A vehicle speed estimation method, system, electronic device and storage medium

Abstract

The application discloses a vehicle speed estimation method and system, an electronic device and a storage medium, belongs to the field of distributed drive electric vehicle chassis drive-by-wire and active safety technology, and solves the technical problems that the existing technology has poor extreme working condition adaptability caused by fixed filter noise parameters, lacks an adaptive credibility evaluation mechanism for multi-source sensor fusion, and is difficult to adaptively process nonlinear dynamic characteristics without parameter calibration. By introducing a dynamic double-parameter covariance injection mechanism and a cross-modal attention mechanism in an adaptive cubature Kalman filter framework, the measurement noise weight can be adaptively adjusted according to the space-time correlation characteristics among multi-source sensor signals, the implicit dynamic characteristics learned by a Transformer are used to dynamically correct the mismatch error of a linear physical model, and therefore, the estimation accuracy, system robustness and full-working-condition stability of vehicle speed estimation in a dynamic complex scene are improved. The application is used for realizing high-precision estimation of vehicle speed.

Description

Technical Field

[0001] This invention relates to the field of distributed drive electric vehicle chassis drive-by-wire and active safety technology, specifically to vehicle speed estimation methods, systems, electronic devices and storage media. Background Technology

[0002] Vehicle longitudinal and lateral speeds are fundamental parameters for the operation of active safety control systems in automotive chassis (such as ABS, ESC, and AEB). However, due to cost and installation limitations, mass-produced passenger vehicles typically cannot directly incorporate high-precision optical or GPS / RTK speed measurement equipment. Instead, they rely on low-cost onboard physical sensors such as wheel speed sensors, inertial measurement units (IMUs), and steering wheel angle sensors for state estimation. Existing vehicle speed estimation methods face three main technical challenges that urgently need to be addressed in this scenario: 1. Traditional filters have fixed noise parameters and poor adaptability to extreme dynamic conditions: Existing vehicle speed estimation generally uses traditional models such as extended Kalman filtering or unscented Kalman filtering. The process noise covariance matrix and measurement noise covariance matrix of these methods usually rely on offline calibration by engineers or static setting through a fixed statistical window. When the vehicle encounters sudden changes in operating conditions (such as rapid acceleration, emergency braking, or abrupt changes in road adhesion coefficient), the fixed noise parameters cannot be adjusted in real time according to the current dynamic intensity of the vehicle, resulting in filter response lag and a significant decrease in estimation accuracy under extreme conditions.

[0003] 2. Lack of adaptive reliability assessment mechanism in multi-source sensor information fusion: Existing methods fuse data from wheel speedometers, inertial measurement units (IMUs) (including accelerometers and gyroscopes), and steering angle sensors with fixed weights. This fails to detect changes in the reliability of each sensor under different operating conditions. For example, wheel speedometer signals become distorted when the tires slip, and the cumulative error of the IMU increases under high dynamic conditions, thereby introducing erroneous measurement information and reducing the robustness of the estimation.

[0004] 3. Difficulty in adaptively handling the highly nonlinear dynamic characteristics of vehicles without complex calibration: When vehicles are making extreme cornering or driving on low-traction surfaces, tire mechanical properties exhibit extremely strong nonlinear saturation characteristics (such as nonlinearity at large slip angles). Existing dynamic-based estimation methods either employ simple linear tire models (leading to severe model mismatch at large slip angles) or complex nonlinear tire models (such as the Pacejka magic formula, whose parameter calibration costs are extremely high and are prone to drift with tire pressure, wear, and road conditions). How to implicitly compensate for these difficult-to-model nonlinear dynamic residuals online using data-driven methods without relying on complex bench calibration of nonlinear tire models is a key problem that the industry has not yet properly solved.

[0005] In summary, existing technologies suffer from several technical problems, including poor adaptability to extreme operating conditions due to fixed filter noise parameters, lack of adaptive reliability assessment mechanisms for multi-source sensor fusion, and difficulty in adaptively handling nonlinear dynamic characteristics without parameter calibration. Summary of the Invention

[0006] This invention solves the technical problems of existing technologies, such as poor adaptability to extreme working conditions due to fixed filter noise parameters, lack of adaptive reliability assessment mechanism for multi-source sensor fusion, and difficulty in adaptively handling nonlinear dynamic characteristics without parameter calibration.

[0007] The vehicle speed estimation method of the present invention includes the following steps: Step 1: Acquire multi-source sensor sequence data of the vehicle; Step 2: Design the Transformer encoder architecture and use it to extract features from multi-source sensor sequence data to obtain feature vectors; The feature vector includes at least the longitudinal vehicle speed prediction value, the lateral vehicle speed prediction value, the confidence score, the process noise correction amount, and the measurement noise correction amount; Step 3: Construct the state-space equations and use adaptive capacitive Kalman filtering as the state estimation framework. Combine the process noise correction amount to calculate the prior mean and prior covariance. Step 4: Calculate the amplified measurement vector based on the longitudinal and lateral vehicle speed prediction values, and construct the amplified measurement covariance corresponding to the amplified measurement vector using the confidence score and measurement noise correction. Step 5: Based on the amplified measurement covariance, the prior mean and prior covariance are corrected to obtain the posterior estimate of the vehicle speed, thus completing the vehicle speed estimation.

[0008] Furthermore, in one embodiment of the present invention, step 2, which utilizes the Transformer encoder architecture to extract features from multi-source sensor sequence data, specifically involves: Step 21: Use the group embedding layer in the Transformer encoder architecture to divide the multi-source sensor sequence data into multiple functional groups, and perform feature embedding on each functional group to obtain the embedding vector of each functional group. Step 22: Concatenate the embedding vectors of each functional group and superimpose the learnable temporal position code to form a token sequence; Step 23: A dual-axis attention mechanism is adopted on the token sequence. The dual-axis attention mechanism calculates self-attention in the time dimension and the sensor dimension respectively, and generates a confidence weight based on the self-attention score in the sensor dimension. Step 24: Input the token sequence processed by the dual-axis attention mechanism into the core layer of the Transformer encoder. Introduce learnable classification feature bits at the beginning of the input sequence and extract the output vector of the last layer corresponding to the learnable classification feature bits as the feature vector.

[0009] Furthermore, in one embodiment of the present invention, the Transformer encoder core layer is composed of multiple identical coding blocks stacked together, and each coding block sequentially includes a multi-head self-attention sub-layer, a first residual connection and layer normalization module, a feedforward neural network sub-layer, and a second residual connection and layer normalization module.

[0010] Furthermore, in one embodiment of the present invention, the Transformer encoder architecture includes three output heads, specifically: The speed prediction head is used to output longitudinal vehicle speed prediction values, lateral vehicle speed prediction values, and confidence scores; Noise estimation head, used to output process noise correction and measurement noise correction; The driving condition recognition head is used to output the probability distribution of the current driving condition.

[0011] Furthermore, in one embodiment of the present invention, the probability distribution of the current driving condition is used to smoothly trigger the kinematic model switching when the probability of the low-speed condition exceeds a threshold, and / or to modulate the process noise correction amount as auxiliary information.

[0012] Furthermore, in one embodiment of the present invention, step 3, in conjunction with the process noise correction amount, calculates the prior mean and prior covariance, specifically as follows: Based on the process noise correction, a process noise covariance matrix is ​​constructed. Based on the state space equation and the process noise covariance matrix, a volume point is generated and propagated using the third-order sphere diameter integral rule, and the prior mean and prior covariance are calculated.

[0013] Furthermore, in one embodiment of the present invention, the construction of the amplification measurement covariance corresponding to the amplification measurement vector in step 4 specifically involves: Based on the measurement noise correction, a measurement noise covariance matrix is ​​constructed, and the confidence weights are mapped to the corresponding sensor components in the measurement noise covariance matrix to obtain the first measurement covariance. The confidence scores are used to construct the second measurement covariance, and the first measurement covariance and the second measurement covariance are combined as the amplified measurement covariance.

[0014] The vehicle speed estimation system of the present invention includes the following modules: The data acquisition module acquires multi-source sensor sequence data of the vehicle; The feature extraction module designs a Transformer encoder architecture and uses it to extract features from multi-source sensor sequence data to obtain feature vectors. The feature vector includes at least the longitudinal vehicle speed prediction value, the lateral vehicle speed prediction value, the confidence score, the process noise correction amount, and the measurement noise correction amount; The state prediction module constructs the state space equation and uses adaptive capacitive Kalman filtering as the state estimation framework. Combined with the process noise correction, it calculates the prior mean and prior covariance. The amplified measurement construction module calculates the amplified measurement vector based on the longitudinal and lateral vehicle speed prediction values, and constructs the amplified measurement covariance corresponding to the amplified measurement vector using the confidence score and measurement noise correction. The state update module, based on the amplified measurement covariance, corrects the prior mean and prior covariance to obtain the posterior estimate of the vehicle speed, thus completing the vehicle speed estimation.

[0015] The electronic device of the present invention includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus. Memory, used to store computer programs; The processor, when executing a program stored in memory, implements any of the vehicle speed estimation methods described above.

[0016] The present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the vehicle speed estimation methods described above.

[0017] This invention solves the technical problems of existing technologies, such as poor adaptability to extreme conditions due to fixed filter noise parameters, lack of adaptive reliability assessment mechanism for multi-source sensor fusion, and difficulty in adaptively handling nonlinear dynamic characteristics without parameter calibration. Specific beneficial effects of this invention include: 1. This invention proposes a vehicle speed estimation method. By introducing a dynamic two-parameter covariance injection mechanism based on Transformer (a neural network architecture based on self-attention mechanism) into the Adaptive Cumulative Kalman Filter (ACKF) framework, it overcomes the technical bottleneck of traditional filtering algorithms where the process noise covariance matrix and measurement noise covariance matrix are highly dependent on offline calibration or empirical settings by engineers. The model can predict the uncertainty of the current vehicle dynamics in real time based on the time sequence of multi-source sensors and output dynamic correction values. When the vehicle encounters extreme conditions such as rapid acceleration, emergency braking, or sudden changes in road adhesion coefficient, the filter can respond instantaneously and adaptively adjust the gain, thereby ensuring high accuracy of vehicle speed estimation under all conditions. 2. This invention proposes a vehicle speed estimation method, designing a grouping embedding and spatiotemporal dual-axis attention mechanism for 16-dimensional multi-source sensors (including wheel speed, IMU, and motor torque) in distributed drive vehicles. This mechanism effectively overcomes the fixed weight deficiency of traditional fusion algorithms, and can automatically mine the potential mechanistic correlations between cross-modal physical quantities (such as wheel speed and longitudinal acceleration) through data-driven methods, and evaluate the dynamic reliability of each sensor in real time. When tire slippage or short-term distortion of local sensor signals occurs, the system can implicitly and smoothly reduce the weight of abnormal measurement signals through scaling dot product operations, without the need to design complex explicit fault detection and isolation rules, greatly improving the anti-interference capability and engineering reliability of the chassis state estimation system. 3. This invention proposes a vehicle speed estimation method that cleverly integrates a linear kinematics / dynamics prediction model with a deep learning model, introducing an augmented measurement equation in the observation update step. Utilizing virtual observations and dynamic confidence scores output by the Transformer speed prediction head, the algorithm can automatically compensate for the strong nonlinear tire mechanical mismatch error generated when the vehicle is driving on extreme cornering or low-friction surfaces. Compared to traditional nonlinear estimation schemes relying on Pacejka's magic formula, this invention not only completely eliminates the complex nonlinear tire parameter calibration costs and drift risks, but also endows the traditional adaptive filtering framework with the ability to handle unknown nonlinear residuals, combining theoretical rigor with strong feasibility for mass production. Attached Figure Description

[0018] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of the vehicle speed estimation method described in Implementation Method 1; Figure 2 This is a diagram of the internal structure of the Transformer encoder described in Embodiment 1; Figure 3 This is a diagram of a three-degree-of-freedom vehicle model as described in Implementation Method 1. Detailed Implementation

[0019] Various embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. The embodiments described with reference to the drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0020] Implementation Method 1: Existing technologies mainly have three problems: First, the process / measurement noise covariance of traditional filters is fixed, resulting in lag and decreased accuracy under extreme conditions; second, multi-source sensor fusion uses fixed weights, which cannot detect changes in reliability when tire slippage or sensor signal distortion occurs, easily introducing erroneous measurements; finally, it is difficult to adaptively process the strong nonlinear dynamic characteristics of vehicles without relying on complex nonlinear tire model calibration, leading to severe model mismatch.

[0021] To address the aforementioned technical problems, this embodiment proposes a vehicle speed estimation method based on Transformer and adaptive capacitive Kalman filtering. This method is particularly suitable for high-precision adaptive estimation of longitudinal and lateral vehicle velocities in distributed electric vehicles with four-wheel independent drive under complex and extreme conditions, involving multi-source sensor information fusion. Figure 1 As shown, it specifically includes the following: Step 1: Acquire multi-source sensor sequence data of the vehicle; In this embodiment, the sensor configuration adopts standard chassis sensors for mass-produced passenger vehicles, including: four-wheel wheel speed sensors (providing the angular velocity of each wheel). Inertial Measurement Unit (IMU) (provides longitudinal acceleration) Lateral acceleration yaw rate Steering angle sensor (provides steering wheel angle) ), motor drive braking torque (providing drive or braking torque to each wheel) , ), subscript These represent the front left, front right, rear left, and rear right wheels, totaling 16 sensor dimensions.

[0022] Using a sliding time window mechanism, the sensor data at the current time k and the data from the previous T-1 times are organized into a time-series input matrix: ; The sensor vector at each time step is: ; The window length T is determined based on the sampling rate and computing resources of the vehicle control system, with a typical value of T=20 (corresponding to a 0.2-second historical window at a sampling rate of 100 Hz).

[0023] Step 2: Design the Transformer encoder architecture and use it to extract features from multi-source sensor sequence data to obtain feature vectors; This implementation addresses the differences in physical characteristics among multi-source sensors by designing a grouped embedding layer and a dual-axis attention mechanism to achieve adaptive extraction and reliability assessment of temporal features from each sensor. Figure 2 As shown.

[0024] Based on the physical principles of sensors, the 16-dimensional sensor input is divided into four functional groups, and features are embedded in each group through an independent fully connected mapping layer: ; ; ; ; in, For the embedding dimension, the learnable temporal location encoding is superimposed after concatenating the embedding vectors from each group. The complete feature vector Token formed at time t is represented as: ; The significance of the group embedding is to preserve the dimensional independence of each sensor group, avoid feature interference caused by directly splicing different physical quantities, and enable the subsequent attention mechanism to perform credibility evaluation at the sensor group level.

[0025] The token sequence (label sequence) obtained from group embedding carries two types of information: the physical quantities of each sensor at different times, and the relationships between sensors at different times. To extract these two types of information separately, a dual-axis attention mechanism is introduced inside the Transformer encoder, performing attention calculations on the time dimension and the sensor dimension respectively. (a) Temporal attention: for a token sequence over T time steps Applying multi-head self-attention first involves using a learnable weight matrix. , , Map the sequence to a query matrix Key matrix Sum matrix : ; ; ; Subsequently, the dynamic correlation features of the sequence in the time dimension are calculated by scaling the dot product. To capture the dynamic characteristics of key historical moments: ; in, , which is the scaling factor (i.e., the dimension of the key matrix). The function is used to normalize the weights of historical time steps. Perform matrix multiplication to find the transpose of the query matrix and the key matrix. If multi-head attention is used, multiple such multiplications will be performed. The results are output after being concatenated along the feature dimensions.

[0026] (b) Sensor Attention: Apply attention to the sensor group feature dimension at the current time k. Define the feature set at the current time as... This matrix is ​​used to calculate the cross-modal correlations between each sensor group in order to evaluate its dynamic confidence weights. First, a sensor query matrix is ​​generated through a linear mapping layer. Key matrix Sum matrix : ; ; ; Subsequently, the correlation score matrix between sensor groups is calculated. and through a linear layer The dynamic reliability weight vector of each sensor group is obtained through mapping. : ; ; in, The confidence weights correspond to the wheel assembly, accelerometer assembly, gyroscope assembly, and steering angle sensor assembly, respectively.

[0027] Credibility weight It has a clear physical meaning: when tire slippage causes a mechanistic conflict between wheel speed and acceleration characteristics, the attention mechanism automatically detects this anomaly through the aforementioned dot product operation and reduces it accordingly. This enables soft fault tolerance for implicit sensor failures.

[0028] In this embodiment, the core layer of the Transformer encoder consists of three identical coding blocks stacked together. Each coding block sequentially includes a multi-head attention sublayer, a first residual connection and layer normalization module, a feedforward neural network (FFN, consisting of two fully connected layers and a ReLU activation function) sublayer, and a second residual connection and layer normalization module. A learnable classification feature bit ([CLS] Token) is introduced at the beginning of the input sequence. As the sequence passes through three layers of deep feature extraction, the final output vector corresponding to the [CLS] Token is finally extracted. , This is the dimension of the feature vector. This vector highly aggregates global temporal features and cross-modal correlations between sensors, and serves as the unified feature input for the subsequent three output heads. For the specific structure, please refer to [reference needed]. Figure 2 .

[0029] Based on the final layer output vector The three types of outputs are generated in parallel through three independent fully connected output heads. The structure, output content, and downstream applications of each output head are described below: First path: Speed ​​prediction head: Consists of two fully connected layers, outputting longitudinal vehicle speed prediction values. Lateral vehicle speed prediction value The units are m / s. Neither of these are the final estimation results, but rather are directly injected into the amplified measurement vector in step 4 as virtual observations. The output... and It can provide additional nonlinear correction information for adaptive capacitive Kalman filtering when the linear model is mismatched. The output header also outputs a confidence score confidence_k∈(0,1), which is used to dynamically adjust the measurement noise weights of the virtual observations in step 4.

[0030] Second path: Noise estimation head: Consists of two fully connected layers, outputting process noise correction. and measurement noise correction amount The output layer activation function uses the Softplus function (a smooth linear rectified function). This ensures that the output value is always positive. These two sets of output values ​​are respectively added to the base process noise covariance in steps 3 and 4. Covariance of baseline measurement noise This constitutes the dynamic process noise covariance matrix. With measurement noise covariance matrix .

[0031] The third approach: Driving condition recognition head: Consists of a two-layer fully connected network, with the output layer activated using Softmax (normalized exponential function), outputting the probability distribution of the current driving condition. It covers five typical driving conditions: steady-state straight driving, rapid acceleration / deceleration, steady-state turning, extreme turning (large sideslip), and driving on low-adhesion surfaces. The output head has two downstream applications: firstly, in the low-speed singularity processing of step 3, it smoothly triggers a kinematic model switch when the probability of the low-speed driving condition exceeds a threshold, avoiding state jumps caused by hard switching; secondly, the driving condition probability can be used as auxiliary information for… Modulation can be applied, for example, by further increasing the process noise corresponding to the lateral speed when extreme cornering conditions are detected. This strengthens the compensation for nonlinear residuals.

[0032] Step 3: Construct the state-space equations and use adaptive capacitive Kalman filtering as the state estimation framework. Combine the process noise correction amount to calculate the prior mean and prior covariance. In this implementation, a three-degree-of-freedom (3-DOF) vehicle dynamics model is established as the state-space equation for the adaptive capacitive Kalman filter estimator. The establishment process is as follows: A three-degree-of-freedom (3-DOF) model of the vehicle is established, including longitudinal motion (along the vehicle's x-axis), lateral motion (along the vehicle's y-axis), and yaw motion (rotation about the vehicle's z-axis). First, a right-handed coordinate system is established with the vehicle's center of mass as the origin. During the vehicle's planar motion, the longitudinal velocity at the center of mass is... Lateral speed and the yaw rate about the z-axis ,like Figure 3 As shown. Based on Newton's second law, the equations of motion for the center of mass plane are established in the vehicle coordinate system as follows: ; in, and These are the vehicle's mass and moment of inertia, respectively. , , , , and These are longitudinal force, lateral force, air resistance, rolling resistance, climbing resistance, and yaw moment, respectively. The formulas for calculating air resistance, rolling resistance, and climbing resistance are as follows: ; in, This is the aerodynamic drag coefficient. The vehicle's frontal area. air density, The rolling resistance coefficient, It is the acceleration due to gravity. The road surface slope. The formulas for calculating the resultant longitudinal and lateral forces and moments of the vehicle are: ; ; ; in, and These are the horizontal distances from the front and rear axles to the center of gravity, respectively. and These are the track widths for the front and rear axles, respectively.

[0033] Using a single-wheel model to describe the dynamic characteristics of each tire, its equation of motion can be given by the following equation: ; in, and These are the effective radius and moment of inertia of the wheel, respectively. and These are driving torque and braking torque, respectively. and These are the tire longitudinal force and rolling resistance, respectively.

[0034] Considering that all tires operate in the linear region most of the time, a small slip angle assumption is made. The relationship between tire lateral force and tire slip angle can be given by the following formula: ; in, and These are the lateral stiffness and lateral angle of each tire.

[0035] Define the state vector, control input vector, and output vector as follows: ; ; ; Establish the continuous-time nonlinear state-space equations: ; in, and The equivalent total lateral stiffness of the front and rear axles are respectively; process noise. , The noise estimation is provided in real time by the Transformer noise estimation head.

[0036] Establish the continuous-time nonlinear measurement equation: ; Among them, measurement noise , It is also provided in real time by the Transformer dynamic covariance injection mechanism.

[0037] This step is the core innovation of the present invention. It adopts adaptive capacitive Kalman filtering (ACKF) as the state estimation framework, and injects the dynamic noise parameters output by Transformer in step 2 and the virtual observations into the filter in real time, so as to realize the organic integration of physical model and data-driven capability.

[0038] Compared to the extended Kalman filter, the adaptive capacitive Kalman filter uses the third-order spherical diameter integral rule to approximate nonlinear state propagation, eliminating the need to calculate the Jacobian matrix. This results in superior numerical stability and estimation accuracy in scenarios with highly nonlinear vehicle dynamics.

[0039] At time k, the output of the noise estimation head is used. and Construct the dynamic process noise covariance matrix With measurement noise covariance matrix : ; ; in, , These are the basic process noise covariance and measurement noise covariance for offline calibration, respectively. The term represents the real-time correction value predicted by the Transformer, which is always non-negative (guaranteed by the Softplus activation function used in the output header). By superimposing the real-time correction value predicted by the Transformer on this base value, the filter can adjust its behavior according to the severity of the current operating conditions (as determined by...). (reflection) and sensor health (by) (Reflects) Adaptive adjustment of gain.

[0040] Furthermore, the sensor confidence weights Mapped to The corresponding sensor components are used to achieve dynamic measurement noise adjustment at the sensor group level: ; in, To prevent small constants from being divided by zero, this design unifies dynamic covariance injection and cross-modal attention credibility within the same mathematical framework, achieving end-to-end coupling.

[0041] The third-order spherical diameter integral rule (Cubature Rule) is used to generate a set of 2n (n=3 for the state dimension) volume points { }: ; in, Let be the i-th standard basis vector (i≤n) or its negative value (i>n). This is achieved through the discrete state transition function. Propagate to each volume point and calculate the prior mean and prior covariance: ; .

[0042] Step 4: Calculate the amplified measurement vector based on the longitudinal and lateral vehicle speed prediction values, and construct the amplified measurement covariance corresponding to the amplified measurement vector using the confidence score and measurement noise correction. By concatenating physical sensor observations with Transformer virtual velocity observations, an augmented measurement vector is constructed: ; ; in, For physical sensor measurements, , For the virtual observations output by the velocity prediction head, This is the Jacobian matrix for measuring the state by a physical sensor. This is the measurement matrix for the state of the virtual observation pair.

[0043] Due to the physical measurement equation Since it is a nonlinear time-varying function, it cannot be updated using a fixed linear measurement matrix H. It utilizes 2n volume points generated during the prediction step. The propagation is performed directly through a nonlinear measurement function to calculate the predicted measurement mean and measurement covariance. ; ; ; In the above derivation, the volume point is directly obtained through a nonlinear function. Propagation can be performed without calculating the Jacobian matrix of the measurement equation with respect to the state. Virtual observations output by the velocity prediction head , Directly corresponding to the state vector , Components, measurement equations Linear: ; The measurement noise corresponding to the Transformer virtual observation is dynamically determined by the confidence score of the velocity prediction head: ; The higher the confidence level, the lower the measurement noise corresponding to the virtual observation, and the greater the contribution of the virtual observation to the filter correction; under extreme nonlinear conditions, With increased Transformer confidence, the filter automatically enhances its dependence on data-driven virtual observations, achieving implicit compensation for nonlinear residuals.

[0044] By merging the covariances of the two types of measurements, an amplified measurement covariance is constructed: ; .

[0045] Step 5: Based on the amplified measurement covariance, the prior mean and prior covariance are corrected to obtain the posterior estimate of the vehicle speed, thus completing the vehicle speed estimation.

[0046] The volume point approximation measurement propagation is used to complete the amplification measurement update process. First, the Kalman gain is calculated. : ; This leads to the posterior correction of the state vector and the update of the covariance matrix: ; ; Output the optimal state estimate at the current moment: ; The estimation results and The output is sent to the upper-level chassis control system (such as ESC, ABS, or AEB), and the posterior state is also transmitted. This serves as the initial value for the filter at the next time step, completing the recursive iterative process of the algorithm.

[0047] This implementation introduces a dynamic two-parameter covariance injection mechanism and a cross-modal attention mechanism into the adaptive capacitive Kalman filter framework. It can adaptively adjust the measurement noise weights according to the spatiotemporal correlation characteristics between multi-source sensor signals, and dynamically correct the mismatch error of the linear physical model by utilizing the implicit dynamic characteristics learned by Transformer. This significantly improves the estimation accuracy, system robustness and stability of vehicle speed estimation in dynamic and complex scenarios.

[0048] Implementation Method 2: A vehicle speed estimation system, comprising the following modules: The data acquisition module acquires multi-source sensor sequence data of the vehicle; The feature extraction module designs a Transformer encoder architecture and uses it to extract features from multi-source sensor sequence data to obtain feature vectors. The feature vector includes at least the longitudinal vehicle speed prediction value, the lateral vehicle speed prediction value, the confidence score, the process noise correction amount, and the measurement noise correction amount; The state prediction module constructs the state space equation and uses adaptive capacitive Kalman filtering as the state estimation framework. Combined with the process noise correction, it calculates the prior mean and prior covariance. The amplified measurement construction module calculates the amplified measurement vector based on the longitudinal and lateral vehicle speed prediction values, and constructs the amplified measurement covariance corresponding to the amplified measurement vector using the confidence score and measurement noise correction. The state update module, based on the amplified measurement covariance, corrects the prior mean and prior covariance to obtain the posterior estimate of the vehicle speed, thus completing the vehicle speed estimation.

[0049] Implementation Method 3: An electronic device includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus. Memory, used to store computer programs; The processor, when executing the program stored in memory, implements the vehicle speed estimation method described above.

[0050] Implementation Method 4: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described vehicle speed estimation method.

[0051] The present invention has provided a detailed description of a vehicle speed estimation method, system, electronic device, and storage medium. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, those skilled in the art will recognize that there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A vehicle speed estimation method, characterized in that, Includes the following steps: Step 1: Acquire multi-source sensor sequence data of the vehicle; Step 2: Design the Transformer encoder architecture and use it to extract features from multi-source sensor sequence data to obtain feature vectors; The feature vector includes at least the longitudinal vehicle speed prediction value, the lateral vehicle speed prediction value, the confidence score, the process noise correction amount, and the measurement noise correction amount; The feature extraction of multi-source sensor sequence data using the Transformer encoder architecture is specifically as follows: Step 21: Use the group embedding layer in the Transformer encoder architecture to divide the multi-source sensor sequence data into multiple functional groups, and perform feature embedding on each functional group to obtain the embedding vector of each functional group. Step 22: Concatenate the embedding vectors of each functional group and superimpose the learnable temporal position code to form a token sequence; Step 23: A dual-axis attention mechanism is adopted on the token sequence. The dual-axis attention mechanism calculates self-attention in the time dimension and the sensor dimension respectively, and generates a confidence weight based on the self-attention score in the sensor dimension. Step 24: Input the token sequence processed by the dual-axis attention mechanism into the core layer of the Transformer encoder. Introduce learnable classification feature bits at the beginning of the input sequence and extract the output vector of the last layer corresponding to the learnable classification feature bits as the feature vector. Step 3: Construct the state-space equations and use adaptive capacitive Kalman filtering as the state estimation framework. Combine the process noise correction amount to calculate the prior mean and prior covariance. Step 4: Calculate the amplified measurement vector based on the longitudinal and lateral vehicle speed prediction values, and construct the amplified measurement covariance corresponding to the amplified measurement vector using the confidence score and measurement noise correction. Step 5: Based on the amplified measurement covariance, the prior mean and prior covariance are corrected to obtain the posterior estimate of the vehicle speed, thus completing the vehicle speed estimation.

2. The vehicle speed estimation method according to claim 1, characterized in that, The Transformer encoder core layer is composed of multiple identical coding blocks stacked together. Each coding block sequentially includes a multi-head self-attention sub-layer, a first residual connection and layer normalization module, a feedforward neural network sub-layer, and a second residual connection and layer normalization module.

3. The vehicle speed estimation method according to claim 1, characterized in that, The Transformer encoder architecture includes three output heads, specifically: The speed prediction head is used to output longitudinal vehicle speed prediction values, lateral vehicle speed prediction values, and confidence scores; Noise estimation head, used to output process noise correction and measurement noise correction; The driving condition recognition head is used to output the probability distribution of the current driving condition.

4. The vehicle speed estimation method according to claim 3, characterized in that, The probability distribution of the current driving condition is used to smoothly trigger the kinematic model switching when the probability of the low-speed condition exceeds the threshold, and / or as auxiliary information to modulate the process noise correction amount.

5. The vehicle speed estimation method according to claim 1, characterized in that, In step 3, the prior mean and prior covariance are calculated by combining the process noise correction amount, specifically as follows: Based on the process noise correction, a process noise covariance matrix is ​​constructed. Based on the state space equation and the process noise covariance matrix, a volume point is generated and propagated using the third-order sphere diameter integral rule, and the prior mean and prior covariance are calculated.

6. The vehicle speed estimation method according to claim 1, characterized in that, In step 4, constructing the amplification measurement covariance corresponding to the amplification measurement vector specifically involves: Based on the measurement noise correction, a measurement noise covariance matrix is ​​constructed, and the confidence weights are mapped to the corresponding sensor components in the measurement noise covariance matrix to obtain the first measurement covariance. The confidence scores are used to construct the second measurement covariance, and the first measurement covariance and the second measurement covariance are combined as the amplified measurement covariance.

7. A vehicle speed estimation system, said system being used to implement the vehicle speed estimation method of claim 1, characterized in that, Includes the following modules: The data acquisition module acquires multi-source sensor sequence data of the vehicle; The feature extraction module designs a Transformer encoder architecture and uses it to extract features from multi-source sensor sequence data to obtain feature vectors. The feature vector includes at least the longitudinal vehicle speed prediction value, the lateral vehicle speed prediction value, the confidence score, the process noise correction amount, and the measurement noise correction amount; The state prediction module constructs the state space equation and uses adaptive capacitive Kalman filtering as the state estimation framework. Combined with the process noise correction, it calculates the prior mean and prior covariance. The amplified measurement construction module calculates the amplified measurement vector based on the longitudinal and lateral vehicle speed prediction values, and constructs the amplified measurement covariance corresponding to the amplified measurement vector using the confidence score and measurement noise correction. The state update module, based on the amplified measurement covariance, corrects the prior mean and prior covariance to obtain the posterior estimate of the vehicle speed, thus completing the vehicle speed estimation.

8. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the vehicle speed estimation method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the vehicle speed estimation method according to any one of claims 1-6.

Images