Tire wear prediction method based on Bayesian theory and deep learning
Through the Bayesian theory and deep learning method, a deep Bayesian residual network model is established, which solves the problems of low accuracy and poor real-time performance in tire wear prediction, and realizes high-precision tire wear prediction, reducing cost and safety risks.
Patent Information
- Application Number
- CN202510392602.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-31
- Publication Date
- 2025-08-29
AI Technical Summary
In the prior art, tire wear prediction methods have problems such as strong subjectivity, low efficiency, inability to monitor in real time and poor accuracy.
Using Bayesian theory and deep learning methods, we obtain the three-axis acceleration signal at the midpoint of the tire crown, perform preprocessing and time-frequency domain feature extraction, and establish a deep Bayesian residual network model, use mean square error and complexity penalty terms to build a loss function, train the model and output the tire wear prediction results.
High-precision tire wear prediction is achieved, avoiding premature tire replacement, reducing material waste and maintenance costs, reducing safety hazards, and improving the accuracy of prediction and early warning capabilities.
Smart Images
Figure CN120561718A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of tire wear prediction, and in particular to a tire wear prediction method based on Bayesian theory and deep learning. Background Art
[0002] As the only part of a vehicle that comes into contact with the ground, the wear of tires directly impacts driving safety, fuel efficiency, and the overall driving experience. With the development of intelligent transportation systems and connected vehicle technologies, tire wear prediction has become a key research area for vehicle maintenance and safety management. Predicting tire wear can provide early warnings, reduce accident risks, extend tire life, and reduce maintenance costs. Tire wear is affected by a variety of factors, including: Driving behavior: Rapid acceleration, sudden braking, and high-speed cornering accelerate tire wear. Road conditions: Road conditions such as pavement material, slope, and potholes significantly impact tire wear. Tire characteristics: A tire's material, structure, and tire pressure determine its wear resistance. Environmental factors: Temperature, humidity, and climatic conditions also affect tire wear rates.
[0003] The research and application of tire wear prediction methods are of great significance for improving driving safety and reducing maintenance costs. The combination of traditional methods and modern technologies provides a variety of solutions for tire wear prediction. Traditional tire wear detection methods mainly rely on manual inspection and regular maintenance, including: Visual inspection: Determine the degree of wear by observing the wear marks on the tire surface. Depth measurement: Use a depth gauge to measure the depth of the tire pattern to evaluate the wear. Regular replacement: Replace tires based on mileage or time period. Although these methods are simple and easy to implement, they have problems such as strong subjectivity, low efficiency, and inability to monitor in real time. Although traditional algorithm prediction models can also achieve the purpose of prediction, they still have the disadvantage of poor accuracy. Summary of the Invention
[0004] The purpose of the present invention is to provide a tire wear prediction method based on Bayesian theory and deep learning to solve the technical problems raised in the background technology.
[0005] To achieve the above object, the present invention provides the following technical solutions:
[0006] S1. Obtain the triaxial acceleration signal of the tire crown midpoint, preprocess it, and extract time-frequency domain features to construct a data set;
[0007] S2. Establish a deep Bayesian residual network structure model, wherein the deep Bayesian residual network structure model is composed of an input layer, a hidden layer, a projection layer, and an output layer in the order of input and output;
[0008] S3, construct the loss function L of the deep Bayesian residual network structure model based on the mean square error MSE;
[0009] S4. Divide the data set into a training set and a test set. Input the training set into the deep Bayesian residual network structure model for training until the loss function L converges. After the model training is completed, input the test set into the model for testing. Finally, the result output by the model output layer is the tire wear prediction result.
[0010] Furthermore, the specific steps of step S3 are:
[0011] S31. Use the negative log-likelihood term as the basis for loss function design, the mean square error MSE as the basic loss function, and calculate the data fitting term L fit , the calculation formula is:
[0012]
[0013] S32, introduce the logarithmic variance mean and calculate the complexity penalty term L pen , the calculation formula is:
[0014]
[0015] S33, the comprehensive data fitting term and complexity penalty term are used to obtain the loss function L, which is calculated as follows:
[0016]
[0017] In the above formula, y i is the true value of the i-th sample, μ i is the predicted mean of the i-th sample, is the prediction variance of the ith sample, and N is the number of samples.
[0018] Furthermore, several hidden layers are set, and each hidden layer is followed by a batch normalization layer and an activation function, and the activation function uses a Leaky ReLU function; the projection layer is also assisted by a residual connection.
[0019] Furthermore, the preprocessing in step S1 includes: coordinate transformation, digital filtering and data segmentation.
[0020] Furthermore, the time-frequency domain feature extraction in step S1 is specifically as follows: using a feature recognition algorithm to find specific peak-to-peak points and peak-to-valley points, and combining the signal statistical characteristics of the non-ground area, in terms of time domain characteristics, the peak value, peak-to-valley value, peak time difference, mean and variance are extracted from the signal after data segmentation; in terms of frequency domain characteristics, the center of gravity frequency, amplitude skewness and frequency standard deviation are calculated.
[0021] Furthermore, the data set includes time-frequency features, time-domain derived features, and 64-dimensional features consisting of tire pressure, wheel speed, and load. The training set accounts for 80% of the data set, and the test set accounts for 20% of the data set.
[0022] Beneficial effects: The present invention can make high-precision predictions of tire wear. By accurately predicting tire life, it can avoid premature replacement of tires that are still usable, reducing material waste and maintenance costs. It can also avoid safety hazards such as decreased grip and tire blowouts that may be caused by excessively worn tires. High-precision predictions can provide early warnings and reduce traffic accidents. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] In order to more clearly illustrate the technical solutions implemented in the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0024] Figure 1 is a flow chart of the tire wear prediction method of the present invention;
[0025] Figure 2 Schematic diagram of the deep Bayesian residual network structure of the present invention;
[0026] Figure 3 Schematic diagram of the structure of the residual connection residual block of the present invention;
[0027] Figure 4 This is a flow chart of the model test of the present invention;
[0028] Figure 5 This is a tire wear prediction result diagram of the present invention;
[0029] Figure 6 This is a comparison chart of the prediction performance evaluation between the present invention and the traditional prediction model. DETAILED DESCRIPTION
[0030] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0031] like Figures 1-6 As shown, the present invention provides a tire wear prediction method based on Bayesian theory and deep learning, and the specific steps are as follows:
[0032] S1. Obtain the three-axis acceleration signal of the tire crown midpoint, preprocess it, and extract time-frequency domain features to construct a data set, where the data set includes time-frequency features, time-domain derived features, and 64-dimensional features consisting of tire pressure, wheel speed, and load. This embodiment collects acceleration signal data of simulated working conditions under 78 uncertain working conditions, where the preprocessing includes: coordinate transformation, digital filtering, and data segmentation.
[0033] Coordinate transformation converts the acceleration signal from the global coordinate system to the follower coordinate system. This is accomplished by selecting target and auxiliary nodes, calculating the direction vector and basis vectors, and projecting them using a transformation matrix. Digital filtering utilizes a Chebyshev II low-pass filter with a 200Hz cutoff frequency to remove high-frequency noise. Data segmentation is performed based on the peaks and valleys of radial acceleration, ensuring independent analysis of each cycle. The radial acceleration signal is used as a benchmark, and its peaks and valleys are used to segment the cycle data. By detecting the peaks and valleys of radial acceleration, the range of each rolling cycle is determined, allowing for synchronous segmentation of the three-axis acceleration.
[0034] Time-frequency domain feature extraction: A feature recognition algorithm is used to find specific peak-to-peak points and peak-to-valley points, and combined with the signal statistical characteristics of the non-grounded area, in terms of time domain characteristics, in order to better identify the time domain waveform characteristics of the three-axis acceleration, the collected multi-cycle rolling signal is first segmented, and the peak value, peak-to-valley value, peak time difference, mean, variance, etc. are extracted from the segmented signal; in terms of frequency domain characteristics, the center of gravity frequency, amplitude skewness, and frequency standard deviation are calculated.
[0035] S2. Establish a deep Bayesian residual network structure model, wherein the deep Bayesian residual network structure model is composed of input layer, hidden layer, projection layer and output layer in the order of data input and output, such as Figure 2 As shown;
[0036] The input layer inputs 64-dimensional features consisting of time-frequency features, time-domain derived features, and tire pressure, wheel speed, and load;
[0037] In this embodiment, four hidden layers are set, and each hidden layer is followed by a batch normalization layer and an activation function. The activation function uses the Leaky ReLU function, which is an improved ReLU activation function that can solve the problem of zero gradient of the standard ReLU in the negative region and accelerate the convergence of the model.
[0038] The projection layer can adjust the dimension of the input features to match the input dimension of the subsequent layers of the network, and assist residual connections, directly passing the input to the output through skip connections, thereby avoiding information loss and gradient vanishing problems in deep networks. It enables the input features to be directly summed with the features after several transformations, which is conducive to promoting the flow of information and the propagation of gradients, thereby improving the performance of deep learning models.
[0039] The output layer outputs the predicted mean and standard deviation, so that the predicted value and uncertainty quantization range can be output later;
[0040] Deep Bayesian residual networks quantify the uncertainty of predictions by modeling the distribution of network weights. Unlike traditional neural networks that output single-point predictions, this model provides a mean and variance for each prediction, allowing it to output a quantified range of uncertainty.
[0041] S3. Construct the loss function L of the deep Bayesian residual network structure model based on the mean square error MSE. The loss function plays a core role in both machine learning and statistical modeling. The loss function defines the learning objective of the model. Choosing a suitable loss function can balance the bias and variance of the model and improve the generalization ability. The loss function not only affects the convergence and generalization ability of the model, but also affects the way the model handles uncertainty. A significant advantage of Bayesian theory is the ability to quantify the uncertainty of predictions, and the loss function plays a key role in this process. By minimizing the expected loss, the uncertainty boundary, confidence interval or risk range can be obtained to improve the accuracy of model predictions;
[0042] In this embodiment, in order to illustrate the implementation of step S3, the specific steps are as follows:
[0043] S31. Use the negative log-likelihood term as the basis for loss function design, the mean square error MSE as the basic loss function, and calculate the data fitting term L fit , the calculation formula is:
[0044]
[0045] Taking into account the uncertainty of the prediction (i.e., the prediction variance), the data fitting term is actually a weighted mean square error, where the weight is the inverse of the prediction variance. It can be seen as an extension of MSE to deal with uncertainty estimation problems;
[0046] S32. To prevent the model from overfitting, the logarithmic variance mean is introduced to calculate the complexity penalty term L pen , by regularizing the balance model uncertainty, the calculation formula is:
[0047]
[0048] S33, the comprehensive data fitting term and complexity penalty term are used to obtain the loss function L, which is calculated as follows:
[0049]
[0050] In the above formula, y i is the true value of the i-th sample, μi is the predicted mean of the i-th sample, is the prediction variance of the i-th sample, N is the number of samples;
[0051] The data fit term and the complexity penalty term together constitute the loss function of the model of this invention. They use different mechanisms to balance the model's fit to the training data with the model's complexity, thereby improving generalization ability and avoiding overfitting or underfitting. The data fit term measures how well the model fits the training data. The data fit term aims to measure the deviation between the model's predictions and the true observations, reflecting the model's ability to fit the known data. The complexity penalty term limits the model's complexity, preventing the model from overfitting the training data and improving generalization ability.
[0052] S4. Divide the data set into a training set and a test set. Input the training set into the deep Bayesian residual network structure model for training until the loss function L converges. After the model training is completed, input the test set into the model for testing. Finally, the output of the model output layer is the tire wear prediction result, where the training set accounts for 80% of the data set and the test set accounts for 20% of the data set.
[0053] The Adam optimization algorithm is added to the model training. In this example, the initial learning rate is set to 0.001 to balance the convergence speed and stability, and the weight decay is set to 1×10 -5 To prevent overfitting, the mean and variance are predicted through forward propagation to quantify the uncertainty.
[0054] During the model training process, the present invention adopts a dimensionality reduction algorithm to process the data after feature extraction from the data set, adopts a variance selection method to perform feature analysis to screen out invalid features, for example, uses the correlation coefficient method to select features, uses a dimensionality reduction algorithm and trial-and-error experiments to analyze the performance of different features, and continuously iterates the model to analyze and screen out features with good performance.
[0055] Since the task of the model is to predict tire wear, there are three types of wear conditions: wear of 0mm, 2.5mm, and 5mm. The data set contains these three types of conditions, and stratified random sampling is required to ensure the rationality of data distribution. Therefore, 5 samples are randomly selected from each type of condition, and a total of 15 sample data are used for testing, such as Figure 5 Shown are the prediction results of 15 random test conditions using the deep Bayesian residual network prediction model. The prediction model outputs the true value, predicted value, and prediction interval of the corresponding sample conditions. The results show that the prediction model can not only output relatively accurate tire wear prediction values in the case of point estimation, but also output the uncertain quantitative prediction range of tire wear, with an effective coverage rate of 93.3%, further improving the generalization ability and prediction accuracy of the tire wear prediction model in uncertain environments.
[0056] In order to better evaluate the prediction performance of the deep Bayesian residual network structure model, quantitative indicators such as root mean square error RMSE and determination coefficient R are used. 2 And the coverage of the prediction interval is defined as follows:
[0057]
[0058] In the above formula, N is the number of samples, y i is the true value of the i-th sample, is the predicted value of the i-th sample, is the average of all true values, I(·) is the indicator function, which is 1 if the condition is met and 0 otherwise. The root mean square error (RMSE) is used to measure the average difference between the predicted value and the true value, and gives higher weight to larger errors; the coefficient of determination (R) 2 It represents the proportion of the variation explained by the model to the total variation, ranging from negative infinity to 1. The closer the value is to 1, the better the model fit is. The coverage of the confidence interval refers to the proportion of actual observations that fall within the predicted confidence interval. It is used to measure the prediction effect of the uncertain quantitative range. The larger the coverage, the stronger the generalization ability.
[0059] To further demonstrate the prediction performance of the deep Bayesian residual prediction model, the commonly used BP neural network prediction method was used as a comparison, using the same data set, partitioning mode, and the same evaluation indicators. By performing a hyperparameter search for the number of hidden layer nodes, learning rate, and number of training times, all possible parameter combinations were traversed to select the optimal configuration. The Levenberg-Marquardt algorithm was used for training, and the loss function was the mean square error. The mean square error of the model was evaluated for different hyperparameter combinations, and the best model was finally selected by minimizing the mean square error. The results of the model evaluation indicators are compared. Figure 6 As shown in the figure, the root mean square error of the prediction model of the present invention is 0.1266mm, which is lower than that of the traditional prediction model, and the determination coefficient is 0.9962, which is higher than that of the traditional prediction model, indicating that the prediction model of the present invention has a better fitting effect. The results show that the established deep Bayesian residual network structure model has good prediction performance.
[0060] The present invention discloses a tire wear prediction method based on Bayesian theory and deep learning, which can make high-precision predictions of tire wear. By accurately predicting tire life, it can avoid premature replacement of tires that are still usable, reducing material waste and maintenance costs; it can also avoid safety hazards such as decreased grip and tire blowouts that may be caused by excessively worn tires. High-precision predictions can provide early warnings and reduce traffic accidents.
[0061] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.
[0062] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.
Claims
1. A tire wear prediction method based on Bayesian theory and deep learning, characterized in that: The following steps are involved: S1. Obtain the triaxial acceleration signal of the tire crown midpoint, preprocess it, and extract time-frequency domain features to construct a data set; S2. Establish a deep Bayesian residual network structure model, wherein the deep Bayesian residual network structure model is composed of an input layer, a hidden layer, a projection layer, and an output layer in the order of input and output; S3, construct the loss function L of the deep Bayesian residual network structure model based on the mean square error MSE; S4. Divide the data set into a training set and a test set. Input the training set into the deep Bayesian residual network structure model for training until the loss function L converges. After the model training is completed, input the test set into the model for testing. Finally, the result output by the model output layer is the tire wear prediction result.
2. The tire wear prediction method according to claim 1, characterized in that: The specific steps of step S3 are: S31. Use the negative log-likelihood term as the basis for loss function design, the mean square error MSE as the basic loss function, and calculate the data fitting term L fit , the calculation formula is: S32, introduce the logarithmic variance mean and calculate the complexity penalty term L pen , the calculation formula is: S33, the comprehensive data fitting term and complexity penalty term are used to obtain the loss function L, which is calculated as follows: In the above formula, y i is the true value of the i-th sample, μ i is the predicted mean of the i-th sample, is the prediction variance of the ith sample, and N is the number of samples.
3. The tire wear prediction method according to claim 1, wherein: There are several hidden layers, and each hidden layer is followed by a batch normalization layer and an activation function, wherein the activation function uses a Leaky ReLU function; the projection layer is also assisted by a residual connection.
4. The tire wear prediction method according to claim 1, wherein: The preprocessing in step S1 includes: coordinate transformation, digital filtering and data segmentation.
5. The tire wear prediction method according to claim 1, wherein: The time-frequency domain feature extraction in step S1 is specifically as follows: using a feature recognition algorithm to find specific peak-to-peak points and peak-to-valley points, and combining the signal statistical characteristics of the non-ground area, in terms of time domain features, extracting the peak value, peak-to-valley value, peak time difference, mean and variance from the signal after data segmentation; In terms of frequency domain characteristics, the centroid frequency, amplitude skewness and frequency standard deviation are calculated.
6. The tire wear prediction method according to claim 1, wherein: The data set includes time-frequency features, time-domain derived features, and 64-dimensional features consisting of tire pressure, wheel speed, and load. The training set accounts for 80% of the data set, and the test set accounts for 20% of the data set.