Automotive braking system monitoring system

Abstract

This application relates to the field of brake monitoring technology and provides a monitoring system for automotive braking systems. By analyzing the temporal correlation between electromyography signals and brake pedal travel signals, it can effectively distinguish the user's braking intentions, avoiding misjudgments of faults due to a single mechanical signal. Through cross-validation of brake fluid status and correction of road adhesion coefficient, fault judgment is no longer limited to a single parameter or fixed operating conditions, significantly improving the adaptability and accuracy of fault identification under complex road conditions. Based on the deviation analysis between the virtual braking force output by the simulation model and the actual braking force, it overcomes the limitation of existing technologies that can only determine the existence of faults. By iteratively updating the alarm threshold through reinforcement learning and based on the fault root cause and probability-triggered hierarchical compensation strategy, while optimizing the reward function with the deviation between the actual braking distance and the target distance, it significantly improves the safety and reliability of automotive braking systems.

Description

Technical Field

[0001] This application relates to the field of brake monitoring technology, and in particular to a monitoring system for automotive brake systems. Background Technology

[0002] The monitoring technology of automotive braking systems has been evolving with the development of the automotive industry. Early brake monitoring mainly relied on the physical feedback of mechanical structures, such as brake pad wear alarm devices, which triggered alarms through mechanical contacts. With the introduction of electronic technology, monitoring methods based on single sensors have gradually developed. For example, wheel speed sensors monitor wheel speed differences to determine brake lock-up, and pressure sensors monitor brake line pressure changes to identify pressure anomalies.

[0003] In recent years, with the application of multi-sensor fusion technology, brake monitoring systems have begun to integrate multi-dimensional data such as wheel speed, pressure, and temperature, achieving fault warnings through simple logical judgments or threshold comparisons. Some advanced systems have introduced basic algorithm models for fault type differentiation. However, existing systems only judge braking behavior based on brake pedal travel or pressure signals, failing to distinguish braking intent and easily leading to false fault diagnosis. Furthermore, most systems rely on a single parameter for fault judgment, neglecting key factors such as brake fluid status and road conditions, resulting in low fault identification accuracy under complex operating conditions. Simultaneously, existing technologies can only identify the existence of a fault but cannot accurately pinpoint its cause, leading to a lack of targeted subsequent repairs or compensation. Most systems have not yet addressed how to differentiate user intent to achieve fault diagnosis, cause labeling, and dynamic compensation under complex operating conditions. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this application provides a monitoring system for automotive braking systems, which includes: a sensor fusion module, a simulation verification module, and a graded compensation module.

[0005] The sensor fusion module is used to acquire electromyography (EMG) signals when the user holds the steering wheel and extract EMG signal features. The EMG signal features are then correlated with the brake pedal travel signal in a time sequence to determine the user's intention. If the braking is not the user's intention, the cross-validation of the brake fluid status is activated to determine the failure probability of the vehicle's braking system. At the same time, the vehicle camera is linked to output the road adhesion coefficient through a target detection algorithm to correct the failure probability.

[0006] The simulation verification module is used to establish a simulation model of the vehicle braking system, receive the fault probability, brake fluid state and road adhesion coefficient, drive the simulation model to output virtual braking force, calculate the deviation rate between the actual braking force and the virtual braking force, and respond to the comparison with the alarm threshold and combine it with the fault probability to mark the root cause of the fault.

[0007] The graded compensation module is used to iteratively update the alarm threshold based on reinforcement learning, and to trigger a graded compensation strategy based on the root cause and probability of the fault. The difference between the actual braking distance after the graded compensation strategy is executed and the target braking distance is used to incrementally update the reward function of reinforcement learning.

[0008] As an optional implementation, the fault probability correction strategy includes:

[0009] The system uses a vehicle-mounted camera to acquire road surface images and then processes these images using a target detection algorithm to extract the texture and geometric features of the road surface.

[0010] The road surface adhesion coefficient is determined based on the texture and geometric features of the road surface, and a corrected mapping table between the road surface adhesion coefficient and the fault type is established.

[0011] The failure probability is corrected by applying correction factors of different road adhesion coefficients to the failure probability based on the correction mapping table.

[0012] As an optional implementation, the user intent determination sub-strategy includes:

[0013] Simultaneously acquire electromyographic signals and brake pedal travel signals when the user holds the steering wheel, extract electromyographic signal features from the electromyographic signals, including muscle contraction intensity, muscle contraction frequency, and contraction duration;

[0014] The displacement change rate and acceleration features are extracted from the travel signal, and the temporal correlation matrix between the electromyographic signal features and the travel signal features is established to determine the temporal matching degree between the electromyographic signal and the travel signal.

[0015] Based on the differences in changes in electromyographic signal characteristics and travel signal characteristics, as well as the temporal matching degree, the user's intention is comprehensively judged. The user's intention includes active braking by the user and braking without user intention.

[0016] As an optional implementation, the sub-strategy for determining the failure probability includes:

[0017] The dielectric constant of the brake fluid, the pressure fluctuation value of the pressure sensor, and the wheel speed difference value of the wheel speed sensor are obtained. Based on the dielectric constant of the brake fluid, the water content and performance degradation coefficient of the brake fluid are determined to judge the condition of the brake fluid.

[0018] A multidimensional evidence vector is constructed, which includes user intent, performance degradation coefficient, pressure fluctuation value and wheel speed difference. The multidimensional evidence vector is then input into a Bayesian network to calculate the posterior probability of the fault type.

[0019] The posterior probability is smoothed over time using the Kalman filter algorithm to determine the fault probability.

[0020] As an optional implementation, the root cause labeling strategy includes:

[0021] The brake master cylinder pressure and wheel deceleration are obtained by pressure sensor and wheel speed sensor respectively, so as to infer the actual braking force;

[0022] Calculate the deviation rate between actual braking force and virtual braking force, and establish a mapping table of alarm threshold, fault probability threshold and fault root cause;

[0023] The deviation rate and failure probability are searched and matched in the mapping table. If a single root cause of failure is matched, the root cause of failure is marked.

[0024] If multiple root causes of failure are matched at the same time, the failure probability and the deviation rate are multiplied together to obtain the evidence weight value for each matched root cause of failure, and the root causes of failure are selected and marked based on the evidence weight value.

[0025] As an optional implementation, the output sub-strategy of the virtual braking force includes:

[0026] Establish a simulation model of the vehicle braking system and receive the failure probability, performance degradation coefficient, and road adhesion coefficient;

[0027] The fault injection parameters of the simulation model are adjusted according to the fault probability, the brake fluid viscosity and compression parameters are corrected based on the performance degradation coefficient, and the tire friction parameters are adjusted according to the road adhesion coefficient.

[0028] The result of judging the user's intent is used as input stimulus to drive the simulation model to run. At the same time, the braking friction torque and wheel braking torque output by the simulation model are obtained to determine the output virtual braking force.

[0029] As an optional implementation, the graded compensation strategy refers to establishing a decision table that includes the root cause of the fault, the probability of the fault, and the compensation level. Based on the marked root cause of the fault and the determined probability of the fault, the compensation level is searched in the decision table, and the vehicle braking system is graded and compensated based on the compensation level.

[0030] As an optional implementation, the alarm threshold update sub-strategy includes:

[0031] Construct a state vector that includes the root cause of the fault, the deviation rate, the road surface adhesion coefficient, and the vehicle speed, and input the state vector into the reinforcement learning agent;

[0032] The agent calculates the expected value of the alarm threshold under different adjustment actions based on the Q-learning algorithm, and selects the adjustment action according to the expected value.

[0033] The alarm thresholds are iteratively updated based on the adjustment actions, and the results are fed back to the mapping table.

[0034] As an optional implementation, the alarm threshold update sub-strategy includes:

[0035] Construct a state vector that includes the root cause of the fault, the deviation rate, the road surface adhesion coefficient, and the vehicle speed, and input the state vector into the reinforcement learning agent;

[0036] The agent calculates the expected value of the alarm threshold under different adjustment actions based on the Q-learning algorithm, and selects the adjustment action according to the expected value.

[0037] The alarm thresholds are iteratively updated based on the adjustment actions, and the results are fed back to the mapping table.

[0038] Compared with existing technologies, the beneficial effects of this application are as follows: By analyzing the temporal correlation between electromyography signals and brake pedal travel signals, the user's braking intention can be effectively distinguished, avoiding misjudgment of faults due to a single mechanical signal, providing a reliable premise for subsequent fault diagnosis. Furthermore, through cross-validation of brake fluid status and correction of road adhesion coefficient, the vehicle's braking system status and external environmental factors are incorporated into the fault probability calculation, making fault judgment no longer limited to a single parameter or fixed operating condition, significantly improving the adaptability and accuracy of fault identification under complex road conditions. Based on the deviation analysis between the virtual braking force output by the simulation model and the actual braking force, a multi-dimensional mapping relationship is constructed in conjunction with the fault probability, breaking through the limitation of existing technologies that can only determine the existence of faults. This allows for precise labeling of fault root causes, laying the foundation for targeted compensation. Through reinforcement learning, alarm thresholds are iteratively updated, and a graded compensation strategy is triggered based on the fault root cause and probability. Simultaneously, the reward function is optimized using the deviation between the actual braking distance and the target distance, achieving continuous self-learning and optimization of the graded compensation strategy, significantly improving the safety and reliability of the vehicle's braking system. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0040] Figure 1 This is a system flowchart of the monitoring system for the automotive braking system provided in an embodiment of this application;

[0041] Figure 2 This is a sub-strategy diagram for determining user intent in the monitoring system of the automotive braking system provided in this application embodiment;

[0042] Figure 3 This is a sub-strategy diagram for updating the alarm threshold of the monitoring system for the automotive braking system provided in this application embodiment. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the embodiments of this application more apparent and understandable, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0044] Example

[0045] like Figure 1 The diagram shown is a system flowchart of a monitoring system for an automotive braking system provided in this application embodiment. The system includes a sensor fusion module, a simulation verification module, and a graded compensation module.

[0046] The sensor fusion module is used to acquire electromyography (EMG) signals when the user holds the steering wheel and extract EMG signal features. The EMG signal features are then correlated with the brake pedal travel signal in a time series to determine the user's intention. If the braking is not the user's intention, the cross-validation of the brake fluid status is activated to determine the failure probability of the vehicle's braking system. At the same time, the onboard camera is linked to output the road surface adhesion coefficient through a target detection algorithm to correct the failure probability.

[0047] Furthermore, such as Figure 2 As shown, the sub-strategies for determining user intent include:

[0048] Simultaneously acquire electromyographic signals and brake pedal travel signals when the user holds the steering wheel, extract electromyographic signal features from the electromyographic signals, including muscle contraction intensity, muscle contraction frequency, and contraction duration;

[0049] The displacement change rate and acceleration features are extracted from the travel signal, and the temporal correlation matrix between the electromyographic signal features and the travel signal features is established to determine the temporal matching degree between the electromyographic signal and the travel signal.

[0050] Based on the differences in changes in electromyographic signal characteristics and travel signal characteristics, as well as the temporal matching degree, the user's intention is comprehensively judged. The user's intention includes active braking by the user and braking without user intention.

[0051] Traditional methods rely solely on brake pedal travel signals to determine braking intent, which struggles to distinguish between active driver input and involuntary braking caused by mechanical malfunctions. Combining this with driver electromyography (EMG) signals enhances accuracy. Distributed EMG sensors are embedded in the circular grip area of ​​the steering wheel, with even spacing to cover the primary force application area of ​​the hand. A travel sensor is also installed at the brake pedal's rotation axis. When the EMG sensors acquire EMG signals, a bandpass filter filters out power frequency interference and high-frequency noise. A sliding window method is then used to extract muscle contraction intensity, frequency, and duration. Muscle contraction intensity represents the peak electrical signal intensity during muscle contraction, frequency represents the number of peak signal occurrences per unit time, and duration represents the duration the signal exceeds the baseline. When the travel sensor acquires travel signals, a low-pass filter eliminates mechanical vibration interference, and displacement rate of change and acceleration are extracted. Displacement rate of change represents the amount of pedal displacement change per unit time, and acceleration represents the change in displacement rate of change. This multi-dimensional extraction of EMG and travel signal features avoids judgment bias caused by single features and provides a rich data foundation for subsequent time-series correlation analysis.

[0052] The temporal correlation between electromyographic (EMG) signals and travel signals is crucial for determining driver intent. This requires quantifying the temporal matching relationship using a matrix to avoid subjective judgment errors. A temporal correlation matrix is ​​constructed using fixed time units, with rows representing the acquisition time of EMG signal features and columns representing the acquisition time of travel signal features. A cross-correlation algorithm is used to calculate the similarity between EMG and travel signal features at different times; a higher similarity value indicates a stronger correlation between the two signals at that time. The similarity values ​​at each time point in the temporal correlation matrix are then weighted and summed to obtain the overall temporal matching degree. The weight allocation is determined based on the degree of influence of EMG and travel signal features on braking intent, with higher weights for muscle contraction intensity and displacement change rate. This transforms the abstract temporal correlation relationship into quantifiable matrix data, making the calculation of the temporal matching degree more objective and consistent.

[0053] Relying solely on timing matching degree or signal feature changes is insufficient to comprehensively determine user intent; a combined analysis of both is necessary to improve accuracy. When the muscle contraction intensity and frequency increase sharply in the electromyographic (EMG) signal characteristics, and the displacement change rate and acceleration increase synchronously in the stroke signal characteristics, while the timing matching degree exceeds the set matching benchmark value, it is determined to be active braking by the user. When the displacement change rate and acceleration increase in the stroke signal characteristics, but the EMG signal characteristics show no significant change, and the timing matching degree is less than the set matching benchmark value, it is determined to be non-user-intended braking. If a contradiction arises between feature changes and timing matching degree, a secondary acquisition and verification process is initiated to eliminate the influence of momentary sensor malfunctions. By comprehensively judging multiple conditions, misjudgments caused by single-factor anomalies are reduced, improving the accuracy of user intent determination. The output result of active or non-user-intended braking directly determines whether to activate cross-validation of the brake fluid status, which is a key trigger condition for initiating fault troubleshooting during non-user-intended braking.

[0054] Furthermore, the sub-strategy for determining the failure probability includes:

[0055] The dielectric constant of the brake fluid, the pressure fluctuation value of the pressure sensor, and the wheel speed difference value of the wheel speed sensor are obtained. Based on the dielectric constant of the brake fluid, the water content and performance degradation coefficient of the brake fluid are determined to judge the condition of the brake fluid.

[0056] A multidimensional evidence vector is constructed, which includes user intent, performance degradation coefficient, pressure fluctuation value and wheel speed difference. The multidimensional evidence vector is then input into a Bayesian network to calculate the posterior probability of the fault type.

[0057] The posterior probability is smoothed over time using the Kalman filter algorithm to determine the fault probability.

[0058] Automotive braking system failures can be caused by a variety of factors, such as brake fluid performance degradation, abnormal pressure transmission, or wheel speed imbalance. It is necessary to acquire key parameters using multiple types of sensors to provide basic data for determining the probability of failure. A dielectric constant sensor is installed at the outlet line of the brake fluid reservoir to monitor changes in the brake fluid's dielectric constant in real time. A pressure sensor is installed at the brake master cylinder outlet to obtain pressure fluctuations in the automotive braking system. Wheel speed sensors are installed at the rims of each wheel to obtain the wheel speed and calculate the wheel speed difference. Based on the data monitored by the dielectric constant sensors and the reference dielectric constant value corresponding to the brake fluid type, it is determined whether the water content of the brake fluid exceeds the standard range. Simultaneously, based on the usage time and the trend of dielectric constant change, a performance degradation coefficient is determined to reflect the brake fluid condition. By comprehensively acquiring key parameters of the automotive braking system through multiple types of sensors, multi-dimensional data support is provided for subsequent determination of the probability of failure.

[0059] A single parameter anomaly is insufficient to accurately determine the fault type and probability. Therefore, multi-dimensional parameters need to be fused and analyzed, and the probability of the fault is calculated using probabilistic reasoning methods. The user intent judgment result, brake fluid performance degradation coefficient, pressure fluctuation value, and wheel speed difference are integrated into a multi-dimensional evidence vector. Each parameter is standardized according to its impact on the fault, with higher weights assigned to non-user-intended braking. This multi-dimensional evidence vector is then input into a pre-trained Bayesian network built on a large amount of historical fault data, including the correlation probability between fault types and each multi-dimensional evidence vector. The posterior probability of different fault types, including brake fluid performance degradation faults, hydraulic system faults, and wheel speed sensor faults, is calculated through Bayesian network reasoning. By fusing multi-dimensional evidence and using probabilistic reasoning, a quantitative assessment of the probability of various faults can be achieved, avoiding the limitations of single-parameter judgment.

[0060] Instantaneous fluctuations in sensor readings can cause jumps in posterior probability calculations. Temporal smoothing is necessary to eliminate these transient interferences and ensure the fault probability better reflects the true fault trend. A Kalman filter algorithm is used to smooth the posterior probability output by the Bayesian network. In the prediction step, the fault probability at the current time is predicted based on the fault probability and fault development trend from the previous time step. Then, in the update step, the predicted fault probability is corrected by combining the posterior probability calculation results from the current time step, resulting in a smoothed fault probability. During the filtering process, filtering parameters are adjusted according to the stability characteristics of the fault type, with a stronger smoothing coefficient used for fault types with poor stability. This reduces the impact of instantaneous sensor noise on fault probability determination, making the fault probability change more stable and better reflecting the true fault development trend. The smoothed fault probability serves as the initial data for fault probability correction, providing a benchmark value for subsequent corrections based on the road surface adhesion coefficient.

[0061] Specifically, the strategies for correcting the failure probability include:

[0062] The system uses a vehicle-mounted camera to acquire road surface images and then processes these images using a target detection algorithm to extract the texture and geometric features of the road surface.

[0063] The road surface adhesion coefficient is determined based on the texture and geometric features of the road surface, and a corrected mapping table between the road surface adhesion coefficient and the fault type is established.

[0064] The failure probability is corrected by applying correction factors of different road adhesion coefficients to the failure probability based on the correction mapping table.

[0065] The manifestations and impacts of vehicle braking system failures vary under different road surface conditions. It is necessary to determine the road adhesion coefficient based on road surface characteristics to provide a basis for fault probability correction. By linking the vehicle's onboard camera to acquire road surface images in real time, and processing these images using target detection algorithms, road surface types are identified, including dry asphalt, wet / slippery surfaces, and snow-covered surfaces. Texture and geometric features are extracted, including texture coarseness and uniformity, and geometric features such as slope and smoothness. The extracted features are then denoised and standardized to eliminate interference from environmental factors such as lighting and shadows. This allows for accurate extraction of road surface features through image recognition technology, providing reliable feature data for subsequent determination of the road adhesion coefficient.

[0066] The road surface adhesion coefficient directly affects the performance and fault behavior of the vehicle braking system. It is necessary to establish its correlation with fault types to achieve reasonable correction of fault probabilities. Based on extracted road texture and geometric features, and combined with a vehicle dynamics model, the road surface adhesion coefficient is comprehensively determined. This model includes parameters such as deceleration and wheel speed changes during vehicle braking. Through numerous fault simulation experiments under different road conditions, the differences between the actual and theoretical probabilities of various fault types under different road surface adhesion coefficients are statistically analyzed to establish a correction mapping table between the road surface adhesion coefficient and fault types. This clarifies the correction ratio for various fault probabilities corresponding to different road surface adhesion coefficients. Therefore, by establishing a correction mapping table based on experimental data, the correction of fault probabilities by the road surface adhesion coefficient becomes more scientific and targeted.

[0067] The manifestation and impact of the same fault vary under different road surface conditions. Therefore, the initial fault probability needs to be corrected based on the road surface adhesion coefficient to better reflect actual road conditions. Based on the calculated road surface adhesion coefficient, the corresponding correction factor for each type of fault is found in the correction mapping table. For faults more easily exposed on surfaces with low road surface adhesion coefficients (including wet and snowy surfaces), a larger correction factor is applied to increase their fault probability. In these cases, the most easily exposed fault is brake fluid performance degradation. For faults less affected by road conditions, such as wheel speed sensor faults, a smaller correction factor is applied. The initial fault probability is then multiplied by the corresponding correction factor to obtain the corrected fault probability. This process ensures that the corrected fault probability fully considers road environment factors, better reflects actual fault occurrence patterns, and improves the accuracy of fault diagnosis. The corrected fault probability is then input into the simulation verification module to provide more accurate basic data for calculating virtual braking force and identifying fault root causes.

[0068] The simulation verification module is used to establish a simulation model of the vehicle braking system. It receives the fault probability, brake fluid state and road adhesion coefficient, drives the simulation model to output virtual braking force, calculates the deviation rate between the actual braking force and the virtual braking force, and compares it with the alarm threshold and combines it with the fault probability to mark the root cause of the fault.

[0069] Furthermore, the output sub-strategy of virtual braking force includes:

[0070] Establish a simulation model of the vehicle braking system and receive the failure probability, performance degradation coefficient, and road adhesion coefficient;

[0071] The fault injection parameters of the simulation model are adjusted according to the fault probability, the brake fluid viscosity and compression parameters are corrected based on the performance degradation coefficient, and the tire friction parameters are adjusted according to the road adhesion coefficient.

[0072] The result of judging the user's intent is used as input stimulus to drive the simulation model to run. At the same time, the braking friction torque and wheel braking torque output by the simulation model are obtained to determine the output virtual braking force.

[0073] Traditional automotive braking system simulation models rely solely on fixed operating parameters, failing to reflect the impact of real-time fault states, brake fluid performance degradation, and changes in road conditions on braking force. This results in significant discrepancies between virtual and actual braking forces, hindering their use in fault diagnosis. This paper constructs a new simulation model for the automotive braking system. This model includes a brake fluid flow simulation unit, a brake execution simulation unit, and a wheel dynamics simulation unit. The brake fluid flow simulation unit simulates the pressure transmission of brake fluid in the pipeline; the brake execution simulation unit simulates the contact friction between the brake pads and brake discs; and the wheel dynamics simulation unit simulates the force and motion state of the wheel. The simulation model receives fault probabilities, brake fluid performance degradation coefficients, and road adhesion coefficients from the sensor fusion module via a CAN bus interface. Fault probabilities include brake fluid failure probabilities, mechanical structure failure probabilities, and sensor failure probabilities. These parameters are converted into directly callable variables by the simulation model's internal data parsing module. The simulation model covers the entire link of the automotive braking system from pressure generation to braking force output. Combined with real-time input parameters, it can dynamically reflect the state changes of the automotive braking system, providing a realistic simulation model foundation for virtual braking force calculation.

[0074] Fault conditions can cause partial malfunctions in the vehicle's braking system. For example, leaks in the piping lead to pressure loss, brake fluid performance degradation reduces its transmission efficiency, and changes in the road surface adhesion coefficient alter the maximum braking force of the tires. If the simulation model's parameters are fixed, they cannot reflect the impact of these factors, resulting in distorted virtual braking force calculations. A pre-defined fault injection parameter library is used. Brake fluid faults correspond to piping pressure loss coefficients (the higher the fault probability, the larger the pressure loss coefficient); mechanical structure faults correspond to the contact gap parameters between the brake pads and brake discs (the higher the fault probability, the larger the gap value); and sensor faults correspond to signal drift parameters (the higher the fault probability, the larger the drift). Then, matching fault injection parameters are retrieved from the fault injection parameter library based on the fault probability. The viscosity and compressibility coefficients in the brake fluid flow simulation unit are corrected based on the correlation curve between the performance degradation coefficient and the brake fluid viscosity and compressibility. Furthermore, the tire friction parameters in the wheel dynamics simulation unit are adjusted based on the mapping relationship between the road surface adhesion coefficient and the tire longitudinal friction coefficient. By dynamically adjusting the fault injection parameters of the simulation model, the model can simulate braking force characteristics under different fault levels, brake fluid performance states, and road conditions, significantly reducing the deviation between virtual and actual braking forces.

[0075] The braking force output directly responds to the user's braking intention. Light pedal depressing corresponds to small braking force, while heavy pedal depressing corresponds to large braking force. If the simulation model input does not reflect the difference in intention, the virtual braking force will not match the actual braking behavior. The judgment result of the user's intention is converted into the input excitation of the simulation model. When the user actively brakes, an analog voltage signal proportional to the stroke signal is input. When braking is not the user's intention, a preset abnormal drive signal, such as a pulse voltage signal, is input. The input excitation is then converted into control commands for the braking execution simulation unit to drive the simulation model. During the simulation model operation, the braking friction torque is obtained through the braking execution simulation unit, and the wheel braking torque is obtained through the wheel dynamics simulation unit. Then, the virtual braking force is calculated by combining the wheel radius parameter, where virtual braking force = wheel braking torque / wheel radius. This allows the magnitude and trend of the virtual braking force to be directly related to the user's operation behavior, improving the relevance and accuracy of the calculation results. The output virtual braking force provides a benchmark value for calculating the deviation rate between the actual braking force and the virtual braking force, and its accuracy directly affects the reliability of fault root cause identification.

[0076] Specifically, the root cause labeling strategy includes:

[0077] The brake master cylinder pressure and wheel deceleration are obtained by pressure sensor and wheel speed sensor respectively, so as to infer the actual braking force;

[0078] Calculate the deviation rate between actual braking force and virtual braking force, and establish a mapping table of alarm threshold, fault probability threshold and fault root cause;

[0079] The deviation rate and failure probability are searched and matched in the mapping table. If a single root cause of failure is matched, the root cause of failure is marked.

[0080] If multiple root causes of failure are matched at the same time, the failure probability and the deviation rate are multiplied together to obtain the evidence weight value for each matched root cause of failure, and the root causes of failure are selected and marked based on the evidence weight value.

[0081] Actual braking force is a core indicator for evaluating the working state of a vehicle's braking system, but it cannot be directly measured by sensors. It needs to be indirectly derived through relevant parameters to provide a basis for comparison with virtual braking force. A pressure sensor is installed at the brake master cylinder outlet, and the pressure signal is converted into a voltage signal through a signal conditioning circuit. Wheel speed sensors are installed on the inner side of each wheel rim to obtain wheel speed signals, which are then processed to obtain wheel deceleration, where wheel deceleration = (current speed - previous speed) / sampling time. Based on the transmission relationship between brake master cylinder pressure and brake line pressure, combined with the effective area of ​​the brake pads, the braking friction force is calculated, where brake line pressure = brake master cylinder pressure × line pressure loss coefficient, and braking friction force = brake line pressure × effective area of ​​brake pads. Then, the dynamic relationship between wheel deceleration and ground braking force is combined, where ground braking force = vehicle mass × deceleration / 4, where 4 is the number of wheels. The average value of braking friction force and ground braking force is taken as the actual braking force. The actual braking force is obtained indirectly through sensor technology and dynamic principles, taking into account the real-time performance and reliability of the data, and avoiding the technical difficulties of direct measurement.

[0082] The deviation rate quantifies the difference between actual braking force and virtual braking force. The greater the difference, the more serious the abnormality of the vehicle's braking system. However, it is necessary to combine this with the failure probability to determine the specific failure type. The mapping table is a key tool for achieving this correlation. The deviation rate between actual braking force and virtual braking force is calculated, where deviation rate = |actual braking force - virtual braking force| / virtual braking force. Then, based on the statistical analysis of historical failure data, alarm thresholds and failure probability thresholds are determined, and a mapping table is established with the root causes of failures. Root causes of failures include brake fluid performance degradation, brake pad sticking, and wheel speed sensor malfunctions. The mapping table clarifies the correspondence between the deviation rate and failure probability and the root causes of failures, upgrading failure diagnosis from qualitative judgment to quantitative analysis and improving the objectivity of root cause identification.

[0083] In single-fault scenarios, such as brake fluid performance degradation alone, a clear matching rule is needed to quickly locate the root cause, providing a direct target for subsequent graded compensation strategies. The calculated deviation rate and fault probability are input into a mapping table, and the search is performed according to the rule of prioritizing deviation rate intervals and using fault probability as an auxiliary verification. When a match is successful, the root cause of the fault is stored in the fault code memory, and the corresponding indicator light signal is triggered. Thus, precise matching enables rapid marking of a single fault root cause, shortening fault diagnosis time and improving response efficiency.

[0084] In complex fault scenarios, such as brake fluid performance degradation accompanied by wheel speed sensor malfunction, multiple root causes simultaneously meet matching conditions. Weight analysis is needed to determine the dominant root cause. When multiple entries in the mapping table simultaneously meet the deviation rate and fault probability conditions, the evidence weight value for each entry is calculated, where evidence weight value = fault probability × deviation rate. The magnitudes of the evidence weight values ​​are then compared, and the root cause with the largest value is selected as the primary root cause, stored in the fault code memory, and secondary root causes are recorded simultaneously. Through the analysis of evidence weight values, the dominant root cause can be accurately identified in multi-fault scenarios, avoiding diagnostic confusion caused by fault correlation. This ensures that the tiered compensation module prioritizes measures for the primary root cause, improving the effectiveness and relevance of the tiered compensation strategy.

[0085] The graded compensation module is used to iteratively update the alarm threshold based on reinforcement learning, and to trigger a graded compensation strategy based on the root cause and probability of the fault. The difference between the actual braking distance after the graded compensation strategy is executed and the target braking distance is used to incrementally update the reward function of reinforcement learning.

[0086] Furthermore, such as Figure 3 As shown, the sub-strategy for updating the alarm threshold includes:

[0087] Construct a state vector that includes the root cause of the fault, the deviation rate, the road surface adhesion coefficient, and the vehicle speed, and input the state vector into the reinforcement learning agent;

[0088] The agent calculates the expected value of the alarm threshold under different adjustment actions based on the Q-learning algorithm, and selects the adjustment action according to the expected value.

[0089] The alarm thresholds are iteratively updated based on the adjustment actions, and the results are fed back to the mapping table.

[0090] Traditional alarm thresholds are mostly fixed values, which cannot adapt to the characteristics of the vehicle braking system under different fault types, road conditions, and vehicle speeds, easily leading to false alarms or missed alarms. A dynamic adjustment mechanism is needed to improve the adaptability of alarm thresholds. The state vector includes the root cause of the fault, the deviation rate between the actual braking force and the virtual braking force, the road adhesion coefficient, and the vehicle speed. The state vector is input into a reinforcement learning agent, which includes a policy network and a value network. The policy network is responsible for outputting the adjustment action of the alarm threshold, and the value network is responsible for evaluating the expected effect of the adjustment action. The state vector comprehensively covers the key factors affecting the safety of the vehicle braking system, providing the agent with a complete decision-making basis. The reinforcement learning framework enables the alarm threshold adjustment to have autonomous learning capabilities and can adapt to complex and changing operating conditions.

[0091] Different adjustment actions have varying effects on optimizing alarm thresholds. Quantitative evaluation is needed to select the optimal action to avoid blind adjustments that could lead to inaccurate thresholds. The agent is pre-defined with three types of adjustment actions: raising the alarm threshold, lowering the alarm threshold, and maintaining the current alarm threshold. The Q-learning algorithm calculates the expected value of each adjustment action in the current state vector by querying a historical experience table, prioritizing the action with the highest expected value. The expected value reflects the potential contribution of the adjustment action to reducing false alarms or missed alarms. If no matching state is found in the historical experience table, an ε-greedy strategy is used to randomly select an action, where the ε value gradually decreases with the number of learning iterations to ensure both exploration of new operating conditions and decision stability. Thus, the Q-learning algorithm achieves quantitative evaluation of adjustment actions, making alarm threshold adjustments more targeted. The ε-greedy strategy balances exploration and utilization, ensuring that the agent can still find the optimal adjustment direction under complex operating conditions.

[0092] The adjustment action needs to be translated into a specific alarm threshold update value, and the updated alarm threshold needs to be applied to the fault root cause marking process in real time to form a closed-loop optimization. The adjustment range of the alarm threshold is preset for different fault root causes; for example, the threshold adjustment step for brake fluid-related faults is smaller than that for mechanical faults. Then, according to the selected adjustment action, the alarm threshold is updated according to the corresponding step size. The updated alarm threshold is fed back to the mapping table of the simulation verification module, overwriting the entries of the original alarm threshold, ensuring that subsequent fault root cause marking uses the latest alarm threshold. The alarm threshold adjustment step size matches the fault type to avoid over-adjustment causing oscillations in the vehicle's braking system. The real-time feedback mechanism ensures that the mapping table remains timely, improving the accuracy of fault root cause marking and providing an updated alarm threshold reference for the graded compensation strategy, ensuring that the triggering conditions of the graded compensation strategy are synchronized with the current state of the vehicle's braking system.

[0093] Specifically, the graded compensation strategy refers to establishing a decision table that includes the root cause of the fault, the probability of the fault, and the compensation level. Based on the marked root cause of the fault and the determined probability of the fault, the compensation level is found in the decision table, and the vehicle braking system is compensated in a graded manner based on the compensation level.

[0094] The severity and required intervention measures vary significantly among different root causes of failures. A structured decision table is needed to clearly define the correspondence between compensation levels and failure states, avoiding confusion in compensation strategies. The decision table's row dimension represents the root cause of the failure, the column dimension represents the failure probability, and the table cells represent the corresponding compensation levels, including Level 1, Level 2, and Level 3 compensation. Extensive bench tests and simulations are used to determine reasonable compensation levels for different root causes under various failure probabilities. For example, a medium probability of brake fluid performance degradation corresponds to Level 2 compensation, while a high probability of mechanical jamming corresponds to Level 3 compensation. The decision table makes the rules for determining compensation levels transparent and standardized, avoiding human error. The correspondence based on experimental data ensures the scientific validity and effectiveness of the tiered compensation strategy.

[0095] The root cause and probability of a fault are the core factors determining the compensation intensity. Precise matching is needed to find the corresponding compensation level, ensuring that the compensation measures match the severity of the fault. The marked root causes and fault probabilities output by the sensor fusion module are obtained from the simulation verification module. The corresponding rows and columns are located in the decision table. If the fault probability is at the intersection of two intervals, a higher compensation level is automatically matched. If there is no perfectly matching root cause in the decision table, a fuzzy matching mechanism is triggered to select the compensation level corresponding to the most similar root cause. Through precise matching and boundary handling rules, the determination of the compensation level ensures that it conforms to preset rules and can cope with edge conditions. The fuzzy matching mechanism improves the adaptability of the decision table to unforeseen fault types.

[0096] Different compensation levels correspond to different fault severity levels, requiring differentiated measures, ranging from minor adjustments to emergency intervention, to avoid over-compensation affecting the driving experience or under-compensation leading to safety risks. Level 1 compensation corresponds to low-probability faults, including adjusting the time window for EMG signal feature extraction and the sliding average filtering of the brake pedal travel signal. Level 2 compensation corresponds to medium-probability faults, including reducing the target pressure of the brake master cylinder and activating the brake fluid circulation pump, thus accelerating brake fluid temperature equalization. Level 3 compensation corresponds to high-probability faults, including switching to an independent electromechanical braking redundancy circuit, triggering automatic emergency braking through the body controller, and simultaneously sending a deceleration command to the powertrain. The tiered compensation strategy matches the fault severity, minimizing interference with normal driving while ensuring safety. The combination of redundancy circuits and emergency braking ensures a safety baseline under high-risk faults, providing actual compensation effect data for subsequent reward function updates, and serving as a key input for reinforcement learning closed-loop optimization.

[0097] Specifically, the incremental update strategy for the reward function includes:

[0098] The actual braking distance after the graded compensation strategy is executed is obtained. Based on the vehicle speed, road surface adhesion coefficient and the state of the vehicle braking system, the target braking distance is calculated through a dynamic model.

[0099] The difference between the actual braking distance and the target braking distance is used to obtain the absolute distance error. A multi-dimensional reward vector is constructed that includes the absolute distance error, compensation energy consumption, and the number of false alarms.

[0100] The multidimensional reward vector is input into the reward function of reinforcement learning, and the reward function is incrementally updated through the temporal difference algorithm to adjust the policy network parameters of the agent.

[0101] Actual braking distance is a core indicator for measuring the effectiveness of compensation. It needs to be compared with the target braking distance to evaluate the effectiveness of the graded compensation strategy and provide a basis for updating the reward function. The actual braking distance after compensation is calculated by fusing vehicle-mounted millimeter-wave radar and camera data. That is, the distance traveled from the start of braking to the vehicle stopping. The input parameters of the dynamic model include vehicle speed, road surface adhesion coefficient, and the state of the vehicle braking system. The state includes the compensated braking pressure and the activation status of the redundant loop. The dynamic model outputs the theoretical target braking distance by simulating the braking process, that is, the expected distance under fault-free conditions. The comparison between the actual braking distance and the target braking distance can intuitively reflect the actual effect of the graded compensation strategy. The dynamic model calculates the target braking distance based on real-time parameters to ensure that the evaluation standard matches the current operating conditions.

[0102] A single indicator cannot comprehensively evaluate the compensation strategy. It is necessary to combine the compensation energy consumption and the number of false alarms to ensure that the reward function can guide the agent to optimize multi-objective performance. The absolute difference between the actual braking distance and the target braking distance is calculated, i.e., the absolute distance error. The energy consumption of the compensation process is calculated through the current sensor and hydraulic pump power meter of the vehicle braking system, such as the power consumption of the brake fluid circulation pump in the secondary compensation. Then, the number of times the alarm threshold is triggered but there is no actual fault within a unit of time is counted, i.e., the number of false alarms. The above three parameters are integrated into a multi-dimensional reward vector according to the preset weights. The weights are set according to the safety priority, with the absolute distance error having the highest weight, followed by the number of false alarms, and the compensation energy consumption having the lowest weight. The multi-dimensional reward vector avoids the performance imbalance caused by the optimization of a single indicator and can simultaneously improve braking safety, reduce energy consumption, and reduce false alarms. The weight setting reflects the design principle of safety priority.

[0103] The reward function needs to be dynamically adjusted based on the actual compensation effect to enable the reinforcement learning agent to gradually optimize its strategy, improve long-term decision-making performance, and avoid insufficient adaptability caused by a fixed reward function. A multi-dimensional reward vector is input into the current reward function, and a target value is calculated by combining the current reward and the expected future reward using a temporal difference algorithm. The target value is compared with the agent's prediction to obtain an error signal. Based on the error signal, the weight parameters of the policy network, such as the neuron connection strength, are adjusted using a backpropagation algorithm, making the agent more likely to choose high-reward actions in similar states. Historical records are retained after each update; if new parameters cause performance degradation, the system can backtrack to historical records. The temporal difference algorithm allows the reward function to consider both immediate effects and long-term gains, avoiding short-sighted decisions. The incremental update mechanism ensures smooth adjustment of the reward function parameters, improving the stability of the agent's policy network. The backtracking mechanism reduces the risk of parameter update failure. The updated policy network parameters directly affect the agent's choice of actions to adjust the alarm threshold, continuously improving system performance.

Claims

1. A monitoring system for an automotive braking system, characterized in that, include: The module consists of a sensor fusion module, a simulation verification module, and a hierarchical compensation module. The sensor fusion module is used to acquire electromyography (EMG) signals when the user holds the steering wheel and extract EMG signal features. The EMG signal features are then correlated with the brake pedal travel signal in a time sequence to determine the user's intention. If the braking is not the user's intention, the cross-validation of the brake fluid status is activated to determine the failure probability of the vehicle's braking system. At the same time, the vehicle camera is linked to output the road adhesion coefficient through a target detection algorithm to correct the failure probability. The simulation verification module is used to establish a simulation model of the vehicle braking system, receive the fault probability, brake fluid state and road adhesion coefficient, drive the simulation model to output virtual braking force, calculate the deviation rate between the actual braking force and the virtual braking force, and respond to the comparison with the alarm threshold and combine it with the fault probability to mark the root cause of the fault. The graded compensation module is used to iteratively update the alarm threshold based on reinforcement learning, and to trigger a graded compensation strategy based on the root cause and probability of the fault. The difference between the actual braking distance after the graded compensation strategy is executed and the target braking distance is used to incrementally update the reward function of reinforcement learning.

2. The monitoring system for the automotive braking system as described in claim 1, characterized in that, The fault probability correction strategy includes: The system uses a vehicle-mounted camera to acquire road surface images and then processes these images using a target detection algorithm to extract the texture and geometric features of the road surface. The road surface adhesion coefficient is determined based on the texture and geometric features of the road surface, and a corrected mapping table between the road surface adhesion coefficient and the fault type is established. The failure probability is corrected by applying correction factors of different road adhesion coefficients to the failure probability based on the correction mapping table.

3. The monitoring system for the automotive braking system as described in claim 2, characterized in that, The sub-strategy for determining user intent includes: Simultaneously acquire electromyographic signals and brake pedal travel signals when the user holds the steering wheel, extract electromyographic signal features from the electromyographic signals, including muscle contraction intensity, muscle contraction frequency, and contraction duration; The displacement change rate and acceleration features are extracted from the travel signal, and the temporal correlation matrix between the electromyographic signal features and the travel signal features is established to determine the temporal matching degree between the electromyographic signal and the travel signal. Based on the differences in changes in electromyographic signal characteristics and travel signal characteristics, as well as the temporal matching degree, the user's intention is comprehensively judged. The user's intention includes active braking by the user and braking without user intention.

4. The monitoring system for the automotive braking system as described in claim 3, characterized in that, The sub-strategy for determining the failure probability includes: The dielectric constant of the brake fluid, the pressure fluctuation value of the pressure sensor, and the wheel speed difference value of the wheel speed sensor are obtained. Based on the dielectric constant of the brake fluid, the water content and performance degradation coefficient of the brake fluid are determined to judge the condition of the brake fluid. A multidimensional evidence vector is constructed, which includes user intent, performance degradation coefficient, pressure fluctuation value and wheel speed difference. The multidimensional evidence vector is then input into a Bayesian network to calculate the posterior probability of the fault type. The posterior probability is smoothed over time using the Kalman filter algorithm to determine the fault probability.

5. The monitoring system for the automotive braking system as described in claim 4, characterized in that, The root cause labeling strategy includes: The brake master cylinder pressure and wheel deceleration are obtained by pressure sensor and wheel speed sensor respectively, so as to infer the actual braking force; Calculate the deviation rate between actual braking force and virtual braking force, and establish a mapping table of alarm threshold, fault probability threshold and fault root cause; The deviation rate and failure probability are searched and matched in the mapping table. If a single root cause of failure is matched, the root cause of failure is marked. If multiple root causes of failure are matched at the same time, the failure probability and the deviation rate are multiplied together to obtain the evidence weight value for each matched root cause of failure, and the root causes of failure are selected and marked based on the evidence weight value.

6. The monitoring system for an automotive braking system as described in claim 5, characterized in that, The output sub-strategy of the virtual braking force includes: Establish a simulation model of the vehicle braking system and receive the failure probability, performance degradation coefficient, and road adhesion coefficient; The fault injection parameters of the simulation model are adjusted according to the fault probability, the brake fluid viscosity and compression parameters are corrected based on the performance degradation coefficient, and the tire friction parameters are adjusted according to the road adhesion coefficient. The result of judging the user's intent is used as input stimulus to drive the simulation model to run. At the same time, the braking friction torque and wheel braking torque output by the simulation model are obtained to determine the output virtual braking force.

7. The monitoring system for an automotive braking system as described in claim 6, characterized in that, The graded compensation strategy refers to establishing a decision table that includes the root cause of the fault, the probability of the fault, and the compensation level. Based on the marked root cause of the fault and the determined probability of the fault, the compensation level is searched in the decision table, and the vehicle braking system is graded and compensated based on the compensation level.

8. The monitoring system for an automotive braking system as described in claim 7, characterized in that, The alarm threshold update sub-strategy includes: Construct a state vector that includes the root cause of the fault, the deviation rate, the road surface adhesion coefficient, and the vehicle speed, and input the state vector into the reinforcement learning agent; The agent calculates the expected value of the alarm threshold under different adjustment actions based on the Q-learning algorithm, and selects the adjustment action according to the expected value. The alarm thresholds are iteratively updated based on the adjustment actions, and the results are fed back to the mapping table.

9. The monitoring system for an automotive braking system as described in claim 8, characterized in that, The incremental update strategy for the reward function includes: The actual braking distance after the graded compensation strategy is executed is obtained. Based on the vehicle speed, road surface adhesion coefficient and the state of the vehicle braking system, the target braking distance is calculated through a dynamic model. The difference between the actual braking distance and the target braking distance is used to obtain the absolute distance error. A multi-dimensional reward vector is constructed that includes the absolute distance error, compensation energy consumption, and the number of false alarms. The multidimensional reward vector is input into the reward function of reinforcement learning, and the reward function is incrementally updated through the temporal difference algorithm to adjust the policy network parameters of the agent.

Images