A data-driven deep learning vehicle motion state estimation method and system

Abstract

The application relates to the technical field of vehicle dynamics control, and discloses a data-driven deep learning vehicle motion state estimation method and system, which comprises the following steps: inputting collected vehicle multi-source sensor data into a pre-constructed TCN-SA-GRU model; and using a whale optimization algorithm to optimize the hyperparameters of the TCN-SA-GRU model, wherein the TCN-SA-GRU model integrates the parallel computing and multi-scale time feature extraction capability of a time convolution network (TCN), the feature extraction capability of a self-attention (SA) mechanism and the time sequence processing capability of a gated recurrent unit (GRU), processes the vehicle sensor data, and obtains vehicle motion state estimation results. The application can realize dynamic and accurate estimation of key motion states such as a vehicle side slip angle and a yaw angular velocity.

Description

Technical Field

[0001] This invention relates to a data-driven deep learning-based method and system for estimating vehicle motion states, belonging to the field of vehicle dynamics control technology. Background Technology

[0002] Real-time and accurate estimation of vehicle motion states, especially key parameters such as vehicle slip angle (VSA) and yaw rate, is a core prerequisite for achieving advanced stability control and ensuring driving safety in autonomous driving systems. However, in complex and ever-changing driving environments, key parameters such as VSA are difficult to obtain directly through conventional low-cost sensors, thus requiring advanced state estimation algorithms for estimation.

[0003] Currently, existing methods for estimating vehicle motion states can be mainly divided into model-based methods and data-driven machine learning methods.

[0004] Model-based methods (such as Kalman filters and their variants, sliding mode observers, etc.) rely on high-precision vehicle dynamics and tire models. However, vehicle dynamics systems are highly nonlinear and time-varying. When the vehicle enters the nonlinear region (such as extreme conditions) or the tire-road conditions change, the performance of these methods deteriorates significantly due to model mismatch, resulting in insufficient robustness.

[0005] Traditional machine learning methods (such as shallow neural networks), while avoiding complex mechanistic modeling, often suffer from insufficient feature extraction capabilities and an inability to effectively capture long-term temporal dependencies when processing vehicle dynamic data with strong temporal correlations. This limits their model performance and makes it difficult to meet the requirements of vehicle control systems in terms of computational efficiency and real-time performance.

[0006] With the development of deep learning technology, especially the emergence of advanced temporal neural network structures such as Temporal Convolutional Networks (TCNs) and Gated Recurrent Units (GRUs), powerful new tools have been provided for feature mining and modeling of high-dimensional time series data. These new technologies have shown significant advantages in parallel computing efficiency, capturing long-range dependencies, and nonlinear mapping capabilities. Therefore, it is urgent to utilize these advanced neural network technologies to construct a novel, data-driven deep learning estimation framework to overcome the limitations of traditional methods and achieve dynamic, accurate, and efficient estimation of vehicle motion states. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide a solution.

[0008] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution.

[0009] In a first aspect, the present invention discloses a data-driven deep learning-based method for estimating the motion state of a vehicle, comprising the following steps: Acquire multi-source sensor data reflecting the motion state of the target vehicle; The multi-source sensor data is input into a pre-trained TCN-SA-GRU hybrid deep learning model to obtain the vehicle motion state estimation result. The training process of the TCN-SA-GRU hybrid deep learning model includes: A standardized training dataset is constructed using a co-simulation environment, and the standardized training dataset contains multi-source sensor data. A TCN-SA-GRU hybrid deep learning model is constructed, which is a deep learning model that sequentially integrates the parallel multi-scale feature extraction capability of Temporal Convolutional Network (TCN), the dynamic feature weighting capability of Self-Attention (SA) mechanism, and the long temporal dependency processing capability of Gated Recurrent Unit (GRU). Using the standardized training data as input, and taking the TCN-SA-GRU hybrid deep learning model as the object, the hyperparameters of the TCN-SA-GRU hybrid deep learning model are globally optimized to minimize the estimation error, thereby training and generating the optimal model and obtaining the trained TCN-SA-GRU hybrid deep learning model.

[0010] Furthermore, the multi-source sensor data includes: four-wheel wheel speed, front wheel steering angle, longitudinal speed, lateral speed, longitudinal acceleration, and lateral acceleration; the vehicle motion state estimation results include vehicle sideslip angle (VSA) and yaw rate.

[0011] Furthermore, the processing procedure of the TCN-SA-GRU hybrid deep learning model includes: Multi-source sensor data is input into a temporal convolutional network (TCN) to capture time-series dependencies, thereby obtaining the extracted features of the original multi-source sensor data. Based on the extracted features of the original multi-source sensor data, the correlation between different sensor signals is calculated and weights are assigned through a self-attention (SA) mechanism to obtain a weighted signal sequence. The weighted signal sequence is input into GRU for deep temporal feature mining to obtain comprehensive features containing long-term and short-term dependencies. The motion state of the target vehicle is estimated based on the comprehensive features containing long-term and short-term dependencies.

[0012] Furthermore, the Whale Optimization Algorithm (WOA) is used to globally optimize the hyperparameters of the TCN-SA-GRU hybrid deep learning model, wherein the hyperparameters include the number of GRU neurons, the learning rate, and the dropout rate.

[0013] Furthermore, the formula for the gated recurrent unit (GRU) is: ; in, and These represent the forward and backward GRUs at time steps, respectively. t The output; x t In time step t The input vector; c t-1 and c t+1 These are the cell states at the previous and next time steps, respectively. h t-1 It is the forward hidden state, representing the output of the forward GRU in the previous time step; h t+1 This is the reverse hidden state, representing the output of the reverse GRU at the next time step; H t This indicates that GRU is at time step t The final output.

[0014] Furthermore, the formula for the self-attention (SA) mechanism is as follows: ; in, Z Given the input sequence, W Q , W K , W V It is a parameter matrix. Q , K , V These are query, key, and value matrices, respectively. d k It is the dimension of the key vector. Attention (·) represents the final output of the self-attention mechanism (i.e., the feature matrix after attention weighting). softmax (·) denotes the normalized exponential function, used to calculate attention weights. T This indicates the matrix transpose.

[0015] Secondly, this invention discloses a data-driven deep learning-based vehicle motion state estimation system, comprising: The data acquisition module is configured to collect data from multiple vehicle sensors in real time. The model processing module is configured to run a pre-built TCN-SA-GRU model. The hyperparameters of the TCN-SA-GRU model are pre-optimized using the Whale Optimized Word of Arithmetic (WOA) algorithm. The TCN-SA-GRU model includes: a Temporal Convolutional Network (TCN) layer, configured to capture the temporal dependencies of the sensor data; a Self-Attention (SA) mechanism layer, configured to calculate the correlation between different signals and assign weights; and a Gated Recurrent Unit (GRU) layer, configured to perform deep temporal feature mining on the weighted signal sequence. The result output module is configured to output the vehicle motion state estimation results after processing by the GRU layer.

[0016] Thirdly, the present invention discloses a computer-readable storage medium storing a computer program thereon, characterized in that: when the computer program is executed by a processor, it implements the steps of the above-mentioned data-driven deep learning vehicle motion state estimation method.

[0017] Fourthly, the present invention discloses a computer device, characterized in that it comprises: Memory, used to store computer programs; A processor is used to execute the computer program to implement the steps of the data-driven deep learning-based vehicle motion state estimation method described above.

[0018] Fifthly, the present invention discloses a computer program product, comprising a computer program, characterized in that: when the computer program is executed by a processor, it implements the steps of the above-described data-driven deep learning vehicle motion state estimation method.

[0019] The beneficial effects achieved by this invention are as follows: (1) This invention proposes a TCN-SA-GRU model, which integrates temporal convolutional network, self-attention mechanism and bidirectional GRU, and utilizes the parallel computing capability of TCN and the dynamic attention of SA mechanism to key features, thereby significantly improving feature extraction capability and estimation accuracy.

[0020] (2) The present invention uses the whale optimization algorithm to automatically optimize the model hyperparameters. Compared with the traditional parameter tuning method, it effectively reduces the training time and further improves the accuracy of estimation.

[0021] (3) The present invention trains the model by constructing a diverse dataset covering a variety of typical traffic scenarios and operating conditions in a joint simulation environment, which makes the model more robust to different road surface adhesion coefficients and vehicle speed changes, and can adapt to more complex and dynamic driving environments. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a schematic diagram of the vehicle kinematics model in an embodiment of the present invention; Figure 3 This is a schematic diagram of the overall architecture of WOA-TCN-SA-GRU in an embodiment of the present invention; Figure 4 Figures (a)-(e) are comparison charts of the estimation results of VSA and yaw rate under different operating conditions in the embodiments of the present invention. Detailed Implementation

[0023] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0024] Example 1: This example introduces a data-driven deep learning-based method for estimating vehicle motion states, such as... Figure 1 and Figure 3 As shown, it includes: Step S1: Data Acquisition and Construction. Using a co-simulation environment, a diverse and representative dataset covering various typical traffic scenarios and operating conditions is constructed, providing a robust foundation for model training and evaluation.

[0025] Step S2: Model Construction and Optimization. Construct the TCN-SA-GRU model architecture. Use the Whale Optimization Algorithm (WOA) to automatically optimize the model hyperparameters, specifically the number of GRU neurons, learning rate, and dropout rate.

[0026] Step S3: Model Training and Input. The model optimized in Step S2 is trained and fine-tuned using the dataset collected in Step S1. During the estimation phase, real-time vehicle sensor data is used as model input, including four-wheel wheel speed, front wheel steering angle, longitudinal speed, lateral speed, longitudinal acceleration, and lateral acceleration.

[0027] Step S4: State Estimation and Output. The input data sequence is processed sequentially through the TCN layer, the SA mechanism layer, and the GRU layer in the model. The TCN network further captures time-series dependencies and utilizes residual connections to improve training efficiency. The SA mechanism is applied to the initial processing of the signal sequence, calculating the correlation between signals from different sensors and assigning appropriate weights. The weighted signal sequence is then fed into the GRU (Gated Recurrent Unit), which considers both historical and future information to deeply mine the sequence data. Finally, based on these integrated features, the algorithm accurately estimates the vehicle's motion state, outputting the vehicle sideslip angle (VSA) and yaw rate.

[0028] In this embodiment, the vehicle kinematic model is as follows: Figure 2 As shown.

[0029] In this embodiment, the model principles involved in steps S2 and S4 are as follows: (1) TCN Algorithm: Temporal Convolutional Networks (TCNs) excel in processing time-series data. Their unique network architecture combines dilated causal convolution and residual connections, achieving efficient parallel processing. The causal convolution formula is: ; in, f It's a filter. x It is input. d It is an expansion factor. k It refers to the size of the nucleus.

[0030] (2) SA Algorithm: The Self-Attention (SA) algorithm enables the model to capture long-range dependencies across locations. It calculates the correlation between the query and key matrices at each time step, transforming the value matrix to produce the output by reweighting the input sequence. Its formula is: ; in, Z Given the input sequence, W Q , W K , W V It is a parameter matrix. Q , K , V These are query, key, and value matrices, respectively. d k It is the dimension of the key vector. Attention (·) represents the final output of the self-attention mechanism (i.e., the feature matrix after attention weighting). softmax (·) denotes the normalized exponential function, used to calculate attention weights. T This indicates the matrix transpose.

[0031] (3) GRU Algorithm: The gated recurrent unit processes the input in both the forward and reverse directions, while utilizing past and future context information. Its formula is: ; in, and These represent the forward and backward GRUs at time steps, respectively. t The output; x t In time step t The input vector; ct-1 and c t+1 These are the cell states at the previous and next time steps, respectively. h t-1 It is the forward hidden state, representing the output of the forward GRU in the previous time step; h t+1 This is the reverse hidden state, representing the output of the reverse GRU at the next time step; H t This indicates that GRU is at time step t The final output.

[0032] (4) Whale Optimization (WOA) Algorithm: The Whale Optimization algorithm is a swarm intelligence method inspired by whales' foraging, grouping, and social behaviors. It combines exploration and exploitation mechanisms, demonstrating strong global search capabilities and convergence in complex optimization problems. Its mathematical model is as follows: ; in, Indicates the first i Only whales are iterating t The position at that time This represents the optimal solution found in the current iteration. α Controlling the influence of the global optimal position. β Adjusting the effect of the difference term to enhance the exploration, R It is a random factor that introduces randomness.

[0033] Example 2 further illustrates the mechanism of the data-driven deep learning-based electric drive chassis health status assessment method of this invention using more specific experimental data. The experiment was conducted using co-simulation software. Specific vehicle parameters are shown in Table 1. The hyperparameter settings of the WOA-TCN-SA-GRU network are shown in Table 2.

[0034] To comprehensively train and validate the algorithm's network model, a dataset based on different driving conditions was first constructed in a co-simulation environment of a visualization simulation software. The vehicle motion state training conditions included various typical driving scenarios in both longitudinal and lateral directions. Specifically, longitudinal driving scenarios included fixed accelerator pedal opening, smooth increase of accelerator pedal opening, and constant deceleration to a stop. Lateral driving scenarios included two-lane changes, sinusoidal steering, and steady-state steering. A summary of the motion state training conditions is shown in Table 3. The model's test condition was selected as the steering wheel angle scanning frequency input condition.

[0035] This section mainly studies WOA-TCN-SA-GRU (WTSG), GRU, GRU and EKF with different road adhesion coefficients. μ ) and vehicle speed ( vThe steering wheel angle scan input shows the estimated vehicle motion state under the given conditions. For example... Figure 4 As shown in (a), (b), and (c), the predicted VSA and yaw rate are compared under the conditions of a constant speed of 60 km / h and road adhesion coefficients of 0.3, 0.6, and 0.9. Figure 4 As shown in (a), (d), and (e), the prediction results of VSA and yaw rate are compared under the conditions of a constant road adhesion coefficient of 0.9 and vehicle speeds of 40 km / h, 50 km / h, and 60 km / h. Table 4 summarizes the root mean square error (RMSE) of all algorithms under different conditions.

[0036] It can be observed that all methods successfully captured the trends of actual VSA and yaw rate changes. For all methods, at the same vehicle speed, the prediction accuracy gradually decreases as the road surface adhesion coefficient decreases. Similarly, at the same road surface adhesion coefficient, the prediction accuracy also decreases accordingly as the vehicle speed decreases. The WTSG method effectively handles the influence of road surface friction coefficient and vehicle speed changes on the prediction results. It shows the smallest error in both VSA and yaw rate estimation, especially maintaining high accuracy under low friction coefficient conditions. The WTSG method enhances feature capture capability through a self-attention mechanism, uses GRU to handle temporal information, and leverages the parallel computing advantages of TCN to better cope with dynamic changes under different vehicle speeds and road surface friction coefficients. In contrast, the estimation error of the GRU method is slightly higher than that of the WTSG method. In yaw rate estimation, GRU is close to WTSG, but in VSA estimation, the error of GRU is significantly increased compared to WTSG. The GRU method performs poorly under low road surface friction coefficients. In VSA estimation, there is a significant positive bias. In yaw rate estimation, its RMSE is an order of magnitude higher than the other two neural network methods. The EKF method exhibits significant estimation errors under low friction conditions, particularly in yaw rate estimation, where it shows marked delays and biases, failing to accurately predict vehicle dynamics. This is attributed to EKF's linear assumption about system dynamics, which leads to larger errors under nonlinear (especially low friction) conditions.

[0037] The experiments described above demonstrate that the proposed WTSG method exhibits superior robustness and accuracy in vehicle motion state estimation, significantly outperforming other comparative methods. It can more accurately characterize the complex dynamic characteristics of vehicles, and maintains stable and reliable estimation performance even under extreme conditions such as low friction coefficients.

[0038] Table 1 Specific Vehicle Parameters ; Table 2 WOA-TCN-SA-GRU Network Hyperparameter Settings ; Table 3 Summary of Training Conditions under Exercise Status ; Table 4. Root mean square error predicted by the algorithm under different operating conditions .

[0039] Example 3, based on the same inventive concept as Example 1, introduces a data-driven deep learning-based vehicle motion state estimation system, including: The data acquisition layer is configured to collect vehicle sensor data in real time. The model processing layer is configured to run a pre-built TCN-SA-GRU model, whose hyperparameters are pre-optimized using the whale optimization algorithm. The model processing layer includes: a Temporal Convolutional Network (TCN) layer configured to capture the temporal dependencies of the sensor data; a Self-Attention (SA) mechanism layer configured to calculate the correlations between different signals and assign weights; and a Gated Recurrent Unit (GRU) layer configured to perform deep temporal feature mining on the weighted signal sequence. The output layer is configured to output the vehicle motion state estimation results processed by the GRU layer. Specific implementation details for each layer are described in the above method embodiments and will not be repeated here.

[0040] Example 4, based on the same inventive concept as other examples, introduces a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the steps of the above-described data-driven deep learning-based vehicle motion state estimation method.

[0041] Example 5, based on the same inventive concept as other examples, introduces a computer device, including: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the above-described data-driven deep learning vehicle motion state estimation method.

[0042] Example 6, based on the same inventive concept as other examples, introduces a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described data-driven deep learning vehicle motion state estimation method.

[0043] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention 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.

[0044] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. 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 illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0045] 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.

[0046] 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.

[0047] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other modifications under the guidance of the present invention, and these modifications are all within the protection scope of the present invention.

Claims

1. A data-driven deep learning-based method for estimating vehicle motion state, characterized in that, Includes the following steps: Acquire multi-source sensor data reflecting the motion state of the target vehicle; The multi-source sensor data is input into a pre-trained TCN-SA-GRU hybrid deep learning model to obtain the vehicle motion state estimation result. The training process of the TCN-SA-GRU hybrid deep learning model includes: A standardized training dataset is constructed using a co-simulation environment, and the standardized training dataset contains multi-source sensor data. A TCN-SA-GRU hybrid deep learning model is constructed, which is a deep learning model that sequentially integrates the parallel multi-scale feature extraction capability of Temporal Convolutional Network (TCN), the dynamic feature weighting capability of Self-Attention (SA) mechanism, and the long temporal dependency processing capability of Gated Recurrent Unit (GRU). Using the standardized training data as input, and taking the TCN-SA-GRU hybrid deep learning model as the object, the hyperparameters of the TCN-SA-GRU hybrid deep learning model are globally optimized to minimize the estimation error, thereby training and generating the optimal model and obtaining the trained TCN-SA-GRU hybrid deep learning model.

2. The data-driven deep learning-based vehicle motion state estimation method according to claim 1, characterized in that, The multi-source sensor data includes: four-wheel wheel speed, front wheel steering angle, longitudinal speed, lateral speed, longitudinal acceleration, and lateral acceleration; the vehicle motion state estimation results include vehicle slip angle (VSA) and yaw rate.

3. The data-driven deep learning-based vehicle motion state estimation method according to claim 1, characterized in that, The processing procedure of the TCN-SA-GRU hybrid deep learning model includes: Multi-source sensor data is input into a temporal convolutional network (TCN) to capture time-series dependencies, thereby obtaining the extracted features of the original multi-source sensor data. Based on the extracted features of the original multi-source sensor data, the correlation between different sensor signals is calculated and weights are assigned through a self-attention (SA) mechanism to obtain a weighted signal sequence. The weighted signal sequence is input into GRU for deep temporal feature mining to obtain comprehensive features containing long-term and short-term dependencies. The motion state of the target vehicle is estimated based on the comprehensive features containing long-term and short-term dependencies.

4. The data-driven deep learning-based vehicle motion state estimation method according to claim 1, characterized in that, The Whale Optimization Algorithm (WOA) is used to globally optimize the hyperparameters of the TCN-SA-GRU hybrid deep learning model, where the hyperparameters include the number of GRU neurons, the learning rate, and the dropout rate.

5. The data-driven deep learning-based vehicle motion state estimation method according to claim 1, characterized in that, The formula for the gated recurrent unit (GRU) is: ； in, and These represent the forward and backward GRUs at time steps, respectively. t The output; x t In time step t The input vector; c t-1 and c t+1 These are the cell states at the previous and next time steps, respectively. h t-1 It is the forward hidden state, representing the output of the forward GRU in the previous time step; h t+1 This is the reverse hidden state, representing the output of the reverse GRU at the next time step; H t Indicates that GRU is at time step t The final output.

6. The data-driven deep learning-based vehicle motion state estimation method according to claim 1, characterized in that, The formula for the self-attention (SA) mechanism is as follows: ； in, Z The input sequence is... W Q , W K , W V It is a parameter matrix; Q , K , V These are the query, key, and value matrix, respectively. d k It is the dimension of the key vector; Attention (·) represents the final output of the self-attention mechanism; softmax (·) denotes the normalized exponential function, used to calculate attention weights; T This indicates the matrix transpose.

7. A data-driven deep learning-based vehicle motion state estimation system, characterized in that, include: The data acquisition module is configured to collect data from multiple vehicle sensors in real time. The model processing module is configured to run a pre-built TCN-SA-GRU model. The hyperparameters of the TCN-SA-GRU model are pre-optimized using the Whale Optimized Word of Arithmetic (WOA) algorithm. The TCN-SA-GRU model includes: a Temporal Convolutional Network (TCN) layer, configured to capture the temporal dependencies of the sensor data; a Self-Attention (SA) mechanism layer, configured to calculate the correlation between different signals and assign weights; and a Gated Recurrent Unit (GRU) layer, configured to perform deep temporal feature mining on the weighted signal sequence. The result output module is configured to output the vehicle motion state estimation results after processing by the GRU layer.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the data-driven deep learning vehicle motion state estimation method according to any one of claims 1 to 6.

9. A computer device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the data-driven deep learning vehicle motion state estimation method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that: When the computer program is executed by the processor, it implements the steps of the data-driven deep learning-based vehicle motion state estimation method according to any one of claims 1 to 6.

Images