Vehicle stability control multi-domain safety fusion method and system based on double neural networks

Abstract

This invention discloses a multi-domain safety fusion method and system for vehicle stability control based on dual neural networks. The method includes: acquiring vehicle operating status and local perception data, and preprocessing external collaborative data to construct an external environment information set; calculating curvature and normalized consistency deviation characteristics using multiple independent perception modes; identifying physical dynamic parameters and adaptively calculating the comprehensive stability risk value by combining road excitation confidence factors; constructing a cascaded topology and constraint coupling path of dual neural networks; inputting the multi-source heterogeneous joint feature space into the first neural network to calculate the fusion safety factor; introducing the fusion safety factor into the variable weight loss function of the second neural network to form structural constraints, dynamically calculating the PID gain, and generating target control commands. This invention achieves deep fusion of four domains: functional safety, expected functional safety, information security, and active safety, solving the problem of nonlinear coupling and adaptive control of multi-source risks.

Description

Technical Field

[0001] This invention relates to the field of road vehicle driving control and intelligent data processing technology, specifically to a multi-domain safety fusion method and system for vehicle stability control based on dual neural networks. Background Technology

[0002] With the rapid development of intelligent data processing and automatic control technologies, modern commercial vehicles and special mobile platforms have evolved into complex control systems integrating perception, decision-making, and execution. These systems often operate in unstructured environments, and the dynamic changes in the load center position pose significant challenges to attitude stability control. However, under the current intelligent connected vehicle architecture, the boundaries of instability risks faced by the system have expanded considerably, no longer limited to a single physical dynamic instability (active safety domain), but evolving into a multi-dimensional hybrid risk encompassing functional safety (random failure of system hardware), expected functional safety (limitations of perception algorithms and environmental interference), and information security (malicious attacks from external networks). Under extreme conditions, vulnerabilities in a single safety domain (such as tampering with external collaborative communication data or blinding environmental perception sensors by strong light) often induce fatal instability accidents at the physical layer. Therefore, constructing a collaborative control system that can deeply integrate the above four safety domains and achieve adaptive tuning of physical control parameters is a key technical challenge that urgently needs to be solved in the field of intelligent control.

[0003] To address the stability control of the aforementioned multidimensional mixed risk patterns, existing technologies mainly focus on offline static analysis, linear weighted quantization, and single physical dynamics control. For example, Chinese patent application CN202510750696.8, "A Risk Assessment Method and Related Equipment for Multidimensional Safety Fusion," proposes to identify hazardous events by acquiring data sets to conduct risk analysis on functional safety, expected functional safety, and information security. However, this approach is essentially an offline static engineering analysis during the system design phase, lacking an online real-time computational closed loop and unable to provide immediate control responses in the event of sudden network attacks or sensor failures. Chinese patent application CN202410976225.4, "A Quantitative Analysis Method for Fusion Safety Attribute Values ​​of Intelligent Connected Vehicles," proposes to calculate the fusion safety attribute values ​​of each system by constructing attribute matrices and judgment matrices. However, this method heavily relies on manually set linear weighting rules or the analytic hierarchy process (AHP), making it difficult to characterize the nonlinear strong coupling characteristics of risks between different safety domains in actual operation, which can easily lead to distortion in system confidence assessment results. Chinese patent application CN202211582320.3, "Anti-tipping method for unmanned operation platform in hilly and mountainous areas based on multi-sensor fusion," proposes to use a neural network model to predict the roll angle and load transfer rate and determine the control strategy accordingly. However, this algorithm assumes that the input data is absolutely reliable and lacks a mechanism to judge the reliability of the data source itself. When encountering environmental interference or malicious attacks that generate false data, the simple physical control algorithm is very likely to output incorrect control commands, which will accelerate the system instability. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-domain safety fusion method and system for vehicle stability control based on dual neural networks, which solves the problems faced by existing vehicles in complex dynamic scenarios, such as the disconnect between safety status assessment and physical control process, the difficulty in nonlinear decoupling of multi-dimensional mixed risks, and the lack of dynamic adaptive capability of control strategies to the confidence of underlying data.

[0005] The technical solution for achieving the objective of this invention is: a multi-domain safety fusion method for vehicle stability control based on dual neural networks, comprising the following steps:

[0006] Step (1): Multi-source information acquisition and preprocessing;

[0007] Step (2): Solving the multi-independent sensing mode and calculating the normalized consistency deviation feature;

[0008] Step (3): Identification of physical parameters and calculation of comprehensive stability risk value;

[0009] Step (4): Construct a cascaded topology of dual neural networks: Construct a cascaded topology containing a first neural network and a second neural network; Define the first neural network as a cross-domain security assessment network and the second neural network as a confidence coupling control network; Construct a cascaded coupling path: Directly connect the output of the first neural network to the input of the second neural network and the variable weight loss function of the second neural network to form a structural constraint of the confidence coupling control network by the cross-domain security assessment network;

[0010] Step (5): Construction of the multi-source heterogeneous feature joint space based on the first neural network and calculation of the fusion safety coefficient: Obtain functional safety status features based on the underlying diagnostic protocol and information security status features based on network packet monitoring; extract environmental interference features, and combine environmental interference features with normalized consistency deviation features to form expected functional safety dimension features; map functional safety status features, information security status features, and expected functional safety dimension features to the same numerical range to construct a multi-source heterogeneous joint feature space; input the vector of the multi-source heterogeneous joint feature space into the first neural network pre-trained based on the asymmetric cost-sensitive mechanism, extract cross-domain fault coupling features between the physical perception layer, information communication layer, and environmental interference layer, and calculate the fusion safety coefficient. ;

[0011] Step (6): Confidence-coupled adaptive control based on the second neural network.

[0012] A multi-domain safety fusion system for vehicle stability control to implement the above method includes:

[0013] Multi-source heterogeneous data acquisition and external environment information set construction module: used to acquire vehicle operating status data, local environmental perception data and external collaborative perception data; perform data validity preprocessing on external collaborative perception data, remove abnormal messages, and form an external environment information set;

[0014] The module for solving multiple independent sensing modes and calculating normalized consistency deviation is used to construct four independent information acquisition modes—roadside communication, vehicle-to-vehicle communication, high-precision map, and dynamic inference—using external environmental information sets and local environmental sensing information. It performs spatiotemporal data alignment processing on the data of the four independent information acquisition modes and solves the geometric curvature features of the road ahead in parallel. It also constructs a normalized mapping relationship between the output values ​​of the four independent information acquisition modes and calculates the normalized consistency deviation feature that characterizes the discreteness of multi-source sensing.

[0015] The physical dynamics parameter identification and adaptive calculation module for comprehensive stability risk value is used to establish longitudinal dynamic equilibrium equations based on operating status data, identify the vehicle mass in real time using the recursive least squares method, and simultaneously calculate the real-time center of gravity height. Substituting the vehicle mass and real-time center of gravity height into the attitude dynamics model, and combining lateral acceleration and roll angle data, the module calculates the basic load transfer index and attitude stability index respectively. Vertical acceleration and pitch angular velocity are introduced as correction variables to calculate an improved load transfer index incorporating vertical dynamic characteristics. Based on the vertical dynamic response data of the vibration damping device, the module calculates the road surface vertical excitation characteristics, uses the Sigmoid nonlinear function to calculate the road surface excitation confidence factor, and performs unified dimension processing on the basic load transfer index, attitude stability index, and improved load transfer index. Based on the road surface excitation confidence factor, adaptive dynamic weighted fusion is performed to output the comprehensive stability risk value.

[0016] The module for constructing a cascaded topology and constraint coupling path for dual neural networks is used to construct a cascaded topology containing a first neural network and a second neural network. The first neural network is defined as a cross-domain security assessment network, and the second neural network is defined as a confidence coupling control network. The module constructs a cascade coupling path by directly connecting the output of the first neural network to the input of the second neural network and the variable weight loss function of the second neural network, forming a structural constraint on the confidence coupling control network by the cross-domain security assessment network.

[0017] The module for constructing and fusing a multi-source heterogeneous joint feature space based on the first neural network is used to extract the functional safety status features of the underlying execution unit and the information security status features of the network message, and to combine the environmental interference features and the normalization consistency deviation features to form the expected functional safety dimension features; it constructs a multi-source heterogeneous joint feature space, inputs the vector of the multi-source heterogeneous joint feature space into the first neural network pre-trained based on the asymmetric cost-sensitive mechanism, extracts the cross-domain fault coupling features between the physical perception layer, the information communication layer and the environmental interference layer, and calculates and outputs the fused safety coefficient that represents the current comprehensive credibility of the vehicle's multiple safety domains;

[0018] The confidence-constrained PID gain dynamic calculation and target control command generation module based on the second neural network is used to calculate the real-time deviation between the comprehensive stability risk value and the preset stability target threshold; construct a control input vector containing the fused safety factor, comprehensive stability risk value and real-time deviation, and input it into the second neural network; use the fused safety factor as the confidence constraint weight, dynamically calculate the gain parameter of the PID controller through the second neural network, substitute the gain parameter and real-time deviation into the PID control algorithm to calculate the target stability control command, and send the target stability control command to the vehicle actuator to implement active intervention.

[0019] The beneficial effects of this invention are:

[0020] (1) This invention constructs four independent information acquisition modes. By calculating the normalized consistency deviation characteristics, it performs in-depth physical and information cross-fusion comparison of multi-source information. When encountering local network tampering or sensor blinding, this mechanism can quantify the discreteness of the multi-source sensing system in real time, effectively avoiding absolute dependence on a single vulnerable information source.

[0021] (2) The present invention introduces a road excitation confidence factor based on vertical dynamic response and performs adaptive dynamic fusion of basic load transfer index, attitude stability index and improved index. This method can automatically identify road bump interference and filter out high-frequency noise, overcoming the limitation of traditional single fixed parameter algorithm that is prone to induce rollover "false alarm" on unstructured road surface.

[0022] (3) This invention constructs a multi-source heterogeneous joint feature space and uses an asymmetric cost-sensitive first neural network to perform feature dimensionality reduction and fusion. By artificially amplifying the loss gradient of the missed samples, the network is driven to converge towards a high-sensitivity direction, breaking the hysteresis of traditional linear superposition evaluation and establishing an agile fusion evaluation mechanism for minor faults and cross-domain covert attacks.

[0023] (4) This invention constructs a dual neural network cascade coupling path, directly embedding the calculated fusion safety coefficient into the variable weight loss function of the second neural network. When the overall confidence of the data decreases, this fusion mechanism automatically assigns a high penalty weight to the maximum overshoot term, forcing the controller to output conservative PID gain parameters, thus completely breaking down the fusion safety defense line from the upper multi-domain state to the lower actuator. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall workflow of the method described in this application.

[0025] Figure 2 This is a schematic diagram of the overall architecture of the method described in this application.

[0026] Figure 3 This is a schematic diagram comparing the time-series changes of multi-source stability risk indicators and existing technical indicators in an embodiment of the present invention.

[0027] Figure 4 This is a schematic diagram illustrating the temporal changes of the first neural network safety assessment and the vehicle's controlled speed response in an embodiment of the present invention; wherein, Figure 4 In the figure, (a) is the time-series variation curve of the fusion security coefficient output by the first neural network. Figure 4 (b) in the figure represents the corresponding longitudinal controlled speed response curve of the vehicle.

[0028] Figure 5This is a schematic diagram illustrating the timing changes of the input state and adaptive PID control gain parameters of the second neural network in an embodiment of the present invention; wherein, Figure 5 In the figure, (a) is the time-series variation curve of the overall stability risk and fusion safety coefficient of the second neural network input layer. Figure 5 (b) in the figure is the time-series curve of the adaptive PID gain adjustment factor of the output layer of the second neural network. Detailed Implementation

[0029] The present invention will now be described in further detail with reference to the accompanying drawings.

[0030] A multi-domain safety fusion method for vehicle stability control based on dual neural networks includes the following steps:

[0031] Step (1): Multi-source information collection and preprocessing: acquire vehicle operating status data, local environmental perception data and external collaborative perception data; perform data validity preprocessing on the external collaborative perception data, remove abnormal messages, and form an external environmental information set;

[0032] Step (2): Multi-independent sensing mode solution and normalized consistency deviation feature calculation: Using the external environment information set and local environment sensing information, four independent information acquisition modes are constructed: roadside communication, vehicle-to-vehicle communication, high-precision map and dynamic inference. The geometric curvature features of the road ahead are solved in parallel. The normalized mapping relationship between the output values ​​of the four modes is constructed, and the normalized consistency deviation feature representing the discreteness of multi-source sensing is calculated.

[0033] Step (3): Physical parameter identification and comprehensive stability risk value calculation: Based on the operating status data, establish the longitudinal dynamic equilibrium equation, use the recursive least squares method to identify the vehicle mass in real time, and simultaneously calculate the real-time center of gravity height; substitute the vehicle mass and real-time center of gravity height into the attitude dynamic model, and combine lateral acceleration and roll angle to solve the basic load transfer index and attitude stability index respectively; introduce vertical acceleration and pitch angular velocity as correction variables to calculate the improved load transfer index containing vertical dynamic characteristics; calculate the road vertical excitation characteristics based on the vertical dynamic response data of the vibration damping device, use the Sigmoid nonlinear function to calculate the road excitation confidence factor, and perform unified dimension processing on the basic load transfer index, attitude stability index, and improved load transfer index; perform adaptive dynamic weighted fusion based on the road excitation confidence factor to output the final comprehensive stability risk value. ;

[0034] Step (4): Construct a cascaded topology of dual neural networks: Construct a cascaded topology containing a first neural network and a second neural network; Define the first neural network as a cross-domain security assessment network and the second neural network as a confidence coupling control network; Construct a cascaded coupling path: Directly connect the output of the first neural network to the input of the second neural network and the variable weight loss function of the second neural network to form a structural constraint of the confidence coupling control network by the cross-domain security assessment network;

[0035] Step (5): Construction of the multi-source heterogeneous feature joint space based on the first neural network and calculation of the fusion safety coefficient: Obtain functional safety status features based on the underlying diagnostic protocol and information security status features based on network packet monitoring; extract environmental interference features, and combine environmental interference features with normalized consistency deviation features to form expected functional safety dimension features; map functional safety status features, information security status features, and expected functional safety dimension features to the same numerical range to construct a multi-source heterogeneous joint feature space; input the vector of the multi-source heterogeneous joint feature space into the first neural network pre-trained based on the asymmetric cost-sensitive mechanism, extract cross-domain fault coupling features between the physical perception layer, information communication layer, and environmental interference layer, and calculate the fusion safety coefficient. ;

[0036] Step (6): Confidence-coupled adaptive control based on the second neural network: Calculate the real-time deviation between the comprehensive stability risk value and the preset stability target threshold; construct a control input vector containing the fusion safety factor, the comprehensive stability risk value and the real-time deviation, and input the control input vector into the second neural network constructed based on the confidence-coupled strategy; use the fusion safety factor as the confidence constraint weight, dynamically solve the gain parameter of the PID controller through the second neural network, substitute the gain parameter and the real-time deviation into the PID control algorithm, calculate the target stability control command, and send it to the vehicle actuator to implement intervention.

[0037] Furthermore, the specific process of step (2) includes:

[0038] Step (21): Solving the geometric curvature features of the road ahead using multiple independent sensing modes: running four independent information acquisition modes in parallel—roadside communication, vehicle-to-vehicle communication, high-precision map, and dynamic inference—to obtain the equivalent curvature value of the road ahead; where:

[0039] The roadside communication mode is as follows: parsing the electronic road sign data frames broadcast from the roadside unit in the external environment information set, and extracting the curvature values ​​from the road geometry field. ;

[0040] The vehicle-to-vehicle communication mode is as follows: parsing historical trajectory data of nearby vehicles from the external environment information set, performing curve fitting using the least squares method, and back-calculating the geometric curvature of the road ahead. ;

[0041] The high-precision map mode involves querying the onboard high-precision map database based on the vehicle's positioning coordinates and matching it with the preset curvature information of the current lane line. ;

[0042] The dynamic self-inference mode is based on the vehicle yaw rate from the vehicle's operating state data. and longitudinal movement rate Real-time curvature was calculated using kinematic formulas. : .

[0043] Step (22): Calculation of normalized consistency deviation feature: Construct a normalized mapping relationship between the output values ​​of the four independent information acquisition modes, and calculate the normalized consistency deviation feature between the output values ​​of the four modes. :

[0044] ,

[0045] in, Number of patterns; For the first The curvature value output by each mode; The arithmetic mean of the curvature of the four modes; A preset curvature normalization constant is used to map the deviation characteristics to... The standardized numerical range; Used to quantify the overall reliability dispersion of current multi-source sensing systems.

[0046] Furthermore, step (3) specifically includes the following steps:

[0047] Step (31): Identification of physical parameters based on dynamic equilibrium equations: Based on the operating state data obtained in step (1), a longitudinal dynamic equilibrium equation is established as a physical constraint model for parameter estimation:

[0048] ,

[0049] in, For vehicle driving force, , , These are rolling resistance, air resistance, and slope resistance, respectively. For longitudinal acceleration, The vehicle weight; obtained through RLS parameter estimation;

[0050] Subsequently, a model for calculating the height of the center of mass was established based on the principle of tilt moment balance:

[0051] ,

[0052] in, For roll stiffness, The height of the roll center. The roll angle is... It is lateral acceleration. The real-time height of the center of mass is obtained by inverse kinematics, which is the acceleration due to gravity. ;

[0053] Step (32): Calculation of basic load transfer index, attitude stability index and improved load transfer index: Call the preset vehicle wheelbase. and pitch moment of inertia Combined with the overall vehicle weight With real-time centroid height Calculate the basic load transfer index, denoted as LTR, and the attitude stability index, denoted as RI; introduce the vehicle body vertical acceleration. With pitch angular velocity Calculate the improved load transfer index, denoted as MLTR, which incorporates vertical dynamic characteristics:

[0054] ,

[0055] in, For pitch rotation inertia, The wheelbase is the distance between the wheels. The yaw rate is angular velocity. It is the acceleration due to gravity;

[0056] Introducing real-time road surface adhesion coefficient Construct a dynamic normalization correction formula:

[0057] ,

[0058] in, Represents any original calculated value from the basic load transfer index LTR, attitude stability index RI, or improved load transfer index MLTR; This represents the corresponding dimensionless index after normalization correction; Preset stability thresholds for each indicator. The pavement adhesion correction factor is used; LTR, RI, and MLTR are processed separately to obtain the corrected dimensionless index. and ;

[0059] Step (33): Adaptive calculation of road excitation confidence factor and comprehensive stability risk value: Calculate the variance of the vertical dynamic response data of the vibration reduction device within the preset sliding time window. The Sigmoid function is used to construct the road surface excitation confidence factor. :

[0060] ,

[0061] in, All are preset adjustment coefficients that are greater than zero; The shape mapping parameters are used to determine the sensitivity of road surface smoothness recognition. The variance benchmark threshold characterizing the boundary of rough road surface determination; road surface excitation confidence factor. Used to characterize the degree of unstructured road surface A larger value indicates a more significant road surface excitation; based on confidence factor. A strategy combining adaptive weighting and maximum fallback is used to calculate the comprehensive stability risk value. :

[0062] ;

[0063] .

[0064] Furthermore, step (4) specifically includes the following steps:

[0065] Step (41): Configure the first neural network as a three-layer feedforward neural network containing a first input layer, a first hidden layer and a first output layer; configure multiple first input neurons in the first input layer, and set the number of multiple first input neurons to be equal to the dimension of the feature vector of the multi-source heterogeneous joint feature space, and use multiple first input neurons to receive the feature vector of the multi-source heterogeneous joint feature space; configure the first hidden layer with a hyperbolic tangent activation function, and use the hyperbolic tangent activation function to extract the cross-domain fault coupling features between the physical perception layer, the information communication layer and the environmental interference layer; configure the first output layer with a single output neuron and a Sigmoid activation function, and use the single output neuron and the Sigmoid activation function to map the cross-domain fault coupling features to a continuous numerical range from zero to one, and output the fusion safety coefficient;

[0066] Step (42): Configure the second neural network as a three-layer backpropagation neural network containing a second input layer, a second hidden layer, and a second output layer; configure multiple second input neurons in the second input layer, setting the number of multiple second input neurons to be equal to the sum of the dimensions of the fusion safety factor, the comprehensive stability risk value, and the real-time deviation, and use multiple second input neurons to receive the control input vector containing the fusion safety factor, the comprehensive stability risk value, and the real-time deviation; configure the hyperbolic tangent activation function in the second hidden layer to fit the nonlinear control surface; configure three parallel output neurons in the second output layer, and use the three parallel output neurons to output the proportional gain, integral gain, and derivative gain of the PID controller respectively;

[0067] Step (43): Construction of constraint coupling path: Establish data flow coupling channel and control flow coupling channel between the first neural network and the second neural network; The data flow coupling channel is constructed by directly hard-wiring the first output layer of the first neural network to the second input layer of the second neural network, so that the fusion safety factor becomes the pre-input variable of the feedforward calculation of the second neural network; The control flow coupling channel is constructed by directly connecting the first output layer of the first neural network to the variable weight loss function expression of the second neural network, so that the fusion safety factor becomes the independent variable factor for dynamically adjusting the weight of the error integral term and the weight of the maximum overshoot term in the variable weight loss function; A dual structural constraint of the first neural network on the second neural network is formed through the data flow coupling channel and the control flow coupling channel.

[0068] Furthermore, step (5) specifically includes the following steps:

[0069] Step (51): Extraction of multi-source heterogeneous state features and construction of multi-source heterogeneous input feature vectors: Based on the underlying diagnostic protocol, extract the signal-to-noise ratio and frame loss rate of the underlying execution unit as functional safety state features; based on network packet monitoring, extract the timestamp anomaly frequency and digital signature verification failure rate of external cooperative communication packets as information security state features; extract the ambient light intensity from the external environment information set. With rain and fog obstruction As environmental disturbance characteristics; the environmental disturbance characteristics and the normalized consistency deviation characteristics calculated in step (2) together constitute the expected functional safety dimension characteristics; the maximum-minimum normalization algorithm is used to uniformly map the functional safety state characteristics, information security state characteristics and expected functional safety dimension characteristics to The dimensionless intervals are spliced ​​together to construct multi-source heterogeneous input feature vectors;

[0070] Step (52): Construction of the asymmetric cost-sensitive loss function and directional optimization training of the first neural network: Input the multi-source heterogeneous input feature vectors into the first neural network; use the asymmetric cost-sensitive loss function as the objective function for offline training of the first neural network; the asymmetric cost-sensitive loss function is defined as a variant of the cross-entropy function, and the specific formula of the asymmetric cost-sensitive loss function is as follows:

[0071] ,

[0072] in, Define the total number of training samples. Define a true safety label Define the probability predicted by the first neural network. A security penalty factor is set for missed detection samples where the predicted probability of the first neural network is greater than the actual security label during the offline training process of the first neural network. The loss gradient of the missed samples is artificially amplified by using a safety penalty factor. The weight matrix and bias vector of the first neural network are iteratively updated by using the backpropagation algorithm based on the gradient descent direction of the asymmetric cost-sensitive loss function, thereby completing the targeted optimization training of the first neural network.

[0073] Step (53): Cross-domain fault coupling feature extraction and fusion safety factor calculation: In the real-time inference stage of the first neural network, the multi-source heterogeneous input feature vector generated in real time is input into the first neural network after directional optimization training. The cross-domain fault coupling features between the physical perception layer, the information communication layer and the environmental interference layer are extracted using the first hidden layer of the first neural network. The scalar value representing the current multi-safety domain comprehensive credibility of the vehicle is output in the first output layer of the first neural network and is defined as the fusion safety factor. .

[0074] Furthermore, step (6) specifically includes the following steps:

[0075] Step (61): Control input vector construction and second neural network variable weight loss function setting: Calculate the real-time deviation between the comprehensive stability risk value and the preset stability target threshold. and its rate of change Integrating safety factor Comprehensive stability risk value Real-time deviation and rate of change The control input vector is constructed by combining the two components; a variable weight loss function based on confidence constraints is used as the objective function for offline training of the second neural network; the variable weight loss function based on confidence constraints is defined as the dynamic weighted sum of the error integral term and the maximum overshoot term; the specific formula of the variable weight loss function based on confidence constraints is as follows:

[0076] ,

[0077] in, Weighted by response speed For stability weights, The overshoot is defined as the step response of the vehicle's physical motion. The response speed weight and stability weight are both constraint functions with the fused safety factor as the independent variable.

[0078] Step (62): Execution of dynamic weight allocation strategy and targeted optimization training of the second neural network: During the offline training of the second neural network, a pre-position confidence safety threshold is set. The system executes a dynamic weight allocation strategy based on the fusion safety coefficient in the control input vector. This strategy includes a low-confidence defense mode and a high-confidence response mode. In the low-confidence defense mode, when the input sample satisfies a fusion safety coefficient less than the pre-positional confidence safety threshold, the stability weight is set to at least 10 times the response speed weight (i.e.,...). The algorithm utilizes a high penalty in a variable-weight loss function based on confidence constraints to artificially amplify the calculation error of overshoot. It then drives the weight matrix and bias vector of the second neural network to iteratively update in the direction of suppressing overshoot through backpropagation. In high-confidence response mode, when the input sample satisfies a fusion safety coefficient greater than or equal to the pre-position confidence safety threshold, the response speed weight is set greater than the stability weight (i.e.,...). The weight matrix and bias vector of the second neural network are iteratively updated in the direction of eliminating real-time bias through the backpropagation algorithm, thus completing the targeted optimization training of the second neural network.

[0079] Step (63): Adaptive gain parameter inference and target stability control command generation: In the real-time inference stage of the second neural network, the real-time generated control input vector is input into the second neural network after directional optimization training, and the proportional gain of the PID controller is output in real time by the three parallel output neurons of the second neural network. Integral gain and differential gain ; to increase the proportional gain Integral gain and differential gain Substitute the values ​​into the incremental PID control algorithm to calculate the target stability control command. The calculation formula for the incremental PID control algorithm is:

[0080] ,

[0081] in, This is the target stability control command from the previous sampling time. and These are the real-time deviations at the first two sampling times; the target stability control command. The signal is sent to the vehicle actuators to implement braking intervention, completing adaptive fusion control covering four major areas: information security, functional safety, expected functional safety, and active safety.

[0082] A multi-domain safety fusion system for vehicle stability control to implement the above method includes:

[0083] Multi-source heterogeneous data acquisition and external environment information set construction module: used to acquire vehicle operating status data, local environmental perception data and external collaborative perception data; perform data validity preprocessing on external collaborative perception data, remove abnormal messages, and form an external environment information set;

[0084] The module for solving multiple independent sensing modes and calculating normalized consistency deviation is used to construct four independent information acquisition modes—roadside communication, vehicle-to-vehicle communication, high-precision map, and dynamic inference—using external environmental information sets and local environmental sensing information. It performs spatiotemporal data alignment processing on the data of the four independent information acquisition modes and solves the geometric curvature features of the road ahead in parallel. It also constructs a normalized mapping relationship between the output values ​​of the four independent information acquisition modes and calculates the normalized consistency deviation feature that characterizes the discreteness of multi-source sensing.

[0085] The physical dynamics parameter identification and adaptive calculation module for comprehensive stability risk value is used to establish longitudinal dynamic equilibrium equations based on operating status data, identify the vehicle mass in real time using the recursive least squares method, and simultaneously calculate the real-time center of gravity height. Substituting the vehicle mass and real-time center of gravity height into the attitude dynamics model, and combining lateral acceleration and roll angle data, the module calculates the basic load transfer index and attitude stability index respectively. Vertical acceleration and pitch angular velocity are introduced as correction variables to calculate an improved load transfer index incorporating vertical dynamic characteristics. Based on the vertical dynamic response data of the vibration damping device, the module calculates the road surface vertical excitation characteristics, uses the Sigmoid nonlinear function to calculate the road surface excitation confidence factor, and performs unified dimension processing on the basic load transfer index, attitude stability index, and improved load transfer index. Based on the road surface excitation confidence factor, adaptive dynamic weighted fusion is performed to output the comprehensive stability risk value.

[0086] The module for constructing a cascaded topology and constraint coupling path for dual neural networks is used to construct a cascaded topology containing a first neural network and a second neural network. The first neural network is defined as a cross-domain security assessment network, and the second neural network is defined as a confidence coupling control network. The module constructs a cascade coupling path by directly connecting the output of the first neural network to the input of the second neural network and the variable weight loss function of the second neural network, forming a structural constraint on the confidence coupling control network by the cross-domain security assessment network.

[0087] The module for constructing and fusing a multi-source heterogeneous joint feature space based on the first neural network is used to extract the functional safety status features of the underlying execution unit and the information security status features of the network message, and to combine the environmental interference features and the normalization consistency deviation features to form the expected functional safety dimension features; it constructs a multi-source heterogeneous joint feature space, inputs the vector of the multi-source heterogeneous joint feature space into the first neural network pre-trained based on the asymmetric cost-sensitive mechanism, extracts the cross-domain fault coupling features between the physical perception layer, the information communication layer and the environmental interference layer, and calculates and outputs the fused safety coefficient that represents the current comprehensive credibility of the vehicle's multiple safety domains;

[0088] The confidence-constrained PID gain dynamic calculation and target control command generation module based on the second neural network is used to calculate the real-time deviation between the comprehensive stability risk value and the preset stability target threshold; construct a control input vector containing the fused safety factor, comprehensive stability risk value and real-time deviation, and input it into the second neural network; use the fused safety factor as the confidence constraint weight, dynamically calculate the gain parameter of the PID controller through the second neural network, substitute the gain parameter and real-time deviation into the PID control algorithm to calculate the target stability control command, and send the target stability control command to the vehicle actuator to implement active intervention.

[0089] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the method defined above.

[0090] Example

[0091] like Figure 1 As shown, this invention provides a vehicle multi-domain safety fusion stability control method based on dual neural networks. Its core execution flow includes: S1 multi-source information acquisition and preprocessing; S2 multi-independent sensing mode calculation and normalized consistency deviation calculation; S3 physical dynamic parameter identification and comprehensive stability risk value calculation; S4 construction of a dual neural network cascaded topology and constraint coupling path; S5 construction of a multi-source heterogeneous joint feature space based on the first neural network and calculation of the fusion safety factor; S6 execution of confidence-coupled dynamic PID gain based on the second neural network and generation of target stability control commands. The specific mathematical models, formula derivations, and logical constraints involved in each step are consistent with the core principles described in the foregoing invention content section. For brevity, repetitive derivations of basic mathematical formulas are omitted here. The actual parameter configuration and control effect of the system under complex working conditions will be verified and explained in detail in the following specific embodiments.

[0092] like Figure 2As shown, the present invention also provides a multi-domain safety fusion system for vehicle stability control to implement the above method. This system is integrated into the chassis domain controller and its main modules include: a multi-source heterogeneous data acquisition and external environment information set construction module; a multi-independent perception mode solution and normalized consistency deviation calculation module; a physical dynamics parameter identification and comprehensive stability risk value adaptive calculation module; a dual neural network cascaded topology and constraint coupling path construction module; and safety factor calculation and target control command generation modules based on the first and second neural networks, respectively. Finally, the control commands are sent to the vehicle actuators to implement active intervention.

[0093] This invention constructs a dual neural network topology that includes a "cross-domain safety assessment" neural network and a "confidence-coupled control" neural network, and builds a multi-dimensional heterogeneous joint feature space covering functional safety, expected functional safety, information security and physical stability (active safety). By using an asymmetric cost-sensitive mechanism and a confidence-coupled strategy, it achieves adaptive dynamic tuning and stability decision-making for control parameters of complex mobile platforms (such as commercial vehicles or special equipment).

[0094] Example

[0095] To verify the control performance of the method described in this invention under complex working conditions, this embodiment selects a double lane change emergency obstacle avoidance scenario for testing. This scenario includes two sharp steering operations, causing the vehicle to exhibit a significant tendency to become unstable. The test environment was set as follows: road surface adhesion coefficient of 0.85 (dry asphalt road surface), and initial vehicle speed of 90 km / h. During the test, the vehicle faced complex perception environment challenges, and the system performed multi-safety domain fusion control through a dual neural network architecture.

[0096] In this embodiment, a dual neural network architecture was constructed and its parameters were calibrated. The specific neural network training hyperparameters are shown in Table 1, and the basic parameter settings of the adaptive controller are shown in Table 2.

[0097] Table 1. Hyperparameters for training the dual neural network model in the example.

[0098]

[0099] Table 2 Basic Parameters of Adaptive Controller in Embodiments

[0100]

[0101] After simulation, the embodiments of this invention output a performance comparison curve of multi-source stability risk indicators. Figure 3 ), First Neural Network Safety Assessment and Vehicle Speed ​​Response Curve ( Figure 4 ), and the second neural network input state and adaptive PID control gain curve ( Figure 5 ).

[0102] Figure 3 The temporal characteristics of different stability indices under the double lane change condition were compared. In the figure, the gray solid line (traditional LTR) exhibits significant high-frequency oscillation noise due to road surface excitation; the black dashed line (traditional RI) shows a physical response lag of approximately 0.4 seconds; while the black dotted line (improved MLTR) effectively smooths out road surface noise. The thick black solid line represents the comprehensive stability risk value generated in this embodiment. This curve combines the steady-state accuracy of MLTR with the dynamic sensitivity of RI, providing a smooth and hysteresis-free risk characterization at the steering peak times of 3.5s and 6.5s. Figure 4 This demonstrates safety assessments and vehicle speed response under a SOTIF (Safety Intended for Functional Context) scenario. For example... Figure 4 As shown in (a), the system is injected with environmental disturbances during the period from 3s to 5s, and the fusion safety factor ( The vehicle accurately dropped to the 0.45 range. It's worth noting the 4-5 second range; during this time, the vehicle is in a state of recovery, and although the risk of physical instability has decreased, environmental interference continues. The system correctly maintained a low position, verifying its ability to decouple and evaluate "physical stability" and "perceived credibility." Figure 4 As shown in (b), the corresponding vehicle speed response is as follows: In the low confidence range (3-5s), the system implemented an active speed reduction strategy, smoothly reducing the vehicle speed from 90 km / h to approximately 55 km / h and maintaining cruising speed. This strategy increases stability margin through speed reduction while avoiding the risk of rear-end collisions caused by blind braking. Figure 5 The dynamic adjustment process of control parameters based on a second neural network is demonstrated. For example... Figure 5 As shown in (a), this represents the core input state of the neural network; as... Figure 5 As shown in (b), this represents the PID gain adjustment factor of the output. , , As can be seen, in the low safety factor range of 3s to 5s, despite the higher stability risk, the control gain ratios are still strongly suppressed by the variable weight loss function (the peak value is limited to around 3.2), effectively preventing over-braking due to perception bias. Here, the output value is the normalized adjustment coefficient, and the actual gain is the product of this coefficient and the base value in Table 2. When the interference disappears after 5s, facing a second emergency steering, the system is in a high-confidence state, and the gain adjustment ratio is quickly unlocked and surges to over 4.8, ensuring sufficient physical correction torque. This process confirms the adaptive error prevention effect of the dual constraint mechanism of data flow and control flow in the dual neural network cascade architecture.

Claims

1. A dual neural network based multi-domain safety fusion method for vehicle stability control, characterized in that, Includes the following steps: Step (1): Multi-source information acquisition and preprocessing: Acquire vehicle operating status data, local environmental perception data and external collaborative perception data; Perform data validity preprocessing on external collaborative sensing data, remove abnormal messages, and form an external environment information set; Step (2): Multi-independent sensing mode solution and normalized consistency deviation feature calculation: Using the external environment information set and local environment sensing information, four independent information acquisition modes are constructed: roadside communication, vehicle-to-vehicle communication, high-precision map and dynamic inference, and the geometric curvature features of the road ahead are solved in parallel. Construct a normalized mapping relationship between the output values ​​of the four modes, and calculate the normalized consistency deviation feature that characterizes the discreteness of multi-source perception; Step (3): Physical parameter identification and comprehensive stability risk value calculation: Based on the operating status data, establish the longitudinal dynamic equilibrium equation, use the recursive least squares method to identify the vehicle mass in real time, and calculate the real-time center of gravity height simultaneously; The vehicle mass and real-time center of gravity height are substituted into the attitude dynamics model. The basic load transfer index and attitude stability index are calculated by combining lateral acceleration and roll angle. Vertical acceleration and pitch rate are introduced as correction variables to calculate the improved load transfer index that includes vertical dynamic characteristics. The vertical excitation characteristics of the road surface are calculated based on the vertical dynamic response data of the vibration reduction device. The road excitation confidence factor is calculated using the Sigmoid nonlinear function. The base load transfer index, attitude stability index, and improved load transfer index are processed with unified dimensions. Adaptive dynamic weighted fusion is performed based on the road excitation confidence factor to output the final comprehensive stability risk value. ; Step (4): Construct a cascaded topology of dual neural networks: Construct a cascaded topology containing a first neural network and a second neural network; Define the first neural network as a cross-domain security assessment network and the second neural network as a confidence coupling control network; Construct a cascaded coupling path: Directly connect the output of the first neural network to the input of the second neural network and the variable weight loss function of the second neural network to form a structural constraint of the confidence coupling control network by the cross-domain security assessment network; Step (5): Construction and fusion of the multi-source heterogeneous joint feature space based on the first neural network and calculation of the security coefficient: obtain functional safety status features based on the underlying diagnostic protocol and obtain information security status features based on network packet monitoring; Environmental disturbance features are extracted and combined with normalized consistency deviation features to form the expected functional safety dimension features. Functional safety status features, information security status features, and expected functional safety dimension features are mapped to the same numerical range to construct a multi-source heterogeneous joint feature space. The vector input of the multi-source heterogeneous joint feature space is then used to input a first neural network pre-trained based on an asymmetric cost-sensitive mechanism to extract cross-domain fault coupling features between the physical perception layer, information communication layer, and environmental interference layer. Finally, a fusion safety coefficient representing the current comprehensive reliability of the vehicle across multiple safety domains is calculated. ; Step (6): Confidence-coupled adaptive control based on the second neural network: calculate the real-time deviation between the comprehensive stability risk value and the preset stability target threshold; construct a control input vector containing the fused safety factor, the comprehensive stability risk value and the real-time deviation, and input the control input vector into the second neural network constructed based on the confidence-coupled strategy; The safety factor is used as the confidence constraint weight. The gain parameter of the PID controller is dynamically calculated through the second neural network. The gain parameter and real-time deviation are substituted into the PID control algorithm to calculate the target stability control command and send it to the vehicle actuator to implement intervention.

2. The method of claim 1, wherein, Step (2) specifically includes the following steps: Step (21): Solve the geometric curvature features of the road ahead using multiple independent sensing modes: The system operates in parallel with four independent information acquisition modes: roadside communication, vehicle-to-vehicle communication, high-precision mapping, and dynamic inference, to obtain the equivalent curvature value of the road ahead; among which: The roadside communication mode is: analyzing the electronic sign data frame broadcast by the roadside unit in the external environment information set, extracting the curvature value in the road geometry field ; The vehicle-to-vehicle communication mode is as follows: parsing historical trajectory data of nearby vehicles from the external environment information set, performing curve fitting using the least squares method, and back-calculating the geometric curvature of the road ahead. ; The high-precision map mode involves querying the onboard high-precision map database based on the vehicle's positioning coordinates and matching it with the preset curvature information of the current lane line. ; The dynamic self-inference mode is based on the vehicle yaw rate from the vehicle's operating state data. and longitudinal movement rate Real-time curvature was calculated using kinematic formulas. : ; Step (22): Calculation of normalized consistency deviation feature: Construct a normalized mapping relationship between the output values ​​of the four independent information acquisition modes, and calculate the normalized consistency deviation feature between the output values ​​of the four modes. : ， in, Number of patterns; For the first The curvature value output by each mode; The arithmetic mean of the curvature of the four modes; A preset curvature normalization constant is used to map the deviation characteristics to... The standardized numerical range; Used to quantify the overall reliability dispersion of current multi-source sensing systems.

3. The method of claim 2, wherein, Step (3) specifically includes the following steps: Step (31): Identification of physical parameters based on dynamic equilibrium equations: Based on the operating state data obtained in step (1), a longitudinal dynamic equilibrium equation is established as a physical constraint model for parameter estimation: ， in, For vehicle driving force, , , These are rolling resistance, air resistance, and slope resistance, respectively. For longitudinal acceleration, The vehicle weight; obtained through RLS parameter estimation; Subsequently, a model for calculating the height of the center of mass was established based on the principle of tilt moment balance: ， in, For roll stiffness, The height of the roll center. The roll angle is... It is lateral acceleration. The real-time height of the center of mass is obtained by inverse kinematics, which is the acceleration due to gravity. ; Step (32): Calculation of basic load transfer index, attitude stability index and improved load transfer index: Call the preset vehicle wheelbase. and pitch moment of inertia Combined with the overall vehicle weight With real-time centroid height Calculate the basic load transfer index and attitude stability index, denoted as LTR and RI respectively; introduce the vehicle body vertical acceleration. With pitch angular velocity Calculate the improved load transfer index, denoted as MLTR, which incorporates vertical dynamic characteristics: ， in, For pitch rotation inertia, The wheelbase is the distance between the wheels. The yaw rate is angular velocity. It is the acceleration due to gravity; Introducing real-time road adhesion coefficient , and constructing dynamic normalization correction formula: ， in, Represents any original calculated value from the basic load transfer index LTR, attitude stability index RI, or improved load transfer index MLTR; This represents the corresponding dimensionless index after normalization correction; Preset stability thresholds for each indicator. The pavement adhesion correction factor is used; LTR, RI, and MLTR are processed separately to obtain the corrected dimensionless index. and ; Step (33): Adaptive calculation of road excitation confidence factor and comprehensive stability risk value: Calculate the variance of the vertical dynamic response data of the vibration reduction device within the preset sliding time window. The Sigmoid function is used to construct the road surface excitation confidence factor. : ， in, All are preset adjustment coefficients that are greater than zero; The shape mapping parameters are used to determine the sensitivity of road surface smoothness recognition. The variance benchmark threshold characterizing the boundary of rough road surface determination; road surface excitation confidence factor. Used to characterize the degree of unstructured road surface A larger value indicates a more significant road surface excitation; based on confidence factor. A strategy combining adaptive weighting and maximum fallback is used to calculate the comprehensive stability risk value. : ; 。 4. The method of claim 3, wherein, Step (4) specifically includes the following steps: Step (41): Configure the first neural network as a three-layer feedforward neural network containing a first input layer, a first hidden layer and a first output layer; configure multiple first input neurons in the first input layer, and set the number of multiple first input neurons to be equal to the dimension of the feature vector of the multi-source heterogeneous joint feature space, and use multiple first input neurons to receive the feature vector of the multi-source heterogeneous joint feature space; configure the first hidden layer with a hyperbolic tangent activation function, and use the hyperbolic tangent activation function to extract the cross-domain fault coupling features between the physical perception layer, the information communication layer and the environmental interference layer; configure the first output layer with a single output neuron and a Sigmoid activation function, and use the single output neuron and the Sigmoid activation function to map the cross-domain fault coupling features to a continuous numerical range from zero to one, and output the fusion safety coefficient; Step (42): Configure the second neural network as a three-layer backpropagation neural network containing a second input layer, a second hidden layer, and a second output layer; configure multiple second input neurons in the second input layer, setting the number of multiple second input neurons to be equal to the sum of the dimensions of the fusion safety factor, the comprehensive stability risk value, and the real-time deviation, and use multiple second input neurons to receive the control input vector containing the fusion safety factor, the comprehensive stability risk value, and the real-time deviation; configure the hyperbolic tangent activation function in the second hidden layer to fit the nonlinear control surface; configure three parallel output neurons in the second output layer, and use the three parallel output neurons to output the proportional gain, integral gain, and derivative gain of the PID controller respectively; Step (43): Constraint coupling path construction: Establish data flow coupling channels and control flow coupling channels between the first neural network and the second neural network; The data flow coupling channel is constructed as follows: the first output layer of the first neural network is directly hardwired to the second input layer of the second neural network, making the fusion safety factor a pre-input variable for the feedforward calculation of the second neural network; the control flow coupling channel is constructed as follows: the first output layer of the first neural network is directly connected to the variable weight loss function expression of the second neural network, making the fusion safety factor an independent variable factor for dynamically adjusting the weight of the error integral term and the weight of the maximum overshoot term in the variable weight loss function; the data flow coupling channel and the control flow coupling channel form a dual structural constraint of the first neural network on the second neural network.

5. The method of claim 4, wherein, Step (5) specifically includes the following steps: Step (51): Extraction of multi-source heterogeneous state features and construction of multi-source heterogeneous input feature vectors: Extract the signal-to-noise ratio and frame drop rate of the underlying execution unit as functional safety state features based on the underlying diagnostic protocol; Based on network packet monitoring, the frequency of timestamp anomalies and the digital signature verification failure rate of external collaborative communication packets are extracted as information security status characteristics; ambient light intensity is extracted from external environmental information sets. With rain and fog obstruction As environmental disturbance characteristics; the environmental disturbance characteristics and the normalized consistency deviation characteristics calculated in step (2) together constitute the expected functional safety dimension characteristics; the maximum-minimum normalization algorithm is used to uniformly map the functional safety state characteristics, information security state characteristics and expected functional safety dimension characteristics to The dimensionless intervals are spliced ​​together to construct multi-source heterogeneous input feature vectors; Step (52): Construction of the asymmetric cost-sensitive loss function and directional optimization training of the first neural network: Input the multi-source heterogeneous input feature vectors into the first neural network; use the asymmetric cost-sensitive loss function as the objective function for offline training of the first neural network; the asymmetric cost-sensitive loss function is defined as a variant of the cross-entropy function, and the specific formula of the asymmetric cost-sensitive loss function is as follows: ， in, Define the total number of training samples. Define a true safety label Define the probability predicted by the first neural network. A security penalty factor is set during the offline training of the first neural network for false negative samples where the predicted probability of the first neural network is greater than the actual security label. The loss gradient of the missed samples is artificially amplified by using a safety penalty factor. The weight matrix and bias vector of the first neural network are iteratively updated by using the backpropagation algorithm based on the gradient descent direction of the asymmetric cost-sensitive loss function, thereby completing the targeted optimization training of the first neural network. Step (53): Cross-domain fault coupling feature extraction and fusion safety factor calculation: In the real-time inference stage of the first neural network, the multi-source heterogeneous input feature vector generated in real time is input into the first neural network after directional optimization training. The cross-domain fault coupling features between the physical perception layer, the information communication layer and the environmental interference layer are extracted using the first hidden layer of the first neural network. The scalar value representing the current multi-safety domain comprehensive credibility of the vehicle is output in the first output layer of the first neural network and is defined as the fusion safety factor. .

6. The method of claim 5, wherein, Step (6) specifically includes the following steps: Step (61): Control input vector construction and second neural network variable weight loss function setting: Calculate the real-time deviation between the comprehensive stability risk value and the preset stability target threshold. and its rate of change Integrating safety factor Comprehensive stability risk value Real-time deviation and rate of change The control input vector is constructed by combining the two components; a variable weight loss function based on confidence constraints is used as the objective function for offline training of the second neural network; the variable weight loss function based on confidence constraints is defined as the dynamic weighted sum of the error integral term and the maximum overshoot term; the specific formula of the variable weight loss function based on confidence constraints is as follows: ， in, Weighted by response speed For stability weights, The overshoot is defined as the step response of the vehicle's physical motion. The response speed weight and stability weight are both constraint functions with the fused safety factor as the independent variable. Step (62): Execution of dynamic weight allocation strategy and targeted optimization training of the second neural network: During the offline training of the second neural network, a pre-position confidence safety threshold is set. The system executes a dynamic weight allocation strategy based on the fusion safety coefficient in the control input vector. This strategy includes a low-confidence defense mode and a high-confidence response mode. In the low-confidence defense mode, when the input sample satisfies a fusion safety coefficient less than the pre-positional confidence safety threshold, the stability weight is set to at least 10 times the response speed weight. By leveraging the high penalty of a variable-weight loss function based on confidence constraints to artificially amplify the calculation error of overshoot, the weight matrix and bias vector of the second neural network are iteratively updated in the direction of suppressing overshoot through backpropagation algorithm. In high-confidence response mode, when the input sample satisfies that the fusion safety coefficient is greater than or equal to the pre-position confidence safety threshold, the response speed weight is set to be greater than the stability weight. The weight matrix and bias vector of the second neural network are iteratively updated in the direction of eliminating real-time bias by driving the backpropagation algorithm, thereby completing the targeted optimization training of the second neural network. Step (63): Adaptive gain parameter inference and target stability control command generation: In the real-time inference stage of the second neural network, the real-time generated control input vector is input into the second neural network after directional optimization training, and the proportional gain of the PID controller is output in real time by the three parallel output neurons of the second neural network. Integral gain and differential gain ; to increase the proportional gain Integral gain and differential gain Substitute the values ​​into the incremental PID control algorithm to calculate the target stability control command. The calculation formula for the incremental PID control algorithm is: ， in, This is the target stability control command from the previous sampling time. and These are the real-time deviations at the first two sampling times; the target stability control command. The signal is sent to the vehicle actuators to implement braking intervention, completing adaptive fusion control covering four major areas: information security, functional safety, expected functional safety, and active safety.

7. A vehicle stability control multi-domain safety fusion system for implementing the method of any one of claims 1-6, characterized by, include: Multi-source heterogeneous data acquisition and external environment information set construction module: used to acquire vehicle operating status data, local environmental perception data and external collaborative perception data; Perform data validity preprocessing on external collaborative sensing data, remove abnormal messages, and form an external environment information set; The multi-independent sensing mode solution and normalization consistency deviation calculation module is used to construct four independent information acquisition modes—roadside communication, vehicle-to-vehicle communication, high-precision map, and dynamic inference—using external environmental information sets and local environmental sensing information. It performs spatiotemporal data alignment processing on the data of the four independent information acquisition modes and solves the geometric curvature features of the road ahead in parallel. We construct a normalized mapping relationship between the output values ​​of four independent information acquisition modes and calculate the normalized consistency deviation feature that characterizes the discreteness of multi-source perception. Physical dynamics parameter identification and adaptive calculation module for comprehensive stability risk value: It is used to establish longitudinal dynamic equilibrium equations based on operating status data, identify the vehicle mass in real time using the recursive least squares method, and simultaneously calculate the real-time center of gravity height. The vehicle mass and real-time center of gravity height are substituted into the attitude dynamics model. Combined with lateral acceleration and roll angle data, the basic load transfer index and attitude stability index are calculated respectively. Vertical acceleration and pitch angular velocity are introduced as correction variables to calculate the improved load transfer index that includes vertical dynamic characteristics. The vertical excitation characteristics of the road surface are calculated based on the vertical dynamic response data of the vibration reduction device. The road surface excitation confidence factor is calculated using the Sigmoid nonlinear function. The basic load transfer index, attitude stability index and improved load transfer index are processed with unified dimensions. Adaptive dynamic weighted fusion is performed based on the road surface excitation confidence factor to output the comprehensive stability risk value. The module for constructing a cascaded topology and constraint coupling path for dual neural networks is used to construct a cascaded topology containing a first neural network and a second neural network. The first neural network is defined as a cross-domain security assessment network, and the second neural network is defined as a confidence coupling control network. The module constructs a cascade coupling path by directly connecting the output of the first neural network to the input of the second neural network and the variable weight loss function of the second neural network, forming a structural constraint on the confidence coupling control network by the cross-domain security assessment network. The module for constructing and fusing a multi-source heterogeneous joint feature space based on the first neural network is used to extract the functional safety status features of the underlying execution unit and the information security status features of the network message, and to combine the environmental interference features and the normalization consistency deviation features to form the expected functional safety dimension features; it constructs a multi-source heterogeneous joint feature space, inputs the vector of the multi-source heterogeneous joint feature space into the first neural network pre-trained based on the asymmetric cost-sensitive mechanism, extracts the cross-domain fault coupling features between the physical perception layer, the information communication layer and the environmental interference layer, and calculates and outputs the fused safety coefficient that represents the current comprehensive credibility of the vehicle's multiple safety domains; The confidence-constrained PID gain dynamic calculation and target control command generation module based on the second neural network is used to calculate the real-time deviation between the comprehensive stability risk value and the preset stability target threshold; and to construct a control input vector containing the fused safety factor, comprehensive stability risk value and real-time deviation, which is then input into the second neural network. The safety factor is used as the confidence constraint weight. The gain parameter of the PID controller is dynamically calculated through the second neural network. The gain parameter and the real-time deviation are substituted into the PID control algorithm to calculate the target stability control command. The target stability control command is then sent to the vehicle actuator to implement active intervention.

Images