An Adaptive Hybrid State Estimation Method for Vehicle Active Safety Control Systems

Abstract

This invention provides an adaptive hybrid state estimation method for vehicle active safety control systems. The method includes collecting sensor data during vehicle operation, including front wheel steering angle, lateral acceleration, side acceleration, yaw rate, longitudinal velocity, and longitudinal acceleration. The sensor data is processed using a deep neural network to output a virtual estimate of the vehicle's sideslip angle and its estimation uncertainty. A state-space equation is constructed based on a three-degree-of-freedom vehicle dynamics model, and model-based state estimation is performed using the virtual estimate. The process noise covariance matrix and measurement noise covariance matrix are dynamically adjusted using an adaptive optimization strategy. The output of the deep neural network and the model-based state estimation result are fused to output a high-precision vehicle state estimate. This method achieves high-precision estimation and strong robustness of vehicle state under non-ideal conditions.

Description

Technical Field

[0001] This invention belongs to the field of intelligent driving and vehicle dynamics control, and in particular relates to an adaptive hybrid state estimation method for vehicle active safety control systems. Background Technology

[0002] In the context of the deep integration of automotive intelligence and connectivity, reliable state perception of intelligent driving systems has become a core research topic for ensuring driving safety and promoting the implementation of autonomous driving technology. As a key enabling link in the intelligent driving technology system, the realization of Advanced Driver Assistance Systems (ADAS) highly depends on the accurate acquisition of vehicle dynamics, such as the center of gravity sideslip angle and yaw rate. The estimation accuracy of these key state information not only directly affects the accuracy of system decisions but also determines the timeliness of intervention by active safety control systems in emergency situations. However, these vehicle states are difficult to obtain directly from sensors, and using high-precision MEMS / IMU and other specialized sensing equipment is costly. Secondly, existing filters based on traditional Kalman algorithms only perform well when sensor noise follows a Gaussian distribution, but do not consider estimation accuracy in non-Gaussian noise environments.

[0003] Against the backdrop of the rapid development of intelligent driving technology, vehicle state estimation, as a core technology supporting driving safety, is increasingly demonstrating its research value. Accurate acquisition of dynamic parameters such as the vehicle's center of gravity and sideslip angle is a fundamental prerequisite for active safety systems to achieve key functions such as rollover and skid prevention. However, due to the high cost of direct sensor measurement and insufficient reliability under complex conditions, existing methods are insufficient to meet the needs of large-scale applications. Therefore, developing a highly robust state estimation algorithm based on multi-source information fusion has become a breakthrough in overcoming the safety bottleneck of intelligent transportation systems and complex scenarios. Artificial intelligence technology has been applied to various research fields and has achieved significant results in intelligent transportation systems and autonomous driving control. Therefore, how to effectively combine... Enhancing the data processing capabilities of data-driven models and improving their interpretability has become a hot topic in vehicle state estimation algorithm research. However, existing research rarely considers the impact of non-ideal conditions, such as time-varying noise, model uncertainty, and non-Gaussian noise. In reality, non-ideal conditions are more common because onboard sensors are often subjected to interference from harsh conditions, generating non-Gaussian noise. Summary of the Invention

[0004] The purpose of this invention is to provide an adaptive hybrid state estimation method for vehicle active safety control systems, which fully leverages the powerful data processing capabilities of data-driven approaches and combines them with the physical law capabilities of model-based Kalman filtering to achieve high-precision estimation and strong robustness of vehicle states under non-ideal conditions.

[0005] To achieve the above objectives, this invention provides an adaptive hybrid state estimation method for vehicle active safety control systems, comprising the following steps: Collect sensor data during vehicle operation, including front wheel steering angle and lateral acceleration. Lateral acceleration, yaw rate γ, longitudinal velocity and longitudinal acceleration ; The sensor data is processed based on a deep neural network to output a virtual estimate of the vehicle sideslip angle and its estimation uncertainty. A state-space equation is constructed based on a three-degree-of-freedom vehicle dynamics model, and a model-based state estimation is performed by combining the virtual estimates. The process noise covariance matrix and measurement noise covariance matrix are dynamically adjusted using an adaptive optimization strategy. By fusing the output of the deep neural network with the model-based state estimation results, a high-precision vehicle state estimate is output.

[0006] Preferably, the deep neural network adopts a sequence network structure that combines a convolutional neural network and a bidirectional long short-term memory network.

[0007] Preferably, the convolutional neural network includes convolutional layers, activation function layers, max pooling layers, and fully connected layers, wherein the activation function is a rectified linear unit.

[0008] Preferably, the state estimation based on the three-degree-of-freedom vehicle dynamics model employs an extended Kalman filter algorithm based on maximum correlation entropy.

[0009] Preferably, the adaptive optimization strategy employs an improved sand cat swarm optimization algorithm to dynamically optimize the noise covariance matrix.

[0010] Preferably, the sand cat swarm optimization algorithm introduces a spiral search strategy factor to enhance the algorithm's global search capability.

[0011] Preferably, the vehicle state includes longitudinal speed, lateral speed, yaw rate, and sideslip angle.

[0012] Preferably, the method further includes inputting the estimated vehicle state into the vehicle active safety control system for electronic stability control decision-making.

[0013] Therefore, this invention employs the aforementioned adaptive hybrid state estimation method for vehicle active safety control systems, combining the high precision of data-driven approaches with the physical consistency of model-driven approaches. The data-driven method can learn complex nonlinear relationships, while the model-driven method avoids overfitting or extrapolation failure inherent in purely data-driven approaches through physical constraints. The data-driven part extracts features from limited data, while the model part fills in the gaps in extreme condition data through physical laws, avoiding unreliable predictions from pure black-box models. Data-driven approaches compensate for errors caused by model simplification, while model-driven approaches correct the generalization deficiencies of data-driven approaches. Simultaneously, end-to-end training achieves synergistic optimization of both approaches, with complementarity being the core principle. Attached Figure Description

[0014] Figure 1 The diagram shows the structure of a vehicle state estimation system based on the hybrid method. Figure 2 Here is a flowchart of the deep neural network training process; Figure 3 This is an extended Kalman filter algorithm based on MCC. Detailed Implementation

[0015] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0016] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0017] Example 1 like Figure 1 and Figure 3 As shown, an adaptive hybrid state estimation method for vehicle active safety control systems includes the following steps: Step 1: Collect sensor data during vehicle operation, including front wheel steering angle and lateral acceleration. Lateral acceleration, yaw rate γ, longitudinal velocity and longitudinal acceleration ; The vehicle-mounted sensor package includes a steering angle sensor, an inertial measurement unit, radar, and wheel speed sensors, which can measure information such as front wheel steering angle, lateral acceleration, side acceleration, and yaw rate, providing a data source for model training and filtering algorithms.

[0018] Adaptive Optimization Strategy Module: This module is mainly responsible for optimizing parameters, specifically the measurement noise covariance matrix of the nonlinear vehicle system when the distributed electric vehicle is driving normally. process noise covariance matrix It is unknown non-Gaussian noise. and The matrix plays a crucial role in the estimation accuracy of the filter. Therefore, this module can optimize the noise covariance matrix according to user needs to improve the estimation accuracy of the algorithm.

[0019] Vehicle critical state output module: Outputs values ​​based on the estimated values ​​from the dynamic model MCCEKF. and The vehicle sideslip angle can be directly obtained. : ; Step 2: Process the sensor data based on a deep neural network to output a virtual estimate of the vehicle sideslip angle and its estimation uncertainty; Deep Neural Network Module: First, train the DNN model offline, such as... Figure 2 As shown, the yaw rate γ of the vehicle and the steering wheel angle of the vehicle are... longitudinal velocity longitudinal acceleration and lateral acceleration The virtual estimate of the sideslip angle is used as input to the model. The estimated level of distrust This serves as the model's output. Simultaneously, driving data under different driving conditions are selected to form a dataset, ensuring the model's generalization ability. The designed DNN model can not only extract local and long-term features of vehicle state data but also highlight important features. A sequential network architecture is constructed using a Convolutional Neural Network (CNN) and a Long Short-Term Memory Network (BiLSTM). Its CNN equation is shown below: Convolutional layers are used to extract input feature points, which are calculated using the following formula: ; In the formula, For input to the convolutional layer, Indicates the first Each output feature map Represents the weight matrix. It is a dot product. This represents the deviation vector.

[0020] Choose the Rectified Linear Unit (ReLU) as the activation function for the CNN: ; This represents an element of the feature map computed from the convolution operation.

[0021] Max pooling is a method that derives the maximum value of a specified region in the feature map from the equation, as shown below: ; In the formula, For max pooling, This represents the deviation. The output of the fully connected layer is calculated as follows: ; This represents the final output vector; Indicates deviation; Represents the weight matrix. The output of max pooling is defined as... .

[0022] BiLSTM (Bidirectional Long Short-Term Memory) provides three key gating structures: the input gate, the output gate, and the forget gate. Two memory states are included: hidden state and memory cell. The forget gate controls the flow of prior information, and its output is calculated as follows: in, The weight matrix from the input features to the forget gate. This is the weight matrix from the hidden state in the previous time step to the forget gate. Represents the sigmoid function; Represents weight; This represents the current input. Assuming 1 ≤ t ≤ m / 2, the output of the previous unit is defined as... The bias is defined as The amount of new data stored in a cell is determined by the input gate, which is calculated using the following formula: ; The weight matrix from the input features to the input gate. This is the weight matrix from the hidden state in the previous time step to the input gate. For the bias term of the input gate, The weight matrix is ​​the input features to the candidate memory units. This is the weight matrix for the hidden state from the candidate memory unit in the previous time step. These are candidate memory units at the current moment.

[0023] The final hidden state of an LSTM (Long Short-Term Memory) cell is calculated by multiplying the modulation cell state by the activated output gate signal: ; This is the output gate for the current moment. For the current LSTM memory cell, BiLSTM in time The final output features It can be obtained in the following ways: ; In the formula, Forward output features, For Long Short-Term Memory (LSTM) networks, This is the inverse output feature.

[0024] Step 3: Construct state-space equations based on the three-degree-of-freedom vehicle dynamics model, and perform model-based state estimation by combining the virtual estimates; The model-based state estimation module: This module uses a three-degree-of-freedom vehicle model as its dynamic basis, establishing lateral, longitudinal, and yaw degrees of freedom. It then proposes discrete-time process equations and measurement equations, while considering non-Gaussian noise interference from onboard sensors. The process model is the modeling model, and the observation model is the actual vehicle. The discrete state-space equations are: ; In the formula, For state vectors, For the observation vector, For control vectors, To measure noise, This is process noise.

[0025] Based on the three-degree-of-freedom model of the vehicle, establish the dynamics and kinematic equations: ; ; ; For the longitudinal force of the vehicle's front wheels, For the front wheel steering angle, The lateral force on the front wheels of the vehicle. For the longitudinal force of the vehicle's rear wheels, Let yaw rate be the vehicle's angular velocity. The lateral force on the front wheels of the vehicle. This refers to the front wheelbase. This refers to the rear wheelbase. The rear wheel track. The front wheel track. This represents the vehicle's moment of inertia.

[0026] The state vector is The observation vector is .in This refers to the output of the DNN model.

[0027] The state matrix is ​​as follows: ; Among them, "uniformly" was changed to This represents the transfer function, where m is the total vehicle mass. ; ; ; The observation matrix is ​​as follows: ; In the formula, This refers to the longitudinal force of the vehicle's tires. This refers to the lateral force on the vehicle's tires. Represents the observation function, Active safety control systems: Vehicle active safety control systems significantly reduce the accident rate through real-time intervention. The sideslip angle, as a core state parameter, directly determines the reliability of the control system under extreme conditions. For example, the core function of electronic stability control is to generate a corrective yaw moment to suppress sideslip by applying braking force to individual wheels or adjusting driving force. If the sideslip angle is inaccurately estimated, the system will fail to identify vehicle instability in a timely manner, leading to a delay in intervention and potentially missing the optimal control window. Incorrect sideslip angle estimation can cause electronic stability control to apply inappropriate braking force or torque, exacerbating vehicle instability. In this case, a phase plane can be used to illustrate stability judgment. ; in, The rate of change of the vehicle's center of gravity sideslip angle. and This is the stability zone boundary coefficient, which is related to vehicle speed, sideslip angle, and road conditions. Therefore, when certain conditions are met, the vehicle is in a stable state, and the stability control system does not need to adjust the braking force of each wheel. When these conditions are not met, the vehicle will activate stability control to ensure vehicle stability and thus guarantee safety. Therefore, accurately estimating the vehicle's center of gravity sideslip angle is crucial for active safety systems.

[0028] Step 4: Dynamically adjust the process noise covariance matrix and measurement noise covariance matrix using an adaptive optimization strategy; The optimization algorithm in the adaptive strategy is an improved sand cat swarm optimization algorithm, which determines the location of prey on the ground and underground by utilizing low-frequency noise detection capabilities. Therefore, the sand cat can determine the optimal location for its prey. ; In the formula, To indicate at time The position at that time The vector of the current position. To represent the local optimal position, A random number between 0 and 1 To indicate that the sand cat is at any time The location.

[0029] This equation allows the algorithm to explore new local optima within the search region. The algorithm reduces the randomness of the search agent by determining a new position between the current position and the prey position through a stochastic process.

[0030] The sand cat's hunting sensitivity range is circular, so the individual's movement direction is determined by random angles on the circle. This avoids getting trapped in local optima. According to the random angles in the following formula, the sand cat can approach its hunting position.

[0031] ; In the formula, To indicate the current time At a certain location below, This indicates that the sand cat is at all times Location, To represent angles, This refers to the current moment.

[0032] Similarly, each member of the population can move along different circular directions in the search space, choosing a random angle for each individual sand cat. This ensures that participating sand cats can approach their prey. This indicates a random location and ensures that participating cats can approach their prey.

[0033] A cube graph is a chaotic graph that generates sequences within the interval (0, 1). This graph uses parameters... The definition is as follows.

[0034] ; In the formula, This is the value of the variable in step n.

[0035] The scatter plot shows that within the interval (0, 1), An optimal chaotic sequence was generated. During the prey search phase, the following spiral search strategy factor was incorporated into the equation. .

[0036] ; Equation (1) is rewritten as: ; In the formula, This indicates the search parameters.

[0037] By integrating spiral search exploration, sand cats possess multiple search paths, allowing them to better adjust their positions. The sand cat swarm will search the search space in a spiral pattern, expanding its ability to explore unknown regions, increasing the likelihood of the algorithm escaping local optima, and effectively improving the overall search performance of the algorithm.

[0038] Step 5: Combine the output of the deep neural network with the model-based state estimation results to output a high-precision vehicle state estimate.

[0039] Therefore, the present invention adopts the above-mentioned adaptive hybrid state estimation method for vehicle active safety control systems, aiming to ensure the accuracy of vehicle state estimation under noise interference environment, providing an innovative solution for vehicle state estimation under noise interference conditions, and has significant engineering application value.

[0040] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An adaptive hybrid vehicle state estimation method for active safety control systems, characterized in that, Includes the following steps: Step 1: Collect sensor data during vehicle operation, including front wheel steering angle and lateral acceleration. Lateral acceleration, yaw rate γ, longitudinal velocity and longitudinal acceleration ; Step 2: Process the sensor data based on a deep neural network to output a virtual estimate of the vehicle sideslip angle and its estimation uncertainty; Step 3: Construct state-space equations based on the three-degree-of-freedom vehicle dynamics model, and perform model-based state estimation by combining the virtual estimates; Step 4: Dynamically adjust the process noise covariance matrix and measurement noise covariance matrix using an adaptive optimization strategy; Step 5: Combine the output of the deep neural network with the model-based state estimation results to output a high-precision vehicle state estimate.

2. The adaptive hybrid vehicle state estimation method for an active safety control system for vehicles according to claim 1, characterized in that, In step 2, the deep neural network adopts a sequence network structure that combines a convolutional neural network and a bidirectional long short-term memory network.

3. The adaptive hybrid vehicle state estimation method for an active safety control system for vehicles according to claim 2, characterized in that, The convolutional neural network includes convolutional layers, activation function layers, max pooling layers, and fully connected layers, wherein the activation function is a rectified linear unit.

4. The adaptive hybrid vehicle state estimation method for an active safety control system for vehicles according to claim 1, characterized in that, In step 3, the state estimation based on the three-degree-of-freedom vehicle dynamics model adopts the extended Kalman filter algorithm based on maximum correlation entropy.

5. The adaptive hybrid vehicle state estimation method for an active safety control system for vehicles according to claim 1, characterized in that, In step 4, the adaptive optimization strategy employs an improved sand cat swarm optimization algorithm to dynamically optimize the noise covariance matrix.

6. The adaptive hybrid vehicle state estimation method for an active safety control system for vehicles according to claim 5, characterized in that, The sand cat swarm optimization algorithm introduces a spiral search strategy factor to enhance the algorithm's global search capability.

7. The adaptive hybrid vehicle state estimation method for an active safety control system for vehicles according to claim 1, characterized in that, In step 5, the vehicle status includes longitudinal speed, lateral speed, yaw rate, and sideslip angle.

8. The adaptive hybrid vehicle state estimation method for an active safety control system for vehicles according to claim 1, characterized in that, The method also includes inputting the estimated vehicle state into the vehicle active safety control system for electronic stability control decisions.

Images