A Smart Tire Slip Angle Estimation Method Based on CNN-Transformer

Abstract

This invention relates to an intelligent tire slip angle estimation method based on CNN-Transformer, belonging to the field of intelligent vehicle perception technology. The method includes: Step S1, acquiring acceleration signals of a fully loaded tire under various slip conditions using a triaxial accelerometer, and performing 200Hz low-pass filtering and 8:2 data partitioning; Step S2, designing a convolutional neural network (CNN) with four one-dimensional convolutional layers to extract deep features from the preprocessed signal and retain temporal information by combining positional encoding; Step S3, using a multi-head cross-attention module to aggregate and enhance the extracted features, filtering out interference information and highlighting key features related to the slip angle; Step S4, inputting the processed test data into a trained CNN-Transformer model to predict the tire slip angle. This invention significantly improves the accuracy and robustness of slip angle estimation by combining the advantages of CNN in local feature extraction and Transformer in global feature attention, and is applicable to various tire specifications and operating conditions.

Description

Technical Field

[0001] This invention belongs to the field of intelligent vehicle perception technology, specifically relating to an intelligent tire slip angle estimation method based on CNN-Transformer. Background Technology

[0002] Tire sideslip angle is a key parameter reflecting vehicle dynamic behavior, directly affecting vehicle stability and handling. Accurate estimation of sideslip angle can not only maintain vehicle stability and reduce accident risk under various complex road conditions, but also provide necessary handling support in emergency situations to ensure driving safety. There are three main existing methods for estimating tire sideslip angle: (1) Based on vehicle dynamics and tire mechanics models, design linear or nonlinear observers and use Kalman filters to estimate tire sideslip angle. This method relies on accurate vehicle and tire parameter models and uses Kalman filters for state estimation. However, in actual driving environments, vehicle parameters and tire characteristics often have uncertainties and dynamic changes, making it difficult for the model to be completely accurate. In addition, Kalman filters have limited performance when dealing with highly nonlinear systems and are difficult to cope with complex vehicle dynamic behavior, thus affecting the accuracy and real-time performance of sideslip angle estimation. (2) Based on prior knowledge of tire parameters and tire-road friction coefficient from onboard sensors for estimation. This method relies on accurate modeling of tire parameters and friction coefficient, but in practical applications, these parameters are affected by various factors, such as road conditions, weather changes, and tire wear, making it difficult to obtain them accurately in real time. Meanwhile, the dynamic change of the friction coefficient makes the prior knowledge-based estimation method insufficient in adaptability and robustness under different working conditions, resulting in the instability and reduced reliability of the estimation results. (3) Prediction method based on deep learning. By establishing a mapping observation model between the sideslip angle and vehicle kinematics, the powerful nonlinear modeling capability and adaptive characteristics of deep learning are used to automatically extract effective features from a large amount of vehicle sensor data. This not only improves the accuracy and robustness of sideslip angle estimation, but also enhances the system's adaptability in complex driving environments, overcoming the limitations of traditional methods in parameter dependence and nonlinear processing.

[0003] While deep learning-based prediction methods have significantly improved the accuracy and robustness of tire sideslip angle estimation compared to existing approaches, there is still room for further optimization. Specifically, traditional convolutional neural networks (CNNs) excel at extracting spatial features, effectively capturing local patterns in vehicle motion data; while Transformer models have a clear advantage in handling long-range dependencies and temporal features, providing a better understanding of global information about vehicle dynamics. However, using either a CNN or a Transformer model alone may not fully leverage the advantages of both, leading to limitations in feature extraction and temporal modeling in complex driving environments. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent tire slip angle estimation method based on CNN-Transformer. The CNN neural network uses four convolutional layers to extract low-level features that are ignored or cannot be captured by the Transformer model, compensating for the limited ability of the Transformer encoder to mine information from multivariate datasets. The extracted spatial features are input into the Transformer module for temporal feature modeling. The position encoding part of the Transformer uses learnable encoding, which is more effective when the input sequence length is short and fixed. In addition, a multi-head attention mechanism is introduced between the four encoding layers to extract shallow information in the encoded sequence, thereby further refining the information learned by different encoders. Then, through a four-layer decoder, the dependencies of these spatial features in the temporal series are analyzed to ensure the accuracy of the estimated slip angle, effectively improving the prediction accuracy of the model and enhancing its adaptability to dynamic changes in complex driving environments.

[0005] To achieve the above objectives, the technical solution of the present invention is: an intelligent tire slip angle estimation method based on CNN-Transformer, comprising the following steps:

[0006] Step S1: In the data acquisition and preprocessing stage, a fully loaded tire is selected, and an intelligent tire system is built based on a triaxial accelerometer. The acceleration signal of the tire is acquired under the side slip condition, covering the changing conditions of load, speed, side slip angle and slip ratio. The acquired triaxial acceleration signal is processed by low-pass filtering, and the processed data is divided into training set A and test set B according to the ratio.

[0007] Step S2: In the feature extraction stage, a multi-layer convolutional neural network feature extractor is designed to extract discriminative features from the preprocessed triaxial acceleration signal.

[0008] Step S3: In the feature aggregation stage, a feature enhancement and aggregation network is constructed using a multi-head cross-attention module. The attention mechanism filters out interference information and highlights features related to the side slip angle to improve the discriminative ability of the features.

[0009] Step S4: Input the data from test set B into the trained neural network model to predict the slip angle of the corresponding tire, and evaluate its prediction performance through the model output.

[0010] In one embodiment of the present invention, step S1 involves collecting experimental data using an intelligent tire system and the MTS Flat-Trac tire platform. The side slip conditions cover different loads, speeds, side slip angles, and slip ratios. Specifically, this includes the dynamic response signals of the tire under different side slip angles and slip ratios under fixed loads and speeds. The experimental data is recorded using a triaxial accelerometer at a sampling frequency of 1600Hz. The collected signal is in the form X(t) = {x(t), y(t), z(t)}, where x(t), y(t), and z(t) represent the acceleration signals along the three axes, and t is the time sampling point.

[0011] In one embodiment of the present invention, in step S1, to improve data quality, the original acceleration signal is subjected to a 200Hz low-pass filter to remove high-frequency noise and low-frequency drift. The formula for the filtered signal is:

[0012]

[0013] Where f represents the frequency of the input signal, which affects the degree of signal attenuation under this filter. c =200Hz is the cutoff frequency of the filter.

[0014] The processed signal is further normalized using the Z-score standardization method to eliminate dimensional differences. The standardization formula is as follows:

[0015]

[0016] Here, μ and σ represent the mean and standard deviation of the acceleration signal, respectively. After the above processing, the signal has zero mean and unit variance, ensuring the consistency of input and data quality for subsequent model training.

[0017] In one embodiment of the present invention, in step S1, the data is divided in an 8:2 ratio, that is, 80% of the data is used for training set A and 20% of the data is used for test set B. The division formula is as follows:

[0018] A∪B=X,

[0019]

[0020] Where X is the triaxial acceleration signal after low-pass filtering.

[0021] In one embodiment of the present invention, step S2, the multi-layer convolutional neural network feature extractor includes:

[0022] 1) The first convolutional layer consists of 32 convolutional kernels, with a kernel size of 3, a stride of 1, padding of 1, and an activation function of ReLU;

[0023] 2) The second convolutional layer consists of 64 convolutional kernels, with a kernel size of 3, a stride of 1, padding of 1, and an activation function of ReLU;

[0024] 3) Max pooling layer, with a pooling window size of 2;

[0025] 4) The third convolutional layer consists of 128 convolutional kernels, with a kernel size of 3, a stride of 1, padding of 1, and the activation function being ReLU;

[0026] 5) The fourth convolutional layer includes feature_dim convolutional kernels, with a kernel size of 3, a stride of 1, padding of 1, and an activation function of ReLU; where feature_dim is the feature dimension, which can be adjusted according to actual needs.

[0027] The mathematical expression for the convolution operation is as follows:

[0028] F i =ReLU(W i ·F i-1 +b i )

[0029] Among them, F i For the output features of the i-th layer, W i and b i These are the weights and bias parameters of the convolutional layer, respectively.

[0030] In one embodiment of the present invention, the specific mathematical expression of the multi-layer convolutional neural network feature extractor structure in step S2 is as follows:

[0031] The input signal x first passes through the first convolutional layer and the ReLU activation function to obtain primary features:

[0032] f (1) =ReLU(Conv1D(x;W1))

[0033] Next, more complex features are extracted through a second convolutional layer and the ReLU activation function:

[0034] f (2) =ReLU(Conv1D(f (1) ;W2))

[0035] Then, a max pooling layer is applied for dimensionality reduction:

[0036] f (3) =MaxPool1D(f (2) )

[0037] After a third convolutional layer and the ReLU activation function, higher-level features are further extracted:

[0038] f (4) =ReLU(Conv1D(f (3) ;W3))

[0039] Finally, the feature dimensions are mapped to the preset feature_dim using a fourth convolutional layer and a ReLU activation function:

[0040] f (5) =ReLU(Conv1D(f (4) ;W4))

[0041] Where W1, W2, W3, and W4 represent the weight parameters of each convolutional layer, Conv1D(·) represents a one-dimensional convolution operation, and MaxPool1D represents a one-dimensional max pooling layer.

[0042] In one embodiment of the present invention, step S3 includes the multi-head cross-attention module:

[0043] A positional encoding is added to each time step of the input. The positional encoding is a trainable parameter of the form: P(t) = Encode(t), where t is the time step length. Positional encoding enhances the model's ability to model the temporal dependencies of time-series signals. The shape of the positional encoding is (1, seq_length, feature_dim), where seq_length is the length of the input time series and feature_dim is the feature dimension, which is adjusted according to actual needs. The positional encoding is added to the input features and then input into a multi-head cross-attention module to generate multiple queries, keys, and values. The calculation formula is as follows:

[0044] Q = X pos W Q K = X pos W K V = X pos W V

[0045] Among them, X pos It is the input feature, W Q W K W V These are the weight matrices for queries, keys, and values, respectively.

[0046] Attention is a mechanism used in deep learning models (especially Transformers) that allows the model to focus on the most relevant information while ignoring less important parts of the input data. In self-attention, different parts of the input influence each other to compute a more efficient representation. The attention weights are calculated and summed using the following formula:

[0047]

[0048] in, It is the dot product between the query and the key, d k The dimension of the attention head; the Softmax function is used to convert a set of values ​​(such as attention scores) into a probability distribution such that their sum is 1;

[0049] The outputs of multiple attention heads are concatenated and then transformed by a linear transformation matrix W. O Mapping back to the original feature dimensions yields the final aggregated feature representation:

[0050] MultiHead(Q,K,V)=Concat(head1,head2,...,head h W O

[0051] Where MultiHead(Q,K,V) represents the concatenated output of multiple attention heads, and head i This represents the output of the i-th attention head; Concat(·) concatenates the outputs of multiple heads.

[0052] In one embodiment of the present invention, in step S3, the pooling operation is average pooling, and the specific formula is as follows:

[0053]

[0054] Where X is the feature vector, L is the length of the time series, and X i Let be the feature vector corresponding to the i-th time step.

[0055] In one embodiment of the present invention, during the training of the neural network model, the mean squared error (MSE) is used as the loss function to optimize the parameters of the CNN-Transformer model. The formula for calculating MSE is:

[0056]

[0057] in, It is the true sideslip angle of the i-th sample. is the predicted lateral tilt angle of the i-th sample, and N is the number of samples.

[0058] In one embodiment of the present invention, in step S4, the neural network model is a regression network used to map the pooled feature vector to the tire's sideslip angle, specifically including:

[0059] A fully connected layer is used to map the feature vector X to the predicted sag angle. The calculation formula is as follows:

[0060]

[0061] In one embodiment of the present invention, after training, the method can achieve high-precision estimation of tire slip angle under different slip conditions, has good generalization ability and real-time performance, and is suitable for dynamic monitoring and control of intelligent tire systems.

[0062] Compared with the prior art, the present invention has the following beneficial effects:

[0063] (1) This invention addresses the limitations of traditional Transformer models in tire slip angle estimation tasks, specifically the encoder limitations and the difficulty in effectively capturing useful information between features. It proposes a hybrid model combining Convolutional Neural Networks (CNNs) and Transformers. By integrating the advantages of CNNs in learning local and related features with the powerful capabilities of Transformers in long-range representation learning, this invention can more comprehensively extract and aggregate key features from tire acceleration signals, thereby significantly improving the accuracy of slip angle prediction and outperforming existing single-model or shallow network methods.

[0064] (2) In the feature extraction stage, this invention designs a convolutional neural network containing four one-dimensional convolutional layers to extract deep, low-level features that traditional Transformer models often ignore or fail to capture. These one-dimensional convolutional layers can effectively capture subtle changes and local patterns in acceleration signals, supplementing the shortcomings of Transformer in processing complex time-series data. Furthermore, by introducing learnable positional encoding, this invention solves the problem of positional information loss in Transformer models when processing input sequences. The application of positional encoding enables the model to explicitly learn and retain the timeliness of time-series data, further improving the temporal correlation of features and the overall predictive ability of the model. Attached Figure Description

[0065] Figure 1 This is a diagram showing the arrangement of the intelligent tire sensors of this invention;

[0066] Figure 2 This is a flowchart illustrating the method implementation of an embodiment of the present invention;

[0067] Figure 3 The result of processing the triaxial acceleration dataset collected in this invention (200Hz low-pass filtering);

[0068] Figure 4 This is a schematic diagram of the CNN-Transforms network proposed in this invention;

[0069] Figure 5 The relationship between MSE and training epochs in neural networks;

[0070] Figure 6 This is the prediction result of the neural network model of this invention on the test set. Detailed Implementation

[0071] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0072] 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.

[0073] 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.

[0074] like Figure 1 The diagram illustrates the specific arrangement of the intelligent tire sensors in this invention. The system comprises three key components: a triaxial accelerometer, a data acquisition module, and a data processing unit. The sensors are mounted along the three axes of the tire to collect acceleration signals in real time under different slip angles.

[0075] like Figure 2 As shown, the method implementation flow of this invention includes four main steps: data acquisition and preprocessing, feature extraction, feature aggregation, and model evaluation. The flowchart clearly illustrates the logical relationships between each step.

[0076] Step S1: Data Acquisition and Preprocessing. Experimental data was acquired using the intelligent tire system and the Measurement and Testing Simulation (MTS) Flat-Trac tire platform. Triaxial acceleration signals of the tire were collected for different sideslip conditions. The acquired signals included dynamic data under different loads, speeds, sideslip angles, and slip ratios, with a data sampling frequency of 1600Hz. The experimental data underwent low-pass filtering (200Hz cutoff frequency) to remove noise signals. 108 features were extracted from the data and standardized to ensure the stability and accuracy of the model training.

[0077] Step S2: Feature Extraction. A convolutional neural network (CNN) model is used to extract features from the preprocessed triaxial acceleration signal, extracting local time-series features.

[0078] Step S3: Feature Aggregation. The encoder module based on the Transformer model further extracts global features. The Transformer uses a multi-head attention mechanism to aggregate the local features extracted by the CNN with global contextual information to highlight key features related to the side slip angle. Simultaneously, the decoder module further optimizes feature representation, improving the model's ability to discriminate features and its robustness to noise.

[0079] Step S4: Model Training and Prediction. An end-to-end CNN-Transformer neural network model is constructed by mapping the aggregated features to the sideslip angle value through a fully connected layer. Mean squared error (MSE) is used as the loss function to train the model on the training set data, progressively optimizing the network parameters. After training, the test set data is input into the model to predict the tire sideslip angle and verify the model's prediction accuracy.

[0080] Figure 3 The triaxial acceleration signal curves after 200Hz low-pass filtering in this embodiment are shown. The filtered signal effectively removes high-frequency noise and low-frequency drift, retains effective information related to sideslip angle estimation, and improves the signal-to-noise ratio.

[0081] Figure 4 This invention presents the CNN-Transformer network structure. The network consists of four convolutional layers for feature extraction, a multi-head cross-attention module for feature aggregation, and a fully connected layer for side slip angle regression prediction. This structure combines the local feature extraction capability of CNNs with the global feature aggregation capability of Transformers, thereby improving the accuracy of side slip angle estimation.

[0082] Figure 5The graph shows the trend of mean squared error (MSE) during training as the number of training epochs increases. As can be seen, the MSE gradually decreases with increasing training epochs, indicating that the model training gradually converges and the prediction accuracy continuously improves.

[0083] Figure 6 The graph shows a comparison between the model's predictions and actual sideslip angle values ​​on test set B. In the graph, blue dots represent the actual measured tire tread depth values, and pink dots represent the model's predictions. As can be observed from the graph, the distributions of the predicted and actual values ​​are quite close, indicating that the model can effectively estimate the tire sideslip angle and has good predictive performance and generalization ability.

[0084] In this embodiment, the specific evaluation index results of the model are quantitatively analyzed as shown in Table 1:

[0085] Table 1

[0086]

[0087] The following section will elaborate on the relevant aspects of this method.

[0088] In step S1, a tire model and its related parameters that meet the research requirements are selected as the experimental subject. Specifically, a Prinx tire with a specification of 245 / 45R19 and a tread pattern of XNEX SPORT EV is selected, and the test is conducted under full load conditions. Based on this tire information, an intelligent tire system is built. The core component of the system is a triaxial accelerometer, which is used to monitor the dynamic response of the tire under different working conditions in real time.

[0089] During the data acquisition phase, tire acceleration signals were collected for different lateral slip conditions. These conditions encompassed variations in load, speed, slip angle, and slip ratio. The intelligent tire system collected triaxial acceleration signals from the tire under these different conditions at a sampling frequency of 1600Hz, yielding raw data in the following format:

[0090] X(t)={x(t),y(t),z(t)}

[0091] Where x(t), y(t), and z(t) represent acceleration signals along the three axes, and t is the time sampling point.

[0092] To improve data quality and model training effectiveness, the acquired raw acceleration signals are preprocessed. First, a 200Hz low-pass filter is used to filter the signal to remove high-frequency noise and low-frequency drift. The filter's transfer function is defined as:

[0093]

[0094] The above filtering process ensures that effective frequency components are retained in the signal, improving the signal-to-noise ratio. Next, the filtered acceleration signal is standardized using the Z-score normalization method to eliminate the influence of different units and magnitudes on model training. The normalization formula is as follows:

[0095]

[0096] Here, μ and σ are the mean and standard deviation of the acceleration signal, respectively. After standardization, the signal has zero mean and unit variance, which helps to accelerate the convergence speed of the model and improve training performance.

[0097] Next, key features are extracted from the preprocessed acceleration signal to characterize the dynamic properties under sideslip conditions. Common features include time-domain features (such as mean, variance, peak value, etc.) and frequency-domain features (such as dominant frequency, spectral energy, etc.), which will be used for subsequent model training and sideslip angle estimation. Finally, the processed dataset is divided into training and test sets proportionally to ensure the model's generalization ability. The specific division ratio is 8:2, that is, 80% of the data is used for training set A and 20% of the data is used for test set B. The division formula is as follows:

[0098] A∪B=X,

[0099]

[0100] Ultimately, the constructed dataset contains acceleration signals in three dimensions, with each dimension corresponding to an independent training set A and a test set B. The dataset is represented as follows:

[0101] X = {X1, X2, X3}

[0102] Among them, X i (i = 1, 2, 3) correspond to acceleration signals along three axes, each X i Divided into training set A i and test set B i The specific ratio is 8:2:

[0103] X i ={A i B i},|A i |:|B i |=8:2

[0104] In step S2, a convolutional neural network (CNN) is used to extract features, performing shallow feature extraction on the preprocessed triaxial acceleration signal. Specifically, a multi-layer CNN structure is designed to effectively capture local patterns and complex features in the time-series data. This CNN feature extractor includes the following key components:

[0105] First, the input data is the triaxial acceleration signal preprocessed in step S1, with a shape of (C, L), where C = 3 represents the number of input channels (corresponding to the three axial accelerations), and L represents the time series length. The input signal first passes through the first convolutional layer, which contains 32 convolutional kernels with a kernel size of 3, a stride of 1, padding of 1, and uses the ReLU activation function. The main function of this layer is to extract primary local features while maintaining the time series length.

[0106] The signal then passes through a second convolutional layer containing 64 kernels, each with a kernel size of 3, a stride of 1, padding of 1, and a ReLU activation function. This second convolutional layer further extracts more complex local features, enhancing their expressive power. Following this, a max-pooling layer with a pooling window size of 2 is applied to reduce feature dimensionality, decrease computational cost, and preserve salient feature information.

[0107] The features processed by the pooling layer continue through the third convolutional layer, which contains 128 kernels with a kernel size of 3, a stride of 1, padding of 1, and uses the ReLU activation function. This third convolutional layer extracts higher-level features, further enhancing their expressive power. Finally, the features pass through the fourth convolutional layer, where the number of kernels is set to `feature_dim` (the feature dimension, which can be adjusted according to actual needs), the kernel size is 3, the stride is 1, padding is 1, and the ReLU activation function is used. The main function of the fourth convolutional layer is to map the feature dimension to a specified dimension, preparing for the subsequent Transformer module.

[0108] In this invention, the core operations in the feature extraction stage include multi-layer convolution and pooling operations. The output of each layer undergoes a non-linear transformation using the ReLU activation function, effectively capturing different levels of features in the time series data. The specific mathematical expression is as follows:

[0109] The input signal x first passes through the first convolutional layer and the ReLU activation function to obtain primary features:

[0110] f (1) =ReLU(Conv1D(x;W1))

[0111] Next, more complex features are extracted through a second convolutional layer and the ReLU activation function:

[0112] f (2) =ReLU(Conv1D(f (1) ;W2))

[0113] Then, a max pooling layer is applied for dimensionality reduction:

[0114] f (3) =MaxPool1D(f (2) )

[0115] After a third convolutional layer and the ReLU activation function, higher-level features are further extracted:

[0116] f (4) =ReLU(Conv1D(f (3) ;W3))

[0117] Finally, the feature dimensions are mapped to the preset feature_dim using a fourth convolutional layer and a ReLU activation function:

[0118] f (5) =ReLU(Conv1D(f (4) ;W4))

[0119] Where W1, W2, W3, and W4 represent the weight parameters of each convolutional layer.

[0120] In step S3, a multi-head cross-attention mechanism is used to enhance and aggregate the features extracted from the convolutional neural network (CNN). Through this process, the model can better focus on features related to the lateral slip angle estimation, highlight important information, and filter out interfering factors, thereby improving prediction accuracy. The specific implementation process includes the following aspects:

[0121] First, the input data comes from the convolutional feature output in step S2, with a shape of (batch_size, seq_length, feature_dim), where batch_size is the batch size, seq_length is the length of the input time series, and feature_dim is the feature dimension output by the convolutional network. These feature tensors will be used as input to the multi-head cross-attention module for further processing.

[0122] To enable the model to understand the temporal information in the input data, we add a positional encoding to each time step of the input. The positional encoding is a learned, trainable parameter with the shape (1, seq_length, feature_dim), representing the position of each time step. By adding the positional encoding to the input features, the model can adjust the weights of the features at each time step based on the positional information in the sequence, thereby enhancing the model's ability to model temporal dependencies.

[0123] Next, input feature X pos After positional encoding, the input features are fed into the multi-head cross-attention module for further processing. In this mechanism, the input features are mapped to multiple queries, keys, and values. The formulas for calculating queries, keys, and values ​​are as follows:

[0124] Q = X pos W Q K = X pos W K V = X pos W V

[0125] Among them, W Q W K W V These are the weight matrices for the query, key, and value, with shape (feature_dim, d...). k ), where d k This refers to the dimension of each attention head. Through this step, the model generates multiple attention heads from the input features, each of which can capture information from different subspaces of the input data.

[0126] The core idea of ​​multi-head cross-attention is to calculate the dot product between the query and the key to obtain the attention weights of the features at each time step, and then perform a weighted sum of these weights to focus on the important feature information. The specific calculation process can be represented as follows:

[0127]

[0128] in, The process involves calculating the dot product between the query and the key, using softmax to normalize and obtain attention weights, and finally applying these attention weights to the values ​​to obtain a weighted feature representation. This process is performed independently for each attention head, resulting in weighted features for different subspaces.

[0129] To fully extract the feature information from different subspaces, we concatenate the outputs of multiple heads and then apply a linear transformation matrix W. O The concatenated features are mapped back to the original feature dimensions to obtain the final aggregated feature representation. The formula is as follows:

[0130] MultiHead(Q,K,V)=Concat(head1,head2,...,head h W O

[0131] Through this multi-head attention mechanism, the model can focus on key information in the input features from different perspectives, further enhancing the expressive power of the features.

[0132] Features processed by the multi-head attention module, while containing rich contextual information, may still have high dimensionality. Therefore, we use pooling operations to reduce the dimensionality of these features. Pooling operations compress the features in the time series into a fixed-dimensional vector by performing average pooling, thereby reducing the dimensionality of the time steps. Through average pooling, we obtain a global, more concise feature representation, as shown in the formula:

[0133]

[0134] Where L is the length of the time series, and Attention_Output i It is the feature vector corresponding to the i-th time step. Pooling operation helps the model shift its focus from local time steps to global feature representation, further improving the generalization ability of the features.

[0135] Finally, the feature vector Final_Output obtained through pooling is fed into the subsequent regression network for prediction of the skid angle. The goal of the regression network is to map this feature vector to the target value—the skid angle of the tire. The specific regression process is completed through a fully connected layer, and the output skid angle value will be used as the final prediction result.

[0136] In step S4, the preprocessed sample data is input into the trained CNN-Transformer model, and the consistency between the actual tire slip angle and the predicted result is compared to verify the model's effectiveness. The test set B data, after undergoing the same preprocessing as the training set, is input into the trained model for forward propagation to obtain the corresponding slip angle predictions. These predictions reflect the model's ability to extract features from the input acceleration signal and estimate the slip angle.

[0137] By comparing the predicted results with the actual values, we use several regression evaluation metrics to quantify the model's predictive accuracy. First, the mean squared error (MSE) is one of the most commonly used metrics for measuring predictive accuracy, calculated as follows:

[0138]

[0139] in, It is the true sideslip angle of the i-th sample. is the predicted lateral slip angle of the i-th sample, and N is the number of samples. The smaller the MSE value, the closer the prediction result is to the true value.

[0140] Secondly, the root mean square error (RMSE) is the square root of the mean square error (MSE), and its calculation formula is as follows:

[0141]

[0142] A smaller RMSE indicates higher prediction accuracy of the model.

[0143] Mean Absolute Error (MAE) measures the absolute error between the predicted and actual values. The formula for calculating it is:

[0144]

[0145] The smaller the MAE value, the smaller the prediction bias and the better the model performance.

[0146] Finally, the coefficient of determination (R²) 2 The formula for measuring the model's fit is:

[0147]

[0148] in, It is the mean of the true sideslip angle. R 2 It measures the strength of the relationship between the model's predictions and the actual data. If R... 2 A value of 1 indicates that the model perfectly fits the data, and all real data variations (i.e., changes) have been successfully predicted by the model.

[0149] 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.

[0150] 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.

[0151] 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.

[0152] 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.

[0153] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present 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 CNN-Transformer-based intelligent tire side slip angle estimation method, characterized in that, Comprising the following steps: Step S1, in the data acquisition and preprocessing stage, select the full load tire, build an intelligent tire system based on three-axis acceleration sensor, collect the acceleration signal of the tire under the condition of side slip, covering the change conditions including load, speed, side slip angle and slip ratio; apply low-pass filter to the collected three-axis acceleration signal for signal processing, and divide the processed data into training set A and test set B according to the proportion; Step S2, in the feature extraction stage, design a multi-layer convolutional neural network feature extractor to extract discriminative features from the preprocessed three-axis acceleration signal; the multi-layer convolutional neural network feature extractor comprises: 1) a first convolutional layer comprising 32 convolutional kernels, the convolutional kernel size is 3, the stride is 1, the padding is 1, and the activation function is ReLU; 2) a second convolutional layer comprising 64 convolutional kernels, the convolutional kernel size is 3, the stride is 1, the padding is 1, and the activation function is ReLU; 3) a maximum pooling layer with a pooling window size of 2; 4) a third convolutional layer comprising 128 convolutional kernels, the convolutional kernel size is 3, the stride is 1, the padding is 1, and the activation function is ReLU; 5) a fourth convolutional layer comprising feature_dim convolutional kernels, the convolutional kernel size is 3, the stride is 1, the padding is 1, and the activation function is ReLU; wherein feature_dim is the feature dimension, which can be adjusted according to actual requirements; Step S3, in the feature aggregation stage, a multi-head cross attention module is used to construct a feature enhancement and aggregation network, which filters out interference information through attention mechanism and highlights the features related to side slip angle to improve the discriminability of the features; Step S4, input the data in test set B into the trained neural network model to predict the side slip angle of the corresponding tire, and evaluate its prediction performance through model output.

2. The CNN-Transformer-based intelligent tire side slip angle estimation method according to claim 1, wherein In step S1, the sampling frequency of the three-axis acceleration sensor is 1600Hz, which is used to collect the three-axis acceleration signal of the tire under different working conditions, and the signal form is X(t)={x(t),y(t),z(t)}, where t is the time sampling point.

3. The CNN-Transformer-based intelligent tire side slip angle estimation method according to claim 1, wherein In step S1, the low-pass filter uses a low-pass filter with a frequency of 200Hz, and its transfer function is defined as: Where f represents the frequency of the input signal, affecting the degree of attenuation of the signal under the low-pass filter, f c = 200 Hz is the cutoff frequency of the filter.

4. The CNN-Transformer-based intelligent tire side slip angle estimation method according to claim 1, wherein, In step S1, the data division is performed according to the proportion of 8:2, i.e. 80% of the data is used for training set A and 20% of the data is used for test set B, and the division formula is: A U B = X', where X' is the three-axis acceleration signal after low-pass filter processing.

5. The CNN-Transformer-based intelligent tire side slip angle estimation method according to claim 1, wherein, In step S2, the specific mathematical expression of the multi-layer convolutional neural network feature extractor structure is as follows: The input signal x is first processed by the first convolutional layer and the ReLU activation function to obtain the primary features: f (1) = ReLU(Conv1D(x; W1)) Then, more complex features are extracted through the second convolutional layer and the ReLU activation function: f (2) = ReLU(Conv1D(f (1) ; W2)) Then, dimensionality reduction is performed by applying the maximum pooling layer: f (3) = MaxPool1D(f (2) ) After the third convolutional layer and the ReLU activation function, high-level features are further extracted: f (4) = ReLU(Conv1D(f (3) ; W3)) Finally, the feature dimension is mapped to the preset feature_dim through the fourth convolutional layer and the ReLU activation function: f (5) = ReLU(Conv1D(f (4) ; W4)) Wherein, W1, W2, W3, W4 represent the weight parameters of each convolution layer respectively, Conv1D(·) represents one-dimensional convolution operation, and MaxPool1D represents one-dimensional maximum pooling layer.

6. The CNN-Transformer-based intelligent tire side slip angle estimation method according to claim 1, wherein In step S3, the multi-head cross attention module comprises: A position encoding is added for each time step of the input, the position encoding is a trainable parameter, and the shape is (1, seq_length, feature_dim). The seq_length is the length of the input time sequence, and the feature_dim is the feature dimension, which is adjusted according to actual requirements. After the position encoding is added to the input feature, the position encoding is input into the multi-head cross attention module to generate a plurality of queries Query, keys Key and values Value, and the calculation formula is: Q = X pos W Q K = X pos W K V = X pos W V where X pos is the input feature, W Q , W K , and W V are the weight matrices for queries, keys, and values, respectively; The attention weight is calculated and weighted summation is performed, and the specific formula is: wherein, is a dot product computation between the query and the key, d k is the dimension of the attention head, Attention is the attention mechanism; the Softmax function is used to convert a set of numerical values into a probability distribution, making their sum equal to 1; After concatenating the outputs of multiple attention heads, the linear transformation matrix W O is applied to map back to the original feature dimension to get the final aggregated feature representation: MultiHead(Q, K, V) = Concat(head1, head2,..., head h )W O wherein MultiHead(Q, K, V) represents the output concatenation result of multiple attention heads, head i represents the output of the i-th attention head; and Concat(·) concatenates the outputs of multiple attention heads.

7. The CNN-Transformer-based intelligent tire side slip angle estimation method according to claim 1, wherein In step S3, the pooling operation is average pooling, and the specific formula is: where X is the feature vector, L is the length of the time series, X i is the feature vector corresponding to the i-th time step.

8. The CNN-Transformer-based intelligent tire side slip angle estimation method according to claim 1, wherein In step S4, the neural network model is a regression network for mapping the feature vector after the pooling to the tire side slip angle, and specifically comprises: one fully connected layer for mapping the feature vector X to the predicted side slip angle The formula is:

9. The CNN-Transformer-based intelligent tire side slip angle estimation method according to claim 1, wherein After training, the method can realize high-precision estimation of the tire side slip angle under different side slip conditions, has good generalization ability and real-time performance, and is suitable for dynamic monitoring and control of an intelligent tire system.

Images