram of spatiotemporal feature extraction and physical constraint fusion according to an embodiment of the present invention;

[0052] Appendix Figure 4 This is a convergence performance curve of an embodiment of the present invention during the model training process;

[0053] Appendix Figure 5 This is a performance ladder diagram of pattern depth prediction according to an embodiment of the present invention.

[0054] Appendix Figure 6 This is a schematic diagram of the wear state classification confusion matrix according to an embodiment of the present invention. Detailed Implementation

[0055] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0056] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0057] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0058] like Figure 1 As shown, the present invention provides a tire tread depth estimation method based on a physical information fusion network, comprising the following steps;

[0059] Step S1: Using an intelligent tire system with a triaxial acceleration sensor attached to the inner liner of the tire, triaxial acceleration data of tires with different wear levels under different load and speed conditions are obtained.

[0060] Step S2: Perform zero-phase-shift filtering to denoise the collected acceleration data, use the feature point matching method to align the phases of signals from different channels, and divide the continuous signal into single-circle samples according to the tire rotation cycle characteristics. Then, resample the samples to unify each circle into standard waveform data of a fixed length.

[0061] Step S3: Construct a spatiotemporal feature extraction neural network and use convolutional and recurrent layers to extract the temporal features of standard waveform data; at the same time, introduce a physical condition fusion mechanism to expand the normalized load and speed data into vectors, and concatenate them with waveform features in the feature extraction stage to obtain deep features containing condition information.

[0062] Step S4: Construct a physical constraint module based on tire stiffness mechanism, set learnable physical parameters, use deep features to predict tire tread depth and vertical deformation, and use the predicted values ​​to calculate the theoretical load based on the mechanical equilibrium formula.

[0063] Step S5: Train the network by establishing a hybrid loss function that includes data error and physical consistency error, update the network weights and physical parameters synchronously, and use probabilistic statistical methods to process the continuous prediction values ​​of the network, outputting the tire wear status level and confidence level.

[0064] In step S1, the raw acceleration data of tires with different wear levels are obtained. Step S1 is implemented as follows:

[0065] First, a 12R22.5 all-steel radial heavy-duty commercial vehicle tire was selected as the experimental subject. A triaxial accelerometer was mounted on the center of the tire crown using a high-strength flexible adhesive. This mounting method ensures a tight mechanical coupling between the sensor and the tire carcass, guaranteeing the sensor's stability and durability under high-speed rotation and repeated ground impact conditions without disrupting the tire's dynamic balance and normal rolling deformation. The sensor's sampling frequency was set to 1600Hz to fully capture the high-frequency vibration characteristics of the tire at the moment of contact with the ground.

[0066] The experiment was conducted on a high-precision drum testing machine. During the experiment, the built-in sensors collected the acceleration vibration signals of the tire along the tangential (X-axis) and radial (Z-axis) directions in real time, and sent the data packets to the host computer processing unit via a wireless transmission module. Simultaneously, the control system of the drum testing machine recorded the current tire vertical load value L and the driving speed value V. To verify the effectiveness of the method throughout the entire life cycle and over a wide load range, this embodiment constructed a multi-dimensional orthogonal experimental condition: In the wear dimension, six tire states with tread depths of 3mm, 6mm, 9mm, 12mm, 15mm, and 17mm were pre-fabricated through physical grinding, where 17mm represents the new tire state and 3mm represents the severely worn state; in the working condition dimension, three load levels of 1000kg (light load), 2300kg (half load), and 3500kg (full load) were set, as well as two steady-state speed levels of 40km / h and 60km / h, thus completing the data collection of tires with different wear levels under various load and speed combinations.

[0067] In step S2, interference from sensor hardware characteristics and environmental noise on the original signal is eliminated, and the asynchronous, variable-length time-domain signal is converted into a standard spatiotemporal tensor that can be processed by a neural network. The implementation method for step S2 is as follows:

[0068] To address the high-frequency mechanical noise caused by the complex electromagnetic environment inside the tire and road surface texture, a zero-phase-shift filtering technique is employed to clean the raw data. While traditional Infinite Impulse Response (IIR) filters effectively remove noise, they introduce nonlinear phase delays, causing key waveform features to drift along the time axis. Therefore, this embodiment constructs a fourth-order Butterworth low-pass filter and employs a bidirectional filtering strategy: first, the discrete signal sequence is forward-filtered; then, the output sequence is inverted along the time axis and filtered a second time; finally, it is inverted back to the forward direction. This process mathematically cancels out the phase shift, ensuring that the smoothed signal after filtering is strictly synchronized with the original physical events in the time dimension.

[0069] To address the inherent time lag between radial and tangential acceleration data, a phase-locking algorithm is proposed. The algorithm uses the radial acceleration signal, which exhibits the most significant signal characteristics, as the absolute reference. It identifies a sequence of negative extreme points in the Z-axis signal that characterize the massive impact at the moment of tire contact with the ground and defines these points as absolute time anchors. Subsequently, a dynamic search window is set on the tangential acceleration signal, centered on each anchor point. Within this window, the algorithm constructs an adaptive feature search threshold, calculated using the following formula:

[0070]

[0071]

[0072] in, This represents the adaptive upper bound threshold for determining the peak value. This represents the adaptive lower bound threshold for determining the trough. This represents the local mean of the tangential strain signal within the search window. This represents the local standard deviation of the tangential strain signal within the search window. This represents the preset statistical sensitivity coefficient, used to adjust the sensitivity of feature point recognition.

[0073] The system searches within the window for conditions greater than The peak feature points and their immediate left side satisfy less than The trough feature points. Once this feature combination is locked, the time difference between its geometric center and the Z-axis anchor point is calculated. This time difference is the physical phase lag, and the system uses this lag to perform a reverse translation operation on the entire X-axis signal.

[0074]

[0075] in, This represents the value of the aligned tangential strain signal at time t. This represents the original tangential strain signal. This represents the calculated time lag.

[0076] After signal alignment, the periodic characteristics of the radial acceleration signal are used to transform continuous time-series data into independent physical samples. The algorithm identifies two adjacent grounding impact anchor points. To ensure that the core information contained in the tire grounding imprint is located at the center of the sample, the slice boundary is defined as follows:

[0077]

[0078] in, This represents the index of the split point of the data in the i-th lap (i.e., the end point of the i-th lap and the start point of the (i+1)-th lap). This represents the Z-axis wave crest anchor point index corresponding to the i-th grounding impact. This represents the Z-axis wave crest anchor point index corresponding to the (i+1)th grounding impact, with the system extracting the interval. The data within is used as a complete single-lap sample.

[0079] Considering that changes in vehicle speed will lead to inconsistent sampling points per lap, a linear interpolation function is constructed to meet the requirement of a fixed input dimension for the neural network and achieve "rotational speed independence" of the data. This function can then be used to interpolate data of arbitrary length. The time-domain slice data is uniformly mapped to a fixed N points (in this embodiment, N=360, corresponding to 360 degrees of a tire's rotation). Finally, to eliminate the interference of signal amplitude differences under different load conditions on the model training weights, the resampled samples are Z-score standardized, calculated using the following formula:

[0080]

[0081] in, This represents the signal value after standardization. Represents the original signal value. This represents the mean of all sampling points in the current single-lap sample. This represents the standard deviation of all sampling points in the current single-lap sample. It is a numerical stability constant used to prevent division errors caused by a standard deviation of zero.

[0082] In step S3, deep features that exhibit a highly nonlinear mapping relationship with tire wear state are extracted from the preprocessed and standardized spatiotemporal waveform tensor. The implementation method for step S3 is as follows:

[0083] First, the standardized single-cycle waveform tensor is input into a temporal convolutional network module to capture local short-term dependencies and micro-texture variations in the signal. To effectively capture long-period vibration modes without significantly increasing the number of network parameters and computational complexity, the main body of this module consists of three stacked temporal convolutional residual blocks, all employing a "causal dilated convolution" structure. For the input time series X and the convolutional filter f, the dilated convolution operation at time t is defined as follows:

[0084]

[0085] in, This represents the output feature value of the convolutional layer at time t. Represents the input sequence at past time points The element value, Let represent the value of the i-th weight in the convolution kernel, and k represent the size of the convolution kernel. In this embodiment, k is set to 0. d represents the dilation rate, which is used to control the interval of sampling points in the convolution kernel.

[0086] In this embodiment, the dilation rates of the three consecutively stacked convolutional blocks are set sequentially as follows: Through this exponentially increasing expansion rate design, the effective receptive field of the network expands rapidly with increasing layer depth, thus covering the long-term evolution characteristics of the tire contact process. Furthermore, to strictly ensure the temporal causality of signal processing—that is, the output at the current moment depends only on the input at the current and historical moments, and the use of future information is strictly prohibited—a Chomp1d causal pruning layer is configured after each convolutional layer to remove the 2d padding points generated at the end of the sequence by the convolution operation. Simultaneously, considering the alternating positive and negative and drastic changes in tire vibration signals, to preserve the gradient information of weak signals and prevent neuron "death," this embodiment uses the SiLU smooth nonlinear activation function instead of the traditional ReLU function. Its mathematical expression is:

[0087]

[0088] Where x represents the linear output value of the convolutional layer. This represents the standard Sigmoid activation function. After the data stream passes through three convolutional blocks, the number of feature channels is progressively increased from the initial 2 dimensions (corresponding to Z-axis acceleration and X-axis acceleration) to 16, 32, and 64 dimensions, ultimately outputting a waveform feature tensor containing rich local texture information. .

[0089] Secondly, considering the large load span of heavy commercial vehicles, directly inputting the original physical quantities into the neural network would lead to gradient update imbalance during backpropagation due to significant numerical differences, potentially even causing gradient explosion. Therefore, this embodiment introduces a physical condition covariate embedding mechanism in the feature extraction stage. A load normalization baseline constant is set. and velocity normalized reference constant The system acquires the current vertical load L and driving speed V in real time, and calculates the dimensionless normalized physical scalar:

[0090]

[0091] in, This represents the normalized load covariates, whose values ​​are constrained to... Within the linear response interval; This represents the normalized velocity covariate. To ensure that the subsequent recurrent neural network can explicitly perceive the current physical boundary conditions at each time step of the recursive computation, a "channel-level pre-fusion" strategy is executed. The above scalar... and A time-series vector is generated by copying and expanding the vector along the time dimension to match the time step of the TCN output feature map. Then, a channel-level concatenation operation is performed before entering the recurrent network layer.

[0092]

[0093] in, This represents the mixed feature tensor input to the subsequent recurrent neural network, with dimension . Where N=360 is the time step and D=64 is the number of waveform feature channels; This represents the waveform feature tensor output by the TCN layer; Indicates by The extended load time-series vector; Indicates by The expanded velocity-time vector; This represents the splicing operator along the feature channel dimension. Through this step, physical condition information is introduced into the feature space, enabling the network to adaptively adjust the extraction weights of waveform features according to the current load and velocity state, significantly solving the problem of easy confusion of single waveform features under varying load conditions.

[0094] Finally, the hybrid feature tensor containing physical condition information is... The input is fed into the Bidirectional Gated Cyclic Unit (Bi-GRU) module. This module uses a gating mechanism to control the information flow, enabling it to capture the global evolution of the signal throughout the entire rotation cycle. For t time steps, the state update calculation within the GRU unit is as follows:

[0095]

[0096] in, The current time step is the mixed feature input vector; This is the hidden state vector from the previous time step; To update the gate, it is determined how much historical information should be retained and transmitted to the current moment; To reset the gate, determine how much historical noise unrelated to the current prediction should be forgotten; This represents the candidate hidden state at the current moment; This represents the final hidden output state at the current moment; All are learnable weight matrices; This represents the Hadamard product (element-wise multiplication). This embodiment employs a bidirectional structure, calculating the forward and backward hidden states separately, and concatenating them to obtain a complete context feature vector of dimension 128. To automatically filter out ground imprint regions containing rich texture information from the entire 360-degree signal, a self-attention mechanism is introduced after the Bi-GRU layer. This mechanism assigns weights by calculating the energy score at each time step; the specific calculation formula is as follows:

[0097]

[0098] in, This represents the non-normalized energy score at time step t, which is calculated by the fully connected layer. represents the normalized attention weight at time step t, reflecting the importance of the feature at that moment; N represents the total time step length; Let C represent the hidden state vector output by the Bi-GRU at time step t; C represents the final generated deep spatiotemporal feature vector. This feature vector C can adaptively ignore invalid signals in the air rotation region and highly focus on the mechanical deformation characteristics during the tire-ground contact process, providing a high signal-to-noise ratio input for subsequent physical parameter inversion.

[0099] In step S4, a physical constraint model based on tire mechanics is embedded at the end of the deep network to give the network physical interpretability; the implementation method of step S4 is as follows:

[0100] To automatically identify the inherent physical properties of tires during network training, two learnable scalar parameters are defined: tire stiffness coefficient. With material structure factor Since physical quantities such as stiffness are always positive in the real world, in order to ensure the physical meaning of the numerical calculation and improve the stability of the optimization, this embodiment adopts an exponential reparameterization strategy to define the parameters:

[0101]

[0102] in, These are the unconstrained weight parameters that the network actually updates during backpropagation. It represents the equivalent vertical stiffness of the tire carcass, reflecting the basic support capacity provided by air pressure and sidewall structure; These are material factors related to the tread compound properties and tread geometry, used to characterize the sensitivity of tread stiffness to changes in wear depth. These two parameters are assigned initial values ​​at the start of training and are automatically identified and updated during the training process.

[0103] The network's end is designed as a dual-head decoupled output structure, which predicts the tire's geometric state and motion state respectively: (1) Branch A directly regresses through a fully connected layer to output the current tread depth estimate. (2) Branch B outputs the vertical deformation of the tire under the current load through the fully connected layer regression. To strictly satisfy the physical constraint that "the deformation must be greater than zero," this branch uses the Softplus activation function for nonlinear mapping:

[0104]

[0105] in, The linear output of the fully connected layer; As a smoothing coefficient, this function guarantees the output value. Furthermore, it has a smooth gradient near zero, thus avoiding the truncation effect of the ReLU function at zero.

[0106] Based on the above predicted values and A physical reasoning path is constructed to inversely derive the theoretical load. This embodiment employs a constitutive theory combining an improved brush model and a series spring model. First, the instantaneous stiffness of the tire tread is defined. With pattern depth The constitutive relation. According to the brush model theory, the higher the pattern block ( The larger the value, the weaker its resistance to shear and compression, and the lower its stiffness. Therefore, the following inverse proportional mapping relationship is established:

[0107]

[0108] in, Indicates the equivalent vertical stiffness of the tread pattern portion; To prevent tiny constants with a denominator of zero.

[0109] Secondly, the tire is considered as a series elastic system consisting of a "carcass" and a "tread". The total vertical stiffness of the tire is calculated using the series spring stiffness coupling formula. :

[0110]

[0111] Finally, based on Hooke's law and the static equilibrium equations, the theoretical load under the current working condition is derived by using the calculated total stiffness and predicted deformation. :

[0112]

[0113] In subsequent training, through constraints By approximating the load L measured by the real sensor, the network can learn the correct pattern depth prediction rules, thus giving the model a strong physical generalization ability.

[0114] In step S5, a closed-loop optimization system is constructed. On the one hand, a hybrid loss function is used to ensure that the neural network simultaneously follows the data distribution pattern and the laws of tire mechanics, achieving synchronous identification of model weights and physical parameters. On the other hand, a Bayesian statistical inference mechanism is introduced to transform the network's continuous regression output into discrete state decisions with confidence assessment. The implementation method of step S5 is as follows:

[0115] To overcome the drawbacks of pure data-driven models, such as overfitting and lack of physical constraints, this embodiment constructs a hybrid loss function that includes both data-driven error and physical consistency error. The mathematical expression for the total loss function is defined as:

[0116]

[0117] in, This represents the total loss value used for backpropagation; Indicates the loss of data-driven terms; Represents the loss of the physical consistency term; This represents the weighting coefficients of the physical constraints, used to balance the magnitudes of the two losses and ensure the stability of the optimization process. Data-driven error term. Mean squared error (MSE) is used to measure the difference between the pattern depth predicted by the network and the true label:

[0118]

[0119] Where M represents the size of the training batch; This represents the pattern depth predicted by the network for the i-th sample; This represents the true pattern depth label corresponding to the i-th sample. This is used to measure the relative difference between the theoretical load derived from the physical layer and the actual load. To address the issue of gradient explosion caused by massive load values ​​under heavy load conditions, this embodiment constructs the loss term using a normalized ratio:

[0120]

[0121] in, This represents the theoretically calculated load of the i-th sample obtained by back-reasoning from the physical model in step S4; This represents the actual load of the i-th sample recorded by a sensor or test bench during the experiment. To prevent numerical stability constants with a denominator of zero, the Adam gradient descent optimization algorithm is used to minimize them during model training. This optimization process not only updates the synaptic weights of the neural network, but also simultaneously calculates information about the physical parameters. and The gradient is used to achieve the desired effect on the tire stiffness coefficient. With material structure factor The online adaptive identification enables the model to automatically learn the physical properties of the current tire.

[0122] To quantify the uncertainty of the estimation results, this embodiment employs a Bayesian probabilistic statistical method to process the continuous regression output of the network during the inference phase. First, a set of standard wear state anchor points covering the entire tire lifecycle is preset:

[0123]

[0124] These correspond to different wear levels. Next, the continuous predicted value z (i.e., ...) of the network output is calculated. The likelihood probability of each anchor point is calculated. The Euclidean distance between the predicted value and the anchor point is converted into a probability density function using a Gaussian kernel function, and then normalized to a posterior probability distribution using the Softmax function.

[0125]

[0126] in, This indicates that the current predicted value z belongs to the j-th wear state anchor point. The posterior probability; This is the bandwidth parameter of the Gaussian kernel function. Finally, the system executes a dual-decision logic:

[0127] (1) State determination: Select the anchor point with the highest posterior probability as the final output pattern depth state:

[0128]

[0129] (2) Confidence assessment: Calculate the information entropy H of the posterior probability distribution:

[0130]

[0131] If the calculated information entropy H is greater than the preset warning threshold This indicates that the predicted value lies in the ambiguous zone between the two state anchor points, and the system will output a "low confidence" warning signal, prompting the user or the back-end management system to conduct manual verification or wait for more data confirmation. This probabilistic output mechanism significantly improves the robustness and security of the monitoring system.

[0132] This paper addresses the challenges of asynchronous multi-source signals and feature distribution drift in heavy-duty commercial vehicles under complex variable load conditions. It proposes a tire tread depth estimation method based on a physical information fusion network. This method can achieve deep coupling between physical mechanism constraints and data-driven features, effectively solving the problems of feature extraction failure and uninterpretable black-box models under wide load domains, and improving the robustness and safety of intelligent tire monitoring systems.

[0133] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0134] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0135] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0136] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0137] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for estimating tire tread depth based on a physical information fusion network, characterized in that, include: Step S1: Using an intelligent tire system with a triaxial acceleration sensor attached to the inner liner of the tire, triaxial acceleration data of tires with different wear levels under different load and speed conditions are obtained. Step S2: Perform zero-phase-shift filtering to denoise the collected acceleration data, use the feature point matching method to align the phases of signals from different channels, and divide the continuous signal into single-circle samples according to the tire rotation cycle characteristics. Then, resample the samples to unify each circle into standard waveform data of a fixed length. Step S3: Construct a spatiotemporal feature extraction neural network and use convolutional and recurrent layers to extract the temporal features of standard waveform data; at the same time, introduce a physical condition fusion mechanism to expand the normalized load and speed data into vectors, and concatenate them with waveform features in the feature extraction stage to obtain deep features containing condition information. Step S4: Construct a physical constraint module based on tire stiffness mechanism, set learnable physical parameters, use deep features to predict tire tread depth and vertical deformation, and use the predicted values ​​to calculate the theoretical load based on the mechanical equilibrium formula. Step S5: Train the network by establishing a hybrid loss function that includes data error and physical consistency error, update the network weights and physical parameters synchronously, and use probabilistic statistical methods to process the continuous prediction values ​​of the network, outputting the tire wear status level and confidence level.

2. The tire tread depth estimation method based on a physical information fusion network according to claim 1, characterized in that, In step S1, the data acquisition includes: selecting an all-steel radial heavy-duty commercial vehicle tire, attaching and installing a triaxial acceleration sensor at the center of the tire crown in the inner liner, and collecting the acceleration vibration signals of the tire along the tangential and radial directions in real time through the sensor, while simultaneously recording the real-time vertical load L and driving speed V of the tire, and constructing an orthogonal experimental condition covering the wear state of the entire life cycle and multi-level load and speed combinations.

3. The tire tread depth estimation method based on a physical information fusion network according to claim 1, characterized in that, In step S2, the signal preprocessing includes: constructing a fourth-order Butterworth low-pass filter and performing zero-phase-shift denoising on the signal using a bidirectional filtering strategy; and calculating the local mean within the dynamic search window of the tangential acceleration signal, using the ground impact extreme point of the radial acceleration signal as the absolute time anchor point. with standard deviation Construct an adaptive feature search threshold: Where k is the statistical sensitivity coefficient; the search satisfies greater than The peak and its immediate neighbor satisfy less than Calculate the time difference between the center and the anchor point of the trough. And perform reverse translation alignment: in, This represents the value of the aligned tangential strain signal at time t. This represents the original tangential strain signal. This represents the calculated time lag.

4. The tire tread depth estimation method based on a physical information fusion network according to claim 3, characterized in that, In step S2, the signal preprocessing further includes: identifying adjacent grounding impact anchor points. and Define the slice boundary as The single-cycle sample is extracted and resampled into a fixed N-point angle domain waveform through linear interpolation, and finally Z-Score normalization is performed.

5. The tire tread depth estimation method based on a physical information fusion network according to claim 1, characterized in that, In step S3, the construction of the spatiotemporal feature extraction neural network includes: using a feature extraction module composed of three stacked temporal convolutional residual blocks, each layer adopting a causal dilated convolutional structure with dilation rates of d=1, 2, and 4 respectively, and configuring a Chomp1d causal pruning layer and a SiLU activation function after the convolutional layer.

6. The tire tread depth estimation method based on a physical information fusion network according to claim 5, characterized in that, In step S3, the physical condition fusion mechanism includes: setting a load normalization benchmark. With speed normalization benchmark Dimensional normalization is performed on the real-time load L and velocity V: scalar and Copy and expand into a time series vector in the time dimension. and Before entering the bidirectional gated recurrent unit layer, the expanded temporal vector is combined with the waveform feature tensor output by the convolutional layer. The components are concatenated along the channel dimension to form a hybrid feature tensor. : in, This represents the splicing operator along the feature channel dimension; Subsequently, a bidirectional gated recurrent unit (Bi-GRU) is used to extract global temporal features, and weights are calculated through a self-attention mechanism to focus on the ground imprint region and output a deep feature vector.

7. The tire tread depth estimation method based on a physical information fusion network according to claim 1, characterized in that, In step S4, the construction of the physical constraint module includes: setting a learnable tire stiffness coefficient. With material structure factor And it is indexed to ensure physical positive definiteness: in, These are the unconstrained weight parameters that the network actually updates during backpropagation; the network output layer branches predict the pattern depth respectively. With vertical deformation ,in The activation function in Softplus is constrained to a positive value: in, For smoothing coefficients, This is the linear output of the fully connected layer.

8. The tire tread depth estimation method based on a physical information fusion network according to claim 7, characterized in that, In step S4, the mechanical equilibrium formula includes: defining the instantaneous stiffness of the tread based on the improved brush model. The total vertical stiffness of the tire is defined based on a series spring model. By combining Hooke's Law to construct mechanical equilibrium equations, the theoretical load can be derived. : 。 9. The tire tread depth estimation method based on a physical information fusion network according to claim 1, characterized in that, In step S5, the hybrid loss function The mathematical expression is: in, For physical weighting coefficients, The error is data-driven and is calculated using mean square error. To account for physical consistency errors and avoid gradient explosion caused by large load values, a normalized ratio is used for construction: in, To infer the load from the theory, For the actual data acquisition load, This is a numerical stability constant; during training, the gradient descent algorithm is used to synchronously update the network weights and physical parameters. and .

10. The tire tread depth estimation method based on a physical information fusion network according to claim 1, characterized in that, In step S5, the probability statistics method includes: pre-setting a set of standard wear state anchor points covering the entire life cycle of the tire. ; Calculate the continuous predicted value z of the network output belonging to each anchor point The Gaussian likelihood probability is calculated and transformed into a posterior probability distribution using a Gaussian kernel function. : Selection probability The largest anchor point is output as the final wear state, and the information entropy H of this distribution is calculated: When the information entropy H is greater than the preset warning threshold, a low confidence warning signal is output.

Images