Electromechanical brake system brake pressure degradation strategy

Abstract

The invention discloses a braking pressure degradation strategy of an electronic mechanical braking system, and belongs to the technical field of vehicle braking control, and the braking pressure degradation strategy of the electronic mechanical braking system comprises the following steps: S1, multi-sensor fusion is carried out to correct a friction coefficient in real time, and a high-precision gridding pavement model is constructed; s2, performing time sequence prediction on braking torque distribution, dynamically constraining a braking pressure threshold value and performing pre-adjustment; s3, reinforcement learning is carried out to generate a prospective degradation sequence, and thermal coupling and pressure fluctuation control are optimized; s4, dynamically encrypting the hierarchical communication architecture, and ensuring that the instruction is complete and consistent with the time sequence through redundancy switching; s5, generating an adversarial network compensation failure signal, and dynamically adjusting the weight of the sensor to resist interference; s6, triggering a man-machine interaction strategy in a grading manner, and dynamically adjusting a safety threshold in combination with a driving behavior; s7, carrying out simulation verification on the digital twin platform, and carrying out fault injection encryption to update full-life-cycle parameters; the method has the beneficial effects that self-adaptive optimization and hierarchical fault tolerance of the brake pressure are realized, and the system safety and response efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle brake control, and in particular to a brake pressure degradation strategy for an electronic mechanical brake system. Background Art

[0002] As an upgrade solution for traditional hydraulic brakes, the electromechanical brake system (EMB) directly controls the distribution of braking force through electronic signals, and has significant advantages in response speed, energy efficiency and integration. However, in complex road environments (such as low-friction roads, dynamic obstacles, and extreme weather), the existing braking system faces the following challenges: Environmental perception limitations: a single sensor is difficult to accurately identify the gradient change of the road friction coefficient, resulting in a lag or misjudgment of the braking strategy; dynamic decision-making defects: traditional control algorithms rely on fixed thresholds and cannot adapt to real-time changes in braking force requirements, which are prone to uneven pressure distribution or thermal decay problems; communication and fault tolerance shortcomings: centralized communication architectures are prone to control failures when delayed or interrupted, and lack robust compensation mechanisms for signal distortion or hardware failures; insufficient personalized adaptation: the safety threshold is not dynamically adjusted in combination with driver behavior characteristics, resulting in low efficiency in human-machine collaboration.

[0003] With the development of autonomous driving technology, the safety and intelligence of the braking system are in urgent need. This strategy solves the problem of dynamic degradation of brake pressure in complex scenarios through multi-dimensional optimization, and is a core technology for improving vehicle active safety and driving experience. Summary of the invention

[0004] The present invention proposes a brake pressure degradation strategy for an electronic mechanical brake system based on multi-sensor fusion, time series prediction and reinforcement learning. Through dynamic environment modeling, hierarchical fault-tolerant communication and adaptive control optimization, safe and efficient braking under complex road conditions is achieved.

[0005] Technical solution: The brake pressure degradation strategy of the electromechanical brake system includes the following steps:

[0006] S1. Real-time collection of road environment data, including dynamic obstacles, three-dimensional road surface topology and material characteristics, is achieved through multi-sensor fusion technology to construct a high-precision grid friction coefficient heat map. The on-board inertial measurement data and filtering algorithm are combined for spatiotemporal registration, the friction coefficient estimate is continuously corrected, and the boundaries and gradient change rates of abnormally low friction areas are marked to form a multi-dimensional road surface model, providing an environmental perception basis for graded braking decisions.

[0007] S2. Based on the friction coefficient model output by S1, the braking torque distribution within a certain time range in the future is predicted through a time series neural network; when the variance of the prediction result exceeds the safety threshold, the three-level warning mechanism is activated, and the pre-adjustment is started when the set distance before the target low-friction area is detected: the backup actuator locking device is released, the front axle priority weight distribution strategy is adopted, and the maximum braking pressure is constrained to the preset safety threshold, so as to establish the boundary conditions for the subsequent dynamic degradation;

[0008] S3. Inherits the constraints of S2, combines the real-time vehicle motion parameters, and generates a forward-looking degradation control sequence through reinforcement learning; the optimization function integrates the brake pressure tracking error, friction efficiency margin, and actuator thermal load index to solve the suboptimal solution under the thermal coupling constraint, giving priority to controlling the wheel pressure fluctuation range in the low friction area, and realizing a closed-loop decision from prediction to execution;

[0009] S4. For dynamic instructions from S1 to S3, a hierarchical communication architecture is deployed: low-power wide area network protocol is used to transmit key instructions, and the self-consistency of the control sequence when communication is interrupted is guaranteed through a fault-tolerant mechanism, and the braking instructions are encrypted; when the communication delay is greater than the set threshold or the bit error rate exceeds the limit, it switches to the local redundant communication mode and synchronously updates the cloud risk assessment model to ensure data integrity and timing consistency in complex environments;

[0010] S5. Based on the potential signal distortion problems of S2 and S3, a multimodal compensation mechanism is constructed: the failure signal is reconstructed using a generative adversarial network, and the pressure confidence interval is generated by integrating the pedal stroke gradient and inertial data; when a hardware failure is detected, the backup control node is activated and the sensor weight is dynamically adjusted, and adversarial training is combined to improve the compensation robustness under disturbance conditions, forming a dual fault-tolerant system;

[0011] S6. Based on the warning status of S2 and the pressure confidence assessment of S5, a hierarchical interaction strategy is implemented: when the braking safety margin is greater than the high threshold, tactile feedback is triggered, when it is in the middle range, an augmented reality warning is projected, and when it is lower than the low threshold, the system takes over; a personalized model is built through cluster analysis of driver behavior data, and the pre-processing trigger timing and safety margin threshold are dynamically adjusted to form a closed-loop control linked with the degradation strategy;

[0012] S7. Build a digital twin platform by integrating all process parameters, rehearse the thermal decay and wear trend of the braking system, and verify the error range of key indicators; generate use cases through fault injection testing, encrypt and upload parameters to the cloud knowledge base, trigger incremental updates of pressure thresholds and friction compensation factors, and ensure the optimization consistency of the system throughout its life cycle.

[0013] Preferably, the S1 specifically comprises the following steps:

[0014] S1-1. Synchronous perception layer: A multi-source data fusion system is formed using millimeter-wave radar, binocular stereo vision sensor and laser radar. The three-dimensional coordinate information of dynamic obstacles is synchronously updated in a cycle of 300 milliseconds. The obstacle recognition accuracy reaches 5 centimeters. The coordinate information is used as the basic input for subsequent road surface analysis.

[0015] S1-2. Road surface texture feature analysis: The road surface texture statistical features are generated based on the three-dimensional laser point cloud computing module, and the surface roughness index is quantified as the Ra value level. At the same time, the near-infrared spectral reflectance data is integrated to identify the asphalt, cement or ice material type. The material identification result is generated by the feature mapping algorithm to generate a square grid friction coefficient heat map with a unit of fifty centimeters;

[0016] S1-3. Dynamic data correction module: Apply the data registration algorithm based on Kalman filter to synchronize the yaw rate data of the inertial navigation system with the laser point cloud in time and space, update the friction coefficient estimation value once per second through the gradient descent optimization algorithm, and transmit the friction coefficient data stream after compensating the sensor drift error to the risk analysis module;

[0017] S1-4. Risk boundary extraction engine: The modified friction coefficient distribution output by S1-3 is processed using an improved Canny edge detection algorithm to identify the boundaries of dangerous areas where the friction coefficient is lower than 0.3, and to track abnormal areas where the friction gradient change rate exceeds 0.1 per second. The generated risk topology layer establishes a data interface with the early warning module of step S2.

[0018] Preferably, the S2 specifically comprises the following steps:

[0019] S2-1. Time series prediction center: construct a long short-term memory time series neural network to analyze the friction coefficient model output by S1, predict the spatial distribution of the braking torque in the next three seconds, and send a three-level warning instruction to the human-computer interaction system through the CAN bus when the standard deviation of the prediction result exceeds 0.3 MPa: the distribution data is used as the decision basis for subsequent pre-adjustment;

[0020] S2-2. Braking force pre-distribution unit: When the laser radar detects the target low-friction area fifty meters ahead, it receives the warning signal of S2-1 and releases the mechanical locking device of the backup hydraulic cylinder through the electromagnetic actuator, and implements the front axle braking force distribution weight to the range of 60% to 80% based on the predicted torque distribution;

[0021] S2-3. Constraint condition generator: obtain the maximum road adhesion coefficient output by S1-3 through the real-time bus, and dynamically calculate the air chamber pressure constraint value based on the coefficient, where the output of the safety threshold calculation model is used as the boundary condition input of the reinforcement learning in step S3.

[0022] Preferably, the S3 specifically includes the following steps:

[0023] S3-1. Reinforcement learning decision core: Establish state boundaries based on the pressure constraint values ​​output by S2-3, construct a Q learning agent to generate an eight-step forward pressure degradation sequence, and transmit the sequence to the wire control brake module via the controller area network;

[0024] S3-2. Safety margin optimizer: constructs a target optimization function based on the instruction sequence generated by S3-1, and the input data of the friction safety margin weighted term comes from the risk topology layer generated by S1-4;

[0025] S3-3. Thermodynamic dynamic weight adjustment: The current pressure command is input into the finite element thermodynamic model, and the thermal stress distribution data output by the model is fed back to the optimization function of S3-2 for dynamic weight adjustment;

[0026] S3-4. Pressure fluctuation control module: Based on the thermal stability boundary calculated in S3-3, a pressure fluctuation suppression algorithm is embedded in the wheel control loop on the low-friction side, and its real-time control data is transmitted to the compensation subsystem on the opposite side through the hydraulic system coupling module.

[0027] Preferably, the S4 specifically comprises the following steps:

[0028] S4-1. Communication protocol classification module: define the hydraulic system control instructions generated by S3-4 as Class A communication data, directly transmit them to the wire control execution terminal through the low-power wide area network protocol, and simultaneously classify the system health status data as Class B information, and upload them to the cloud monitoring platform through the narrowband Internet of Things protocol;

[0029] S4-2. Fault-tolerant control activator: When the communication link of S4-1 fails, the sequential logic controller pre-stored in the local flash memory is enabled, which reconstructs the braking system control command queue based on the historical records of the three most recent valid warning signals of S2-1;

[0030] S4-3. Dynamic encryption verification unit: The current transmission data packet is processed by the encryption key generated by the S3-1 agent, encrypted using the improved Advanced Encryption Standard 256-bit algorithm, and the road fingerprint information obtained by S1-2 is added to the cyclic redundancy check code as a dynamic salt value;

[0031] S4-4. Network redundancy switching mechanism: Establish a network quality perception thread. When it is detected that the end-to-end delay of the S4-1 transmission channel exceeds two hundred milliseconds, it automatically switches to the dual-channel redundancy mode of the CAN flexible data rate protocol. The switching instruction synchronously triggers the emergency control node activation process of S5-3.

[0032] Preferably, the S5 specifically includes the following steps:

[0033] S5-1. Construction of signal reconstruction training set: Based on the corrected sensor data output by S1-3, a training data set for the signal reconstruction model is constructed. The generator module of the generative adversarial network learns the joint time-frequency features of each sensor signal in step S1 through the back propagation algorithm;

[0034] S5-2. Dynamic confidence interval calculator: Combined with the front axle braking force distribution parameters implemented by S2-2, the pedal travel-acceleration confidence interval is calculated in real time, and its lower limit value is dynamically adjusted according to the pressure fluctuation suppression intensity output by S3-4;

[0035] S5-3. Emergency mode decision center: When the emergency mode triggered by S4-4 is activated, the reinforcement learning controller that takes over the S3 step calculates the optimal weight coefficient in real time based on the signal data reconstructed by S5-1, and the coefficient is transmitted back to the decision center through the safety channel of S4-3;

[0036] S5-4. Anti-interference training enhancement module: In the system self-learning stage, the risk scenario characteristic signal identified by S1-4 is injected to construct a composite training set containing Gaussian noise and pulse interference, and its disturbance amplitude is dynamically adjusted according to the safety margin calculated by S2-3.

[0037] Preferably, the S6 specifically comprises the following steps:

[0038] S6-1. Risk assessment grading mechanism: The prediction results of the brake safety margin optimized by S3-2 are divided into three levels, among which the judgment threshold of the red warning level is analyzed for spatial overlap with the high-risk area distribution map generated by S1-4;

[0039] S6-2. Human-machine collaborative response channel: establish a human-machine collaborative control loop, the warning signal of S6-1 drives the execution device in real time through the communication link established by S4-1, and the yellow level virtual track projection function calls the road grid map data generated by S1-2;

[0040] S6-3. Driving behavior modeling engine: Builds a driver profile model based on the thermodynamic parameters output by S3-3. By analyzing nearly 100 braking operation data recorded by S2-1, it generates a personalized correction coefficient and feeds it back to the Q learning decision tree of S3-1.

[0041] S6-4. Safety strategy dynamic corrector: Dynamically corrects the pre-processing trigger conditions. Its core parameters integrate the real-time road adhesion coefficient of S2-3 and the lower limit of the confidence interval calculated by S5-2 to form a closed-loop safety control strategy.

[0042] Preferably, the S7 specifically includes the following steps:

[0043] S7-1. Multi-physics simulation modeling layer: Construct a multi-physics joint simulation environment, use the thermal boundary conditions output by S3-4 and the driver behavior model of S6-3 as initial parameters, and solve the three-dimensional unsteady-state heat conduction differential equation through implicit difference algorithm;

[0044] S7-2. Verification benchmark adaptive adjuster: defines the system verification benchmark, where the pressure tracking error threshold is positively correlated with the temperature compensation coefficient of S2-3, and the temperature prediction deviation allowable range is dynamically shrunk according to the model accuracy of S7-1;

[0045] S7-3. Fault mode injection analyzer: construct a fault injection test set under the closed-loop strategy framework formed in S6-4, in which the sample weights of the short-circuit fault mode are adaptively allocated according to the adversarial training results of S5-4;

[0046] S7-4. Distributed knowledge storage architecture: Create a distributed parameter storage architecture. The decision parameters generated by S3-1 reinforcement learning are uploaded to the blockchain node through the S4-3 encrypted channel, while a compressed knowledge copy matching the S5-1 signal features is retained locally.

[0047] Compared with the prior art, the advantages of the present invention are:

[0048] (1) Accurate environmental modeling and real-time correction: Integrating multi-source data such as millimeter-wave radar and lidar to construct a high-precision grid-based friction coefficient thermal map, combined with Kalman filter dynamic correction, significantly improves the road state perception accuracy (error < 5 cm), providing a reliable basis for braking decisions;

[0049] (2) Prediction-optimization closed-loop control architecture: The LSTM time series neural network is used to predict the braking torque distribution, and reinforcement learning is combined to generate a forward-looking degradation sequence to achieve full-link closed-loop optimization from environmental perception to execution. The prediction cycle covers the next 3 seconds, and the braking force distribution error is reduced by more than 40%;

[0050] (3) Multimodal compensation and personalized interaction: Based on the generative adversarial network to reconstruct the failure signal, combined with the driver behavior cluster analysis to dynamically adjust the safety margin threshold, provide tactile feedback, AR warning and other multi-level human-computer interaction, the system takeover response time is shortened to within 0.5 seconds, and the user error rate is reduced by 60%; DETAILED DESCRIPTION

[0051] Example

[0052] 1. The brake pressure degradation strategy of the electromechanical brake system is characterized by comprising the following steps:

[0053] S1. Real-time collection of road environment data, including dynamic obstacles, three-dimensional road surface topology and material characteristics, is achieved through multi-sensor fusion technology to construct a high-precision grid friction coefficient heat map. The on-board inertial measurement data and filtering algorithm are combined for spatiotemporal registration, the friction coefficient estimate is continuously corrected, and the boundaries and gradient change rates of abnormally low friction areas are marked to form a multi-dimensional road surface model, providing an environmental perception basis for graded braking decisions.

[0054] S2. Based on the friction coefficient model output by S1, the braking torque distribution within a certain time range in the future is predicted through a time series neural network; when the variance of the prediction result exceeds the safety threshold, the three-level warning mechanism is activated, and the pre-adjustment is started when the set distance before the target low-friction area is detected: the backup actuator locking device is released, the front axle priority weight distribution strategy is adopted, and the maximum braking pressure is constrained to the preset safety threshold, so as to establish the boundary conditions for the subsequent dynamic degradation;

[0055] S3. Inherits the constraints of S2, combines the real-time vehicle motion parameters, and generates a forward-looking degradation control sequence through reinforcement learning; the optimization function integrates the brake pressure tracking error, friction efficiency margin, and actuator thermal load index to solve the suboptimal solution under the thermal coupling constraint, giving priority to controlling the wheel pressure fluctuation range in the low friction area, and realizing a closed-loop decision from prediction to execution;

[0056] S4. For dynamic instructions from S1 to S3, a hierarchical communication architecture is deployed: low-power wide area network protocol is used to transmit key instructions, and the self-consistency of the control sequence when communication is interrupted is guaranteed through a fault-tolerant mechanism, and the braking instructions are encrypted; when the communication delay is greater than the set threshold or the bit error rate exceeds the limit, it switches to the local redundant communication mode and synchronously updates the cloud risk assessment model to ensure data integrity and timing consistency in complex environments;

[0057] S5. Based on the potential signal distortion problems of S2 and S3, a multimodal compensation mechanism is constructed: the failure signal is reconstructed using a generative adversarial network, and the pressure confidence interval is generated by integrating the pedal stroke gradient and inertial data; when a hardware failure is detected, the backup control node is activated and the sensor weight is dynamically adjusted, and adversarial training is combined to improve the compensation robustness under disturbance conditions, forming a dual fault-tolerant system;

[0058] S6. Based on the warning status of S2 and the pressure confidence assessment of S5, a hierarchical interaction strategy is implemented: when the braking safety margin is greater than the high threshold, tactile feedback is triggered, when it is in the middle range, an augmented reality warning is projected, and when it is lower than the low threshold, the system takes over; a personalized model is built through cluster analysis of driver behavior data, and the pre-processing trigger timing and safety margin threshold are dynamically adjusted to form a closed-loop control linked with the degradation strategy;

[0059] S7. Build a digital twin platform by integrating all process parameters, rehearse the thermal decay and wear trend of the braking system, and verify the error range of key indicators; generate use cases through fault injection testing, encrypt and upload parameters to the cloud knowledge base, trigger incremental updates of pressure thresholds and friction compensation factors, and ensure the optimization consistency of the system throughout its life cycle.

[0060] The S1 specifically includes the following steps:

[0061] S1-1. Synchronous perception layer: A multi-source data fusion system is formed using millimeter-wave radar, binocular stereo vision sensor and laser radar. The three-dimensional coordinate information of dynamic obstacles is synchronously updated in a cycle of 300 milliseconds. The obstacle recognition accuracy reaches 5 centimeters. The coordinate information is used as the basic input for subsequent road surface analysis.

[0062] S1-2. Road surface texture feature analysis: The road surface texture statistical features are generated based on the three-dimensional laser point cloud computing module, and the surface roughness index is quantified as the Ra value level. At the same time, the near-infrared spectral reflectance data is integrated to identify the asphalt, cement or ice material type. The material identification result is generated by the feature mapping algorithm to generate a square grid friction coefficient heat map with a unit of fifty centimeters;

[0063] Calculation of surface roughness Ra value:

[0064]

[0065] Glossary: ​​Ra value

[0066] Chinese full name: Arithmetic Average Roughness; English full name: Arithmetic Average Roughness;

[0067] Parameter Description: z i : Height measurement value of a single laser point cloud; μ z : average height in the grid area; N: number of laser points in the current grid;

[0068] Technical effect: Ra value quantifies the microscopic morphology of the road surface and assists in distinguishing the material type (for example, the Ra value of ice surface is significantly lower than that of asphalt);

[0069] Bilinear interpolation grid mapping formula:

[0070]

[0071] Parameter description: (x1, y1), (x2, y2): coordinates of the center of the adjacent grid; μ 11 ,μ 21 ,μ 12 ,μ 22 : Friction coefficient of adjacent grids;

[0072] Technical effect: Bilinear interpolation ensures smooth transition of friction coefficient between 50cm grids, improving the continuity of thermal maps and S2-1 prediction accuracy;

[0073] S1-3. Dynamic data correction module: Apply the data registration algorithm based on Kalman filter to synchronize the yaw rate data of the inertial navigation system with the laser point cloud in time and space, update the friction coefficient estimation value once per second through the gradient descent optimization algorithm, and transmit the friction coefficient data stream after compensating the sensor drift error to the risk analysis module;

[0074] Kalman filter state update equation:

[0075]

[0076] Parameter Description: The estimated value of the state at step k; K k : Kalman gain matrix; z k : sensor observation value; H: observation matrix

[0077] Technical effect: Kalman filtering eliminates sensor random noise and drift errors by fusing inertial navigation data with point cloud timing information;

[0078] Gradient descent optimization objective function:

[0079]

[0080] Parameter description: θ: friction coefficient estimation parameter; y i : true friction coefficient;

[0081] f θ (x i ): Friction coefficient prediction model based on sensor data;

[0082] Technical effect: Gradient descent updates the friction coefficient once per second, correcting local deviations of the road model in real time and providing accurate input for risk boundary extraction of S1-4;

[0083] S1-4. Risk boundary extraction engine: The modified friction coefficient distribution output by S1-3 is processed using an improved Canny edge detection algorithm to identify the boundaries of dangerous areas where the friction coefficient is lower than 0.3, and to track abnormal areas where the friction gradient change rate exceeds 0.1 per second. The generated risk topology layer establishes a data interface with the early warning module of step S2.

[0084] Canny edge detection gradient amplitude calculation:

[0085]

[0086] Parameter description: G: friction coefficient gradient amplitude; μ: road friction coefficient;

[0087] Technical effect: By calculating the spatial variation of the friction coefficient gradient, the boundary of the jump area (dangerous area) with a friction coefficient lower than 0.3 can be accurately identified, and the pre-allocation strategy can be triggered by linking with the early warning module of S2;

[0088] The S2 specifically includes the following steps:

[0089] S2-1. Time series prediction center: construct a long short-term memory time series neural network to analyze the friction coefficient model output by S1, predict the spatial distribution of the braking torque in the next three seconds, and send a three-level warning instruction to the human-computer interaction system through the CAN bus when the standard deviation of the prediction result exceeds 0.3 MPa: the distribution data is used as the decision basis for subsequent pre-adjustment;

[0090] LSTM unit state update formula:

[0091]

[0092] Glossary: ​​LSTM;

[0093] Chinese full name: Long Short-Term Memory Network; English full name: Long Short-Term Memory;

[0094] Parameter Description: C t : The current unit status; f t ,i t : The activation values ​​of the forget gate and the input gate; Candidate memory cells;

[0095] Technical effect: The long-term and short-term dependencies of the spatiotemporal distribution of the friction coefficient are captured through LSTM, and the variance of the braking torque distribution in the next 3 seconds is predicted (standard deviation threshold 0.3MPa), which is used to activate the pre-allocation strategy of S2-2;

[0096] S2-2. Braking force pre-distribution unit: When the laser radar detects the target low-friction area fifty meters ahead, it receives the warning signal of S2-1 and releases the mechanical locking device of the backup hydraulic cylinder through the electromagnetic actuator, and implements the front axle braking force distribution weight to the range of 60% to 80% based on the predicted torque distribution;

[0097] Front axle weight dynamic distribution model:

[0098]

[0099] Parameter description: μ front ,μ rear : Friction coefficient of the road surface corresponding to the front and rear axles (from S1-3); η safe=0.8 (maximum safety weight limit);

[0100] Technical effect: Adaptive distribution based on the friction difference between the front and rear axles avoids steering failure caused by front axle overload and meets the weight constraint of 60% to 80%, providing initial action space for S3-1;

[0101] S2-3. Constraint condition generator: obtain the maximum road adhesion coefficient output by S1-3 through the real-time bus, and dynamically calculate the air chamber pressure constraint value based on the coefficient, where the output of the safety threshold calculation model is used as the boundary condition input of the reinforcement learning in step S3.

[0102] Maximum brake pressure safety threshold model:

[0103] P max = k·μ peak ·A brake

[0104] Parameter Description: max : Maximum allowable braking pressure; μ peak : Peak road adhesion coefficient (from S1-3); A brake : Braking area;

[0105] Technical effect: Dynamically limit the upper pressure limit to avoid tire locking on low-friction roads, and provide a constraint boundary for S3's reinforcement learning.

[0106] The S3 specifically includes the following steps:

[0107] S3-1. Reinforcement learning decision core: Establish state boundaries based on the pressure constraint values ​​output by S2-3, construct a Q learning agent to generate an eight-step forward pressure degradation sequence, and transmit the sequence to the wire control brake module via the controller area network;

[0108] Q-learning update rule:

[0109]

[0110] Parameter description: Q(s,a): state-action value function; α: learning rate; γ: discount factor;

[0111] Technical effect: The forward-looking 8-step pressure degradation sequence generation ensures the selection of suboptimal strategies under thermodynamic constraints (such as preferentially suppressing wheel pressure fluctuations on the low-friction side);

[0112] S3-2. Safety margin optimizer: constructs a target optimization function based on the instruction sequence generated by S3-1, and the input data of the friction safety margin weighted term comes from the risk topology layer generated by S1-4;

[0113] Objective optimization function:

[0114] J=w1∫(P cmd -P actual ) 2 dt+w2·min(μ reserve )

[0115] Parameter description: w1, w2: weight coefficient; μ reserve : Friction safety margin (from S1-4);

[0116] Technical effect: Balance pressure tracking accuracy and friction margin to prevent the system from entering an irreversible slip zone;

[0117] S3-3. Thermodynamic dynamic weight adjustment: The current pressure command is input into the finite element thermodynamic model, and the thermal stress distribution data output by the model is fed back to the optimization function of S3-2 for dynamic weight adjustment;

[0118] Finite element thermal node temperature update equation:

[0119]

[0120] Parameter Description: The temperature of node i at time step n; q i : Brake heat flux density (related to the current pressure command); Ω(i): The set of adjacent nodes of node i;

[0121] Technical effect: Dynamically predict the thermal stress distribution and feed it back to the S3-2 optimization function (such as weight w2∝max(T)) to ensure that the brake thermal load does not exceed the material limit;

[0122] S3-4. Pressure fluctuation control module: Based on the thermal stability boundary calculated in S3-3, a pressure fluctuation suppression algorithm is embedded in the wheel control loop on the low-friction side, and its real-time control data is transmitted to the compensation subsystem on the opposite side through the hydraulic system coupling module.

[0123] The S4 specifically comprises the following steps:

[0124] S4-1. Communication protocol classification module: define the hydraulic system control instructions generated by S3-4 as Class A communication data, directly transmit them to the wire control execution terminal through the low-power wide area network protocol, and simultaneously classify the system health status data as Class B information, and upload them to the cloud monitoring platform through the narrowband Internet of Things protocol;

[0125] S4-2. Fault-tolerant control activator: When the communication link of S4-1 fails, the sequential logic controller pre-stored in the local flash memory is enabled, which reconstructs the braking system control command queue based on the historical records of the three most recent valid warning signals of S2-1;

[0126] S4-3. Dynamic encryption verification unit: The current transmission data packet is processed by the encryption key generated by the S3-1 agent, encrypted using the improved Advanced Encryption Standard 256-bit algorithm, and the road fingerprint information obtained by S1-2 is added to the cyclic redundancy check code as a dynamic salt value;

[0127] AES-256 round key generation function:

[0128]

[0129] Parameter description: SubWord: byte replacement operation; RotWord: word rotation; Rcon: round constant matrix;

[0130] Technical effect: Combined with dynamic salt value (road fingerprint), it enhances encryption uniqueness, prevents S4-1 data transmission from being tampered with, and ensures input reliability of S5-2 confidence interval calculation;

[0131] S4-4. Network redundancy switching mechanism: Establish a network quality perception thread. When it is detected that the end-to-end delay of the S4-1 transmission channel exceeds two hundred milliseconds, it automatically switches to the dual-channel redundancy mode of the CAN flexible data rate protocol. The switching instruction synchronously triggers the emergency control node activation process of S5-3.

[0132] End-to-end delay tolerance model:

[0133]

[0134] Parameter description: d reaction : Braking system reaction distance; v: Vehicle speed;

[0135] Technical effect: When the delay exceeds 200ms, it switches to CAN FD dual-channel mode to ensure real-time response of the S5 emergency control node.

[0136] The S5 specifically includes the following steps:

[0137] S5-1. Construction of signal reconstruction training set: Based on the corrected sensor data output by S1-3, a training data set for the signal reconstruction model is constructed. The generator module of the generative adversarial network learns the joint time-frequency features of each sensor signal in step S1 through the back propagation algorithm;

[0138] GAN generator loss function:

[0139]

[0140] Parameter description: D: discriminator network, G: generator network; z: noise vector;

[0141] Technical effect: Reconstruct failure signals (such as pedal travel sensor abnormality) to form the emergency control data source of S5-3;

[0142] S5-2. Dynamic confidence interval calculator: Combined with the front axle braking force distribution parameters implemented by S2-2, the pedal travel-acceleration confidence interval is calculated in real time, and its lower limit value is dynamically adjusted according to the pressure fluctuation suppression intensity output by S3-4;

[0143] S5-3. Emergency mode decision center: When the emergency mode triggered by S4-4 is activated, the reinforcement learning controller that takes over the S3 step calculates the optimal weight coefficient in real time based on the signal data reconstructed by S5-1, and the coefficient is transmitted back to the decision center through the safety channel of S4-3;

[0144] S5-4. Anti-interference training enhancement module: In the system self-learning stage, the risk scenario characteristic signal identified by S1-4 is injected to construct a composite training set containing Gaussian noise and pulse interference, and its disturbance amplitude is dynamically adjusted according to the safety margin calculated by S2-3.

[0145] Online anti-interference objective function:

[0146]

[0147] Parameter Description: D adv : Adversarial discriminator, used to identify noise patterns; λ: robustness adjustment factor (dynamically adjusted according to the S2-3 safety margin);

[0148] Technical effect: Enhance the ability of generative adversarial networks to resist Gaussian noise and pulse interference, and improve the reliability of signal reconstruction in S5-3 emergency mode.

[0149] The S6 specifically comprises the following steps:

[0150] S6-1. Risk assessment grading mechanism: The prediction results of the brake safety margin optimized by S3-2 are divided into three levels, among which the judgment threshold of the red warning level is analyzed for spatial overlap with the high-risk area distribution map generated by S1-4;

[0151] Risk level threshold dynamic correction formula:

[0152] ξ th =μ reserve -k σ ·σ history

[0153] Parameter description: th : Level determination threshold; σ history : Historical braking safety margin standard deviation; k σ (Empirical coefficient, the default value is 1.5);

[0154] Technical effect: Combining the real-time margin optimized by S3-2 with historical volatility, the red / yellow warning boundary is adaptively adjusted to avoid false triggering of emergency takeover;

[0155] S6-2. Human-machine collaborative response channel: establish a human-machine collaborative control loop, the warning signal of S6-1 drives the execution device in real time through the communication link established by S4-1, and the yellow level virtual track projection function calls the road grid map data generated by S1-2;

[0156] S6-3. Driving behavior modeling engine: Builds a driver profile model based on the thermodynamic parameters output by S3-3. By analyzing nearly 100 braking operation data recorded by S2-1, it generates a personalized correction coefficient and feeds it back to the Q learning decision tree of S3-1.

[0157] Clustering objective function:

[0158]

[0159] Parameter Description: S i : driving behavior cluster of type i; μ i :Cluster Center;

[0160] Technical effect: Based on the clustering of 100 braking operation data, a driver profile is generated and fed back to the Q-learning decision tree to adjust the trigger threshold of the downgrade strategy in a personalized manner;

[0161] S6-4. Safety strategy dynamic corrector: Dynamically corrects the pre-processing trigger conditions. Its core parameters integrate the real-time road adhesion coefficient of S2-3 and the lower limit of the confidence interval calculated by S5-2 to form a closed-loop safety control strategy.

[0162] The S7 specifically comprises the following steps:

[0163] S7-1. Multi-physics simulation modeling layer: Construct a multi-physics joint simulation environment, use the thermal boundary conditions output by S3-4 and the driver behavior model of S6-3 as initial parameters, and solve the three-dimensional unsteady-state heat conduction differential equation through implicit difference algorithm;

[0164] Three-dimensional unsteady heat conduction equation:

[0165]

[0166] Parameter description: T: temperature field distribution; k: thermal conductivity of material;

[0167] Technical effect: Combined with the thermal boundary conditions of S3-4, the thermal decay trend of the brake disc is predicted, and the adaptive verification benchmark accuracy of S7-2 is verified;

[0168] S7-2. Verification benchmark adaptive adjuster: defines the system verification benchmark, where the pressure tracking error threshold is positively correlated with the temperature compensation coefficient of S2-3, and the temperature prediction deviation allowable range is dynamically shrunk according to the model accuracy of S7-1;

[0169] S7-3. Fault mode injection analyzer: construct a fault injection test set under the closed-loop strategy framework formed in S6-4, in which the sample weights of the short-circuit fault mode are adaptively allocated according to the adversarial training results of S5-4;

[0170] Fault sample weight allocation rules:

[0171]

[0172] Parameter Description: L adv,i : adversarial training loss value of the i-th failure mode; α adv =2(attenuation coefficient);

[0173] Technical effect: Increase the test weight for the weak links of S5-4 training, improve the verification coverage, and ensure the effectiveness of the S7-4 knowledge base update;

[0174] S7-4. Distributed knowledge storage architecture: Create a distributed parameter storage architecture. The decision parameters generated by S3-1 reinforcement learning are uploaded to the blockchain node through the S4-3 encrypted channel, while a compressed knowledge copy matching the S5-1 signal features is retained locally.

[0175] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. The brake pressure degradation strategy of the electromechanical brake system is characterized by: The following steps are involved: S1. Real-time collection of road environment data, including dynamic obstacles, three-dimensional road surface topology and material characteristics, is achieved through multi-sensor fusion technology to construct a high-precision grid friction coefficient heat map. The on-board inertial measurement data and filtering algorithm are combined for spatiotemporal registration, the friction coefficient estimate is continuously corrected, and the boundaries and gradient change rates of abnormally low friction areas are marked to form a multi-dimensional road surface model, providing an environmental perception basis for graded braking decisions. S2. Based on the friction coefficient model output by S1, the braking torque distribution within a certain time range in the future is predicted through a time series neural network; when the variance of the prediction result exceeds the safety threshold, the three-level warning mechanism is activated, and the pre-adjustment is started when the set distance before the target low-friction area is detected: the backup actuator locking device is released, the front axle priority weight distribution strategy is adopted, and the maximum braking pressure is constrained to the preset safety threshold, so as to establish the boundary conditions for the subsequent dynamic degradation; S3. Inherits the constraints of S2, combines the real-time vehicle motion parameters, and generates a forward-looking degradation control sequence through reinforcement learning; the optimization function integrates the brake pressure tracking error, friction efficiency margin, and actuator thermal load index to solve the suboptimal solution under the thermal coupling constraint, giving priority to controlling the wheel pressure fluctuation range in the low friction area, and realizing a closed-loop decision from prediction to execution; S4. For dynamic instructions from S1 to S3, a hierarchical communication architecture is deployed: low-power wide area network protocol is used to transmit key instructions, and the self-consistency of the control sequence when communication is interrupted is guaranteed through a fault-tolerant mechanism, and the braking instructions are encrypted; when the communication delay is greater than the set threshold or the bit error rate exceeds the limit, it switches to the local redundant communication mode and synchronously updates the cloud risk assessment model to ensure data integrity and timing consistency in complex environments; S5. Based on the potential signal distortion problem of S2 and S3, a multimodal compensation mechanism is constructed: the failure signal is reconstructed using a generative adversarial network, and the pressure confidence interval is generated by integrating the pedal stroke gradient and inertial data; When a hardware failure is detected, the backup control node is activated and the sensor weight is dynamically adjusted. Combined with adversarial training, the compensation robustness under disturbance conditions is improved to form a dual fault-tolerant system. S6. Based on the warning status of S2 and the pressure confidence assessment of S5, a hierarchical interaction strategy is implemented: when the braking safety margin is greater than the high threshold, tactile feedback is triggered, when it is in the middle range, an augmented reality warning is projected, and when it is lower than the low threshold, the system takes over; a personalized model is built through cluster analysis of driver behavior data, and the pre-processing trigger timing and safety margin threshold are dynamically adjusted to form a closed-loop control linked with the degradation strategy; S7. Build a digital twin platform by integrating all process parameters, rehearse the thermal decay and wear trend of the braking system, and verify the error range of key indicators; generate use cases through fault injection testing, encrypt and upload parameters to the cloud knowledge base, trigger incremental updates of pressure thresholds and friction compensation factors, and ensure the optimization consistency of the system throughout its life cycle.

2. The brake pressure degradation strategy of the electromechanical brake system according to claim 1, characterized in that: The S1 specifically includes the following steps: S1-1. Synchronous perception layer: A multi-source data fusion system is formed using millimeter-wave radar, binocular stereo vision sensor and laser radar. The three-dimensional coordinate information of dynamic obstacles is synchronously updated in a cycle of 300 milliseconds. The obstacle recognition accuracy reaches 5 centimeters. The coordinate information is used as the basic input for subsequent road surface analysis. S1-2. Road surface texture feature analysis: The road surface texture statistical features are generated based on the three-dimensional laser point cloud computing module, and the surface roughness index is quantified as the Ra value level. At the same time, the near-infrared spectral reflectance data is integrated to identify the asphalt, cement or ice material type. The material identification result is generated by the feature mapping algorithm to generate a square grid friction coefficient heat map with a unit of fifty centimeters; S1-3. Dynamic data correction module: Apply the data registration algorithm based on Kalman filter to synchronize the yaw rate data of the inertial navigation system with the laser point cloud in time and space, update the friction coefficient estimation value once per second through the gradient descent optimization algorithm, and transmit the friction coefficient data stream after compensating the sensor drift error to the risk analysis module; S1-4. Risk boundary extraction engine: The modified friction coefficient distribution output by S1-3 is processed using an improved Canny edge detection algorithm to identify the boundaries of dangerous areas where the friction coefficient is lower than 0.3, and to track abnormal areas where the friction gradient change rate exceeds 0.1 per second. The generated risk topology layer establishes a data interface with the early warning module of step S2.

3. The brake pressure degradation strategy of the electromechanical brake system according to claim 1, characterized in that: The S2 specifically includes the following steps: S2-1. Time series prediction center: construct a long short-term memory time series neural network to analyze the friction coefficient model output by S1, predict the spatial distribution of the braking torque in the next three seconds, and send a three-level warning instruction to the human-computer interaction system through the CAN bus when the standard deviation of the prediction result exceeds 0.3 MPa: the distribution data is used as the decision basis for subsequent pre-adjustment; S2-2. Braking force pre-distribution unit: When the laser radar detects the target low-friction area fifty meters ahead, it receives the warning signal of S2-1 and releases the mechanical locking device of the backup hydraulic cylinder through the electromagnetic actuator, and implements the front axle braking force distribution weight to the range of 60% to 80% based on the predicted torque distribution; S2-3. Constraint condition generator: obtain the maximum road adhesion coefficient output by S1-3 through the real-time bus, and dynamically calculate the air chamber pressure constraint value based on the coefficient, where the output of the safety threshold calculation model is used as the boundary condition input of the reinforcement learning in step S3.

4. The brake pressure degradation strategy of the electromechanical brake system according to claim 1, characterized in that: The S3 specifically includes the following steps: S3-1. Reinforcement learning decision core: Establish state boundaries based on the pressure constraint values ​​output by S2-3, construct a Q learning agent to generate an eight-step forward pressure degradation sequence, and transmit the sequence to the wire control brake module via the controller area network; S3-2. Safety margin optimizer: constructs a target optimization function based on the instruction sequence generated by S3-1, and the input data of the friction safety margin weighted term comes from the risk topology layer generated by S1-4; S3-3. Thermodynamic dynamic weight adjustment: The current pressure command is input into the finite element thermodynamic model, and the thermal stress distribution data output by the model is fed back to the optimization function of S3-2 for dynamic weight adjustment; S3-4. Pressure fluctuation control module: Based on the thermal stability boundary calculated in S3-3, a pressure fluctuation suppression algorithm is embedded in the wheel control loop on the low-friction side, and its real-time control data is transmitted to the compensation subsystem on the opposite side through the hydraulic system coupling module.

5. The brake pressure degradation strategy of the electromechanical brake system according to claim 1, characterized in that: The S4 specifically comprises the following steps: S4-1. Communication protocol classification module: define the hydraulic system control instructions generated by S3-4 as Class A communication data, directly transmit them to the wire control execution terminal through the low-power wide area network protocol, and simultaneously classify the system health status data as Class B information, and upload them to the cloud monitoring platform through the narrowband Internet of Things protocol; S4-2. Fault-tolerant control activator: When the communication link of S4-1 fails, the sequential logic controller pre-stored in the local flash memory is enabled, which reconstructs the braking system control command queue based on the historical records of the three most recent valid warning signals of S2-1; S4-3. Dynamic encryption verification unit: The current transmission data packet is processed by the encryption key generated by the S3-1 agent, encrypted using the improved Advanced Encryption Standard 256-bit algorithm, and the road fingerprint information obtained by S1-2 is added to the cyclic redundancy check code as a dynamic salt value; S4-4. Network redundancy switching mechanism: Establish a network quality perception thread. When it is detected that the end-to-end delay of the S4-1 transmission channel exceeds two hundred milliseconds, it automatically switches to the dual-channel redundancy mode of the CAN flexible data rate protocol. The switching instruction synchronously triggers the emergency control node activation process of S5-3.

6. The brake pressure degradation strategy of the electromechanical brake system according to claim 1, characterized in that: The S5 specifically includes the following steps: S5-1. Construction of signal reconstruction training set: Based on the corrected sensor data output by S1-3, a training data set for the signal reconstruction model is constructed. The generator module of the generative adversarial network learns the joint time-frequency features of each sensor signal in step S1 through the back propagation algorithm; S5-2. Dynamic confidence interval calculator: Combined with the front axle braking force distribution parameters implemented by S2-2, the pedal travel-acceleration confidence interval is calculated in real time, and its lower limit value is dynamically adjusted according to the pressure fluctuation suppression intensity output by S3-4; S5-3. Emergency mode decision center: When the emergency mode triggered by S4-4 is activated, the reinforcement learning controller that takes over the S3 step calculates the optimal weight coefficient in real time based on the signal data reconstructed by S5-1, and the coefficient is transmitted back to the decision center through the safety channel of S4-3; S5-4. Anti-interference training enhancement module: In the system self-learning stage, the risk scenario characteristic signal identified by S1-4 is injected to construct a composite training set containing Gaussian noise and pulse interference, and its disturbance amplitude is dynamically adjusted according to the safety margin calculated by S2-3.

7. The brake pressure degradation strategy of the electromechanical brake system according to claim 1, characterized in that: The S6 specifically comprises the following steps: S6-1. Risk assessment grading mechanism: The prediction results of the brake safety margin optimized by S3-2 are divided into three levels, among which the judgment threshold of the red warning level is analyzed for spatial overlap with the high-risk area distribution map generated by S1-4; S6-2. Human-machine collaborative response channel: establish a human-machine collaborative control loop, the warning signal of S6-1 drives the execution device in real time through the communication link established by S4-1, and the yellow level virtual track projection function calls the road grid map data generated by S1-2; S6-3. Driving behavior modeling engine: Builds a driver profile model based on the thermodynamic parameters output by S3-3. By analyzing nearly 100 braking operation data recorded by S2-1, it generates a personalized correction coefficient and feeds it back to the Q learning decision tree of S3-1. S6-4. Safety strategy dynamic corrector: Dynamically corrects the pre-processing trigger conditions. Its core parameters integrate the real-time road adhesion coefficient of S2-3 and the lower limit of the confidence interval calculated by S5-2 to form a closed-loop safety control strategy.

8. The brake pressure degradation strategy of the electromechanical brake system according to claim 1, characterized in that: The S7 specifically comprises the following steps: S7-1. Multi-physics simulation modeling layer: Construct a multi-physics joint simulation environment, use the thermal boundary conditions output by S3-4 and the driver behavior model of S6-3 as initial parameters, and solve the three-dimensional unsteady-state heat conduction differential equation through implicit difference algorithm; S7-2. Verification benchmark adaptive adjuster: defines the system verification benchmark, where the pressure tracking error threshold is positively correlated with the temperature compensation coefficient of S2-3, and the temperature prediction deviation allowable range is dynamically shrunk according to the model accuracy of S7-1; S7-3. Fault mode injection analyzer: construct a fault injection test set under the closed-loop strategy framework formed in S6-4, in which the sample weights of the short-circuit fault mode are adaptively allocated according to the adversarial training results of S5-4; S7-4. Distributed knowledge storage architecture: Create a distributed parameter storage architecture. The decision parameters generated by S3-1 reinforcement learning are uploaded to the blockchain node through the S4-3 encrypted channel, while a compressed knowledge copy matching the S5-1 signal features is retained locally.

Images