AI-based brake-by-wire system failure prediction method and system

Abstract

The present application relates to the technical field of vehicle braking, in particular to a brake-by-wire system fault prediction method and system based on AI. The method constructs a dynamic correlation graph by acquiring system data, and simulates fault propagation based on the graph to obtain a fault propagation chain and a benchmark degradation time. The degradation time is corrected to determine the earliest failure node time by predicting the node degradation rate in combination with the current working condition, thereby realizing fault prediction. The present application can predict the occurrence of faults in advance and improve the reliability of the braking system.

Description

Technical Field

[0001] This invention relates to the field of vehicle braking technology, and in particular to an AI-based method and system for predicting faults in brake-by-wire systems. Background Technology

[0002] As the core of active safety in modern vehicles, the reliability of the brake-by-wire system directly affects driving safety. This system typically consists of an electronic control unit, sensors, actuators, and a communication network. It transmits control commands via electrical signals to drive hydraulic or electric mechanisms to generate braking force. With increasing vehicle intelligence and system complexity, potential failure modes have become more diverse, placing higher demands on early fault prediction and diagnosis.

[0003] Currently, the industry commonly employs threshold-based alarms and diagnostic fault codes for fault handling in brake-by-wire systems. The system continuously monitors key parameters such as pressure, displacement, and current using sensors. Once the monitored value exceeds a preset static threshold, an alarm is triggered and the corresponding fault code is recorded. This method relies on prior knowledge of known fault modes, and its diagnostic logic is typically solidified during the system design phase.

[0004] However, these conventional methods have significant limitations. Static threshold settings for faults are often based on standard operating conditions or laboratory environments, making it difficult to cover the complex and varied dynamic conditions encountered in real-world driving, such as extreme temperatures, continuous acceleration and deceleration, or different road surface adhesion coefficients. Under dynamic conditions, the normal response range of system components can also shift, leading to false alarms or missed alarms in monitoring based on fixed thresholds, thus affecting the accuracy of predictions.

[0005] Existing methods lack the ability to dynamically model the failure evolution process. Within a brake-by-wire system, components are tightly coupled through force and signal; an initial failure at one node can propagate along a specific physical path, triggering a chain reaction. Conventional threshold alarm methods can only indicate the final node or instantaneous state of the failure, failing to reveal the failure propagation chain or predict the remaining time from initial anomaly to functional failure. This makes it difficult to accurately determine the timing of preventative maintenance and risk intervention. Summary of the Invention

[0006] This invention provides an AI-based method and system for predicting faults in brake-by-wire systems, which can solve the problems in the prior art.

[0007] A first aspect of this invention provides an AI-based method for predicting faults in a brake-by-wire system, comprising: Acquire the measurement data, control command data, and vehicle response data of the actuator of the brake-by-wire system, extract the signal delay and force transmission delay between the data, identify the variable transmission path that conforms to the physical causal order, construct a dynamic correlation graph, and establish a causal edge only when the change time of the previous node is earlier than the response time of the subsequent node and the time difference is within the boundary value of the physical reachable range. The boundary value is determined by the mean and standard deviation of the statistical historical delay samples. When a deviation of the node state variable is detected in the dynamic correlation graph, a fault propagation simulation is performed along the causal edge to obtain the fault propagation chain and the baseline value of the deterioration time of each node. Extract current operating condition parameters, calculate the predicted deterioration rate of each node in the fault propagation chain under the current operating condition through multiple regression analysis, calculate the corrected deterioration time with the baseline value of the deterioration time, and determine the node time that reaches the failure threshold earliest as the fault prediction value. After the fault actually occurs, the boundary values ​​and multiple regression analysis parameters are updated based on the actual fault deviation.

[0008] The steps for extracting signal delay and force transmission delay between data include: The moment when the actuator response signal in the measurement data of the execution unit begins to rise is marked as the execution response moment, and the moment when the amplitude of the control signal in the control command data exceeds the activation threshold is marked as the command issuance moment. The time difference between the execution response moment and the command issuance moment is calculated as the signal delay. The pressure rise rate is identified by performing differential operation on the pressure signal in the measurement data of the execution unit and the moment when it exceeds the establishment criterion is identified as the pressure establishment moment. The deceleration signal in the vehicle response data is filtered and the deceleration is identified by the moment when it exceeds the response criterion as the deceleration response moment. The time difference between the pressure establishment moment and the deceleration response moment is calculated as the force transmission delay.

[0009] Based on time delay identification, a dynamic correlation graph is constructed by identifying variable propagation paths that conform to the physical causal order. Causal edges are established only when the change time of the preceding node is earlier than the response time of the following node and the time difference is within the boundary value of the physically reachable range. The boundary value is determined by the mean and standard deviation of the statistical historical time delay samples. Extract control layer variables, execution layer variables, and vehicle layer variables as candidate nodes from execution unit measurement data, control command data, and vehicle response data; By calculating the rate of change of the time series data of each candidate node, the moment that exceeds the change criterion is identified as the change moment; each candidate node is taken as the target node, and the change moments of other candidate nodes and the target node are compared within a preset time window. Candidate nodes whose change moments are earlier than the target nodes are marked as potential predecessor nodes, and candidate nodes whose change moments are later than the target nodes are marked as potential successor nodes. The transmission path type is determined based on the variable hierarchy of potential predecessor and successor nodes. When the transmission path is from the control layer to the execution layer, the boundary value is associated with the signal delay. When the transmission path is from the execution layer to the vehicle layer, the boundary value is associated with the force transmission delay. The boundary value is determined by the mean and standard deviation of the statistical historical delay samples. A causal edge is established when the change time of the predecessor node is earlier than the response time of the successor node, the time difference falls within the boundary value, and the association verification is passed. Construct a dynamic association graph using the confirmed nodes and causal edges.

[0010] The steps of calculating the rate of change of the time series data of each candidate node and identifying moments exceeding the change criterion as change moments, and then comparing the change moments of other candidate nodes with those of the target nodes within a preset time window, include: Collect time series data of each candidate node under normal historical conditions, calculate the statistical distribution characteristics of the historical change rate samples of each candidate node, and set adaptive change criteria for each candidate node based on the statistical distribution characteristics. The sliding window rate of change is calculated for the current time series data of each candidate node. When the rate of change exceeds the adaptive change criterion of the candidate node, it is recorded as a candidate change moment. The candidate change moment is continuously verified. When the duration of the change state exceeds the duration criterion, it is confirmed as a change moment. The size of the preset time window is determined based on the statistical characteristics of the signal delay and the force transmission delay. When the target node is a control layer variable, the upper bound of the time window is determined based on the signal delay. When the target node is an execution layer variable, the upper bound of the time window is determined by superimposing the signal delay and the force transmission delay. Within the upper bound of the time window, the change times of other candidate nodes are searched and compared with the change times of the target node.

[0011] When a deviation in the node state variable is detected in the dynamic correlation graph, the steps for obtaining the fault propagation chain and the degradation time reference value of each node by simulating fault propagation along the causal edge include: When the deviation between a node's state variable and the statistical baseline value of the node's historical normal operating state exceeds the deviation criterion, it is marked as an initial abnormal node. The deviation criterion is determined by the difference between the fluctuation range of the node under the historical normal state and the distribution under the abnormal state. Obtain successor nodes from the initial abnormal node along its outgoing edges, and calculate the propagation influence of the state deviation of the initial abnormal node on each successor node based on the causal influence strength of the outgoing edges. For each successor node, a mapping function is fitted between the disturbance input and the degradation response data in the historical fault samples, and the degradation time baseline value of the successor node is calculated based on the mapping function. When the propagation impact of a subsequent node exceeds the disturbance propagation criterion, it continues to propagate along its outgoing edge, and the calculation of the propagation impact and the degradation time base value are repeated until the propagation impact does not exceed the disturbance propagation criterion. All nodes whose transmitted influence exceeds the disturbance propagation criterion and their causal edges are organized into a fault propagation chain according to the diffusion order, and the degradation time reference value of each node is recorded.

[0012] The steps of extracting current operating condition parameters, calculating the predicted degradation rate of each node in the fault propagation chain under the current operating condition through multiple regression analysis, calculating the corrected degradation time by comparing it with the baseline degradation time value, and determining the earliest node time to reach the failure threshold as the fault prediction value include: Extract the current braking request intensity, ambient temperature, and road surface adhesion status as operating parameters; For each node in the fault propagation chain, based on the node's hierarchical position in the fault propagation chain, identify the set of operating parameters and the degradation state variables of the preceding node that affect the node's degradation rate. Establish a ternary mapping relationship between the set of operating parameters, the degradation state variables of the preceding node, and the degradation rate of the node through multivariate regression analysis. Collect degradation rate samples of the node under different historical degradation accumulations, establish the mapping relationship between historical degradation accumulation and degradation rate gain coefficient through regression analysis, and calculate the degradation rate gain coefficient based on the current historical degradation accumulation of the node. Substituting the current operating condition parameter set and the degradation state variables of the preceding node into the ternary mapping relationship yields the basic degradation rate. Multiplying the basic degradation rate by the degradation rate gain coefficient yields the predicted degradation rate of the node. Divide the baseline degradation time of the node by the predicted degradation rate to obtain the corrected degradation time; determine the node with the shortest corrected degradation time and its time as the predicted fault arrival time.

[0013] After a fault actually occurs, the steps for updating the boundary values ​​and multiple regression analysis parameters based on the actual fault deviation include: Collect a snapshot of the system state at the actual moment the fault occurs, and extract the actual fault propagation path and the actual degradation time of each node; The actual fault propagation path is compared with the predicted fault propagation chain. If the paths are inconsistent, the mean and standard deviation of the corresponding propagation path are recalculated to update the boundary values. If the paths are consistent but the deviation between the actual degradation time and the baseline degradation time exceeds the baseline deviation threshold, the mapping function between the disturbance input and degradation response of the node is refitted. If the paths are consistent and the baseline degradation time is accurate but the deviation between the predicted fault arrival time and the actual occurrence time exceeds the prediction deviation threshold, multiple regression analysis is re-executed based on the newly added fault samples to update the ternary mapping relationship and the gain coefficient mapping relationship. For parameters that need to be updated, the accuracy of the parameter in historical predictions is extracted as the confidence weight, and the historical value of the parameter is combined with the updated value calculated based on the new fault samples using a weighted average method. When the operating parameters at the time of the fault exceed the historical operating parameter distribution range, the fault sample is marked as an abnormal operating condition sample and will not participate in parameter updates.

[0014] A second aspect of this invention provides an AI-based fault prediction system for brake-by-wire systems, comprising: The data association unit is used to acquire the measurement data, control command data and vehicle response data of the actuator of the brake-by-wire system, extract the signal delay and force transmission delay between the data, identify the variable transmission path that conforms to the physical causal order and construct a dynamic association graph. Causal edges are established only when the change time of the previous node is earlier than the response time of the subsequent node and the time difference is within the boundary value of the physical reachable range. The boundary value is determined by the mean and standard deviation of the statistical historical delay samples. The fault propagation unit is used to simulate the fault propagation chain and the degradation time base value of each node by performing fault propagation along the causal edge when a deviation of the node state variable is detected in the dynamic correlation graph. The fault prediction unit is used to extract current operating condition parameters, calculate the predicted value of the deterioration rate of each node in the fault propagation chain under the current operating condition through multiple regression analysis, calculate the corrected deterioration time with the deterioration time reference value, and determine the earliest node time to reach the failure threshold as the fault prediction value. The parameter update unit is used to update the boundary values ​​and multiple regression analysis parameters according to the actual fault deviation after the fault actually occurs.

[0015] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0016] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0017] This method, by constructing a dynamic correlation graph and simulating fault propagation, can accurately identify potential fault nodes and their propagation paths in a brake-by-wire system. By dynamically correcting the degradation rate using real-time operating parameters, it can adapt to complex changes in vehicle operating conditions. Multiple regression analysis quantifies the specific impact of different operating conditions on the node degradation process, making the corrected degradation time closer to the actual decay process. This method can accurately locate the most vulnerable link in the propagation chain and use the predicted failure time of that node as an early warning point for system-level faults, achieving a leap from qualitative analysis to quantitative prediction.

[0018] The closed-loop feedback mechanism enables the construction criteria and rate prediction model of the dynamic correlation graph to continuously evolve with use, and the prediction accuracy gradually improves over time, effectively overcoming the prediction drift problem caused by differences in initial data or environment. This method elevates fault diagnosis from a passive response to an active early warning stage, providing crucial decision-making basis for preventative maintenance of brake-by-wire systems. By knowing potential fault points and failure times in advance, maintenance plans can be rationally arranged, avoiding the risk of vehicle loss of control due to sudden failures. This method enhances the safety and reliability of brake-by-wire systems, providing important assurance for the stable operation of intelligent driving vehicles. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating an AI-based fault prediction method for brake-by-wire systems. Figure 2 Flowchart for constructing dynamic association graphs. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0022] Figure 1This is a flowchart illustrating the AI-based fault prediction method for brake-by-wire systems according to an embodiment of the present invention. Figure 1 As shown, the AI-based fault prediction method for brake-by-wire systems includes: Acquire the measurement data, control command data, and vehicle response data of the actuator of the brake-by-wire system, extract the signal delay and force transmission delay between the data, identify the variable transmission path that conforms to the physical causal order, construct a dynamic correlation graph, and establish a causal edge only when the change time of the previous node is earlier than the response time of the subsequent node and the time difference is within the boundary value of the physical reachable range. The boundary value is determined by the mean and standard deviation of the statistical historical delay samples. When a deviation of the node state variable is detected in the dynamic correlation graph, a fault propagation simulation is performed along the causal edge to obtain the fault propagation chain and the baseline value of the deterioration time of each node. Extract current operating condition parameters, calculate the predicted deterioration rate of each node in the fault propagation chain under the current operating condition through multiple regression analysis, calculate the corrected deterioration time with the baseline value of the deterioration time, and determine the node time that reaches the failure threshold earliest as the fault prediction value. After the fault actually occurs, the boundary values ​​and multiple regression analysis parameters are updated based on the actual fault deviation.

[0023] In one optional implementation, the step of extracting the signal delay and force transmission delay between data includes: The moment when the actuator response signal in the measurement data of the execution unit begins to rise is marked as the execution response moment, and the moment when the amplitude of the control signal in the control command data exceeds the activation threshold is marked as the command issuance moment. The time difference between the execution response moment and the command issuance moment is calculated as the signal delay. The pressure rise rate is identified by performing differential operation on the pressure signal in the measurement data of the execution unit and the moment when it exceeds the establishment criterion is identified as the pressure establishment moment. The deceleration signal in the vehicle response data is filtered and the deceleration is identified by the moment when it exceeds the response criterion as the deceleration response moment. The time difference between the pressure establishment moment and the deceleration response moment is calculated as the force transmission delay.

[0024] For example, the measurement data acquired by the execution unit originates from multiple sensor nodes distributed throughout the brake-by-wire system. These sensors include, but are not limited to, displacement sensors within the brake calipers, pressure sensors in the hydraulic lines, motor current sensors, and temperature sensors. During data acquisition, the sensor sampling frequency is typically set between 100Hz and 1000Hz to ensure the capture of rapid dynamic changes during braking execution. Control command data originates from PWM signals or voltage command signals output by the brake control unit. The timestamps of these command signals are recorded using a high-precision clock source, achieving a time resolution at the microsecond level. Vehicle response data primarily includes wheel speed signals, vehicle acceleration signals, and multi-axis deceleration signals acquired through the inertial measurement unit. These signals are transmitted via the vehicle network to the data acquisition module and synchronized with the control command data.

[0025] In the specific implementation of signal delay extraction, the measurement data of the execution unit first needs to be preprocessed to eliminate sensor noise and high-frequency interference introduced by sampling. For actuator response signals, typical actuators include motor-driven brake calipers or hydraulic pumps, whose response signals can be motor position feedback signals or hydraulic pressure signals. The determination of the execution response time adopts a sliding window algorithm, continuously monitoring the signal within a time window length of 10 milliseconds to 50 milliseconds. When the signal amplitude starts to rise from the reference value and the rate of change of three consecutive sampling points is positive, the time corresponding to the first sampling point is marked as the execution response time. To improve the accuracy of time marking, linear interpolation is performed on the signal points in the rising edge region. The execution response time is accurately located by the intersection of the interpolated continuous signal with the reference value plus a noise threshold. This noise threshold is set according to the standard deviation of signal fluctuation under static conditions, and is usually taken as 3 times the standard deviation as the judgment value.

[0026] Processing control command data requires distinguishing between different control signal types. For digital command signals, the activation threshold is set at 50% of the logic high level. For analog voltage command signals, the activation threshold is determined based on the actuator's startup characteristics, typically set within the range of 5% to 10% of the rated operating voltage. The timing of command issuance is also marked using a rising edge detection algorithm, but since control commands usually have low noise levels, a single-point criterion can be used for marking. In practical applications, control command signals may exhibit multi-step changes; in such cases, it is necessary to identify the moment when the activation threshold is first exceeded as the valid command issuance moment. Signal delay is calculated by subtracting the command issuance time from the timestamp of the execution response time. This delay reflects the entire signal transmission delay from the controller outputting the command to the actuator producing a measurable response, including the combined effects of signal transmission delay, actuator mechanical response delay, and sensor measurement delay.

[0027] Extracting the force transmission delay involves analyzing the physical process of braking force from the actuator to the vehicle's deceleration effect. The pressure signal is acquired from a pressure sensor installed at the brake caliper inlet or master cylinder output. The raw pressure signal includes pipeline pressure pulsations and periodic fluctuations introduced by the hydraulic pump operation. To accurately identify the effective braking pressure build-up process, the pressure signal is first bandpass filtered. The lower cutoff frequency of the filter is set to 0.5Hz to remove zero-point drift, and the upper cutoff frequency is set to 50Hz to retain the main dynamic characteristics of the pressure build-up process while suppressing high-frequency noise. The filtered pressure signal is then numerically differentiated using a central difference scheme to reduce numerical errors, calculating the pressure change rate between adjacent sampling points. The pressure rise rate is evaluated within a window consisting of five consecutive sampling points. The average pressure change rate within the window is calculated, and when this average exceeds the build-up criterion, effective pressure build-up is confirmed.

[0028] The establishment criteria need to consider the differences in different braking conditions. Under light braking conditions, the pressure build-up rate is typically between 1 MPa per second and 5 MPa per second, while under emergency braking conditions, it can reach over 10 MPa per second. To adapt to the needs of multiple operating conditions, the establishment criteria adopt an adaptive threshold strategy, dynamically adjusting the criterion value according to the magnitude of the control command amplitude. In specific implementation, the control command amplitude is divided into several levels, each corresponding to a preset establishment criterion benchmark value. The actual criterion value is corrected based on the benchmark value according to factors such as the current temperature and the cumulative operating time of the system. The pressure build-up moment is marked as the timestamp when the pressure rise rate first exceeds the establishment criterion. To improve positioning accuracy, three sampling points are taken before and after exceeding the criterion, and the pressure rise rate at these points is subjected to quadratic polynomial fitting. The intersection of the fitted curve and the establishment criterion is used to determine a more accurate pressure build-up moment.

[0029] The deceleration signal in the vehicle response data originates from a longitudinal acceleration sensor, typically installed near the vehicle's center of gravity to capture representative vehicle motion. The raw deceleration signal is affected by road excitation, powertrain vibration, and sensor noise, necessitating effective filtering. A low-pass filter is chosen, with a cutoff frequency set between 10Hz and 20Hz to preserve the dynamic characteristics of braking deceleration while effectively suppressing high-frequency vibration noise. A Butterworth filter is used to achieve a flat passband response; the filter order is determined based on the signal-to-noise ratio requirements, typically a second- to fourth-order filter. Filtering can be performed offline in batch processing to perform forward and backward filtering on historical data to eliminate phase delay, or online in real-time filtering can be used to meet real-time fault prediction requirements. The fixed phase delay introduced by real-time filtering needs to be compensated for in subsequent delay calculations.

[0030] Identifying the deceleration response moment requires distinguishing between braking deceleration and deceleration changes caused by other factors. Before braking initiation, the vehicle may be in an acceleration, constant speed, or coasting state. The deceleration baseline value is determined based on the average value of the deceleration signal within a time window before braking triggering. The response criterion is set as an incremental threshold superimposed on the baseline value. This incremental threshold is typically between 0.05g and 0.2g, with the specific value determined based on the vehicle type and braking system characteristics. The deceleration signal is continuously monitored after filtering. When the deceleration values ​​at more than 5 consecutive sampling points exceed the response criterion, a valid braking deceleration response is confirmed, and the moment the criterion is first exceeded is marked as the deceleration response moment. To avoid misjudgment caused by transient noise, after marking the deceleration response moment, it is also necessary to verify that the deceleration remains above the criterion for a subsequent period. The verification time window length is typically set to 50 to 100 milliseconds.

[0031] Force transmission delay is obtained by calculating the time difference between the deceleration response moment and the pressure build-up moment. Physically, this delay represents the time required for the braking torque generated by the braking actuator, through tire-road friction and vehicle dynamics response, to ultimately manifest as measurable vehicle deceleration. Force transmission delay is influenced by various factors, including the coefficient of friction between the tire and road surface, vehicle load distribution, suspension system stiffness, and the distribution ratio of braking force among the wheels. Under standard operating conditions on dry asphalt roads with vehicle loads close to rated values, the typical force transmission delay ranges from 20 to 80 milliseconds. Significant differences exist in force transmission delay for different road conditions and load states; on low-adhesion roads, due to the longer tire slip ratio change process, the force transmission delay may extend to over 150 milliseconds.

[0032] In practical calculations, multiple sets of signal delay and force transmission delay data can be obtained from a single braking event. This is because control commands may be adjusted multiple times during braking, resulting in multiple actuator responses and changes in vehicle deceleration. For such multi-stage braking conditions, delay extraction needs to be performed independently for each control adjustment event, establishing a correspondence between the control adjustment sequence and the delay sequence. When control commands undergo continuous small adjustments, to avoid errors introduced by frequent delay calculations, a threshold for the magnitude of control command changes is set. A new round of delay extraction is triggered only when the change in command exceeds 3% of the rated value. The extracted delay data is stored as raw samples in a historical database, providing data support for subsequent dynamic correlation graph construction and boundary value statistics. This delay data is also used to evaluate the response performance and health status of the brake-by-wire system.

[0033] In one alternative implementation, the steps of constructing a dynamic association graph include: Extract control layer variables, execution layer variables, and vehicle layer variables as candidate nodes from execution unit measurement data, control command data, and vehicle response data; By calculating the rate of change of the time series data of each candidate node, the moment that exceeds the change criterion is identified as the moment of change. Each candidate node is treated as a target node. Within a preset time window, the change times of other candidate nodes and the target node are compared. Candidate nodes whose change times are earlier than the target node are marked as potential predecessor nodes, and candidate nodes whose change times are later than the target node are marked as potential successor nodes. The transmission path type is determined based on the variable hierarchy of potential predecessor and successor nodes. When the transmission path is from the control layer to the execution layer, the boundary value is associated with the signal delay. When the transmission path is from the execution layer to the vehicle layer, the boundary value is associated with the force transmission delay. The boundary value is determined by the mean and standard deviation of the statistical historical delay samples. A causal edge is established when the change time of the current sequence node is earlier than the response time of the subsequent sequence node, the time difference falls within the boundary value, and the association verification is passed. Construct a dynamic association graph using the confirmed nodes and causal edges.

[0034] Combination Figure 2 The flowchart illustrating the construction of the dynamic correlation graph is provided. During the operation of the online braking system, the process from data acquisition to the construction of the dynamic correlation graph involves a systematic process of node identification and causal edge establishment. Data acquisition covers three levels: the control layer includes variables such as brake pressure commands output by the electronic control unit, ABS intervention signals, and brake force distribution ratios; the execution layer includes measurement data such as hydraulic regulating valve opening, motor speed, brake caliper displacement, and wheel cylinder pressure; and the vehicle layer includes response parameters such as wheel speed, vehicle deceleration, yaw rate, and tire longitudinal slip ratio. These variables, according to physical causal relationships, constitute a complete transmission chain from control decision-making to execution actions and then to vehicle response.

[0035] After aligning the collected multi-level data according to timestamps, the candidate node extraction stage begins. For control-level variables, the target pressure command output by the brake master controller is extracted as the control node. This command is typically transmitted in CA7 message format with a sampling frequency of 100Hz. For execution-level variables, the actual opening feedback of the hydraulic regulating valve, the hydraulic pump motor current, and the readings of the front and rear wheel cylinder pressure sensors are selected as execution nodes. These variables directly reflect the action state of the hydraulic actuators. For vehicle-level variables, the wheel speed sensor data of the four wheels, the output of the vehicle's longitudinal accelerometer, and the cumulative braking distance are extracted as response nodes. Each candidate node retains a historical data window of the most recent 200 seconds, forming a time-series sample for subsequent analysis.

[0036] An adaptive threshold strategy is used to identify the moment of change. For the time series data of each candidate node, the first-order difference between adjacent sampling points is calculated as the rate of change. Taking the opening of a hydraulic control valve as an example, a valid change is considered to have occurred when the difference in opening between two adjacent samples exceeds 5% and the change continues for at least 3 sampling periods. The change criteria are set differently according to the variable type: the change criterion for pressure variables is set to 0.5 MPa, the change criterion for velocity variables is set to 0.3 m / s, and the change criterion for acceleration variables is set to 0.2 m / s². The moment that meets the criteria is marked as the moment of change for that node, and the direction of change is recorded to distinguish between rising and falling edges. To avoid noise interference, median filtering preprocessing is introduced, and the window length is set to 5 sampling points.

[0037] A causal relationship search is conducted using the wheel cylinder pressure node as the target node. A preset time window is set to 500ms forward and 500ms backward, within which the change times of other candidate nodes are retrieved. The brake pressure command node, which increases 120ms before the target time, is marked as a potential predecessor node; the hydraulic regulating valve opening node, which begins to increase 80ms before the target time, is also included in the potential predecessor set; and the wheel speed node, which decreases only 150ms after the target time, is marked as a potential successor node. By traversing all candidate nodes, a potential predecessor set and a potential successor set are established for each target node. For cases with multiple potential predecessor nodes, the node with the time difference closest to the typical propagation delay is retained as the primary predecessor, and the other nodes are considered auxiliary causal sources.

[0038] The transmission path type is determined based on the node's hierarchical level. When a potential preceding node belongs to the control layer and a subsequent node belongs to the execution layer, it is identified as a control-execution transmission path. Its physical mechanism involves an electrical signal being processed by the ECU to drive a hydraulic valve. The boundary value of this type of path is related to signal delay, which includes CA7 bus transmission delay, ECU calculation delay, and drive circuit response delay. By statistically analyzing the delay samples from control commands to hydraulic valve responses in 1000 historical braking events, a mean of 65ms and a standard deviation of 12ms were obtained. Therefore, the boundary value is set as follows: That is, the time difference must fall between 41ms and 89ms for a causal relationship to be considered valid. Represents the mean. Representing the standard deviation, when the potential preceding node belongs to the execution layer and the subsequent node belongs to the vehicle layer, it is identified as an execution-response transmission path. Its physical mechanism is that hydraulic pressure acts on the brake disc through the brake caliper, generating frictional torque, which in turn changes the wheel's dynamic state. The boundary values ​​of this type of path are related to the force transmission delay, which is affected by the compressibility of the brake fluid, the mechanical clearance of the caliper, and the tire deformation characteristics. Statistical analysis of historical samples shows that the mean delay from wheel cylinder pressure establishment to wheel deceleration response is 95ms, the standard deviation is 18ms, and the boundary value range is determined to be 59ms to 131ms.

[0039] Building upon boundary value determination, an association verification mechanism is added to eliminate spurious causal relationships. Association verification includes two dimensions: correlation testing and consistency testing. The correlation test calculates the Pearson correlation coefficient between the changes in preceding and subsequent nodes; a coefficient with an absolute value greater than 0.6 passes verification. Taking brake pressure command and wheel cylinder pressure as an example, data on the magnitude of changes in both were extracted from 20 consecutive change events, and the calculated correlation coefficient was 0.87, indicating a strong correlation. The consistency test requires that the direction of change in the preceding node is consistent with the direction of the physical response of the subsequent node. For example, an increase in brake pressure command should lead to an increase in wheel cylinder pressure. If the command increases but the pressure decreases, it is considered an anomaly, and no causal edge is established. For the transmission from the execution layer to the vehicle layer, an increase in pressure should lead to a decrease in wheel speed, consistent with the physical laws of braking.

[0040] The process of establishing causal edges records complete transmission characteristic parameters. Each causal edge carries the identifiers of the preceding and succeeding nodes, average transmission delay, transmission gain, and confidence score. Transmission gain is defined as the ratio of the change in the succeeding node to the change in the preceding node, reflecting transmission efficiency. Taking the transmission from hydraulic valve opening to cylinder pressure as an example, a 10% increase in valve opening corresponds to a 1.2 MPa increase in cylinder pressure, resulting in a transmission gain of 0.12 MPa / %. The confidence score comprehensively considers delay stability, correlation coefficient, and historical successful matching rate, calculated using a weighted summation method. A smaller delay standard deviation, a higher correlation coefficient, and a greater number of historical successful matches result in a higher confidence score. The score ranges from 0 to 1; causal edges with a score higher than 0.75 are marked as strong causal relationships and used for identifying critical fault propagation paths.

[0041] The dynamic correlation graph is stored using a directed graph data structure. Nodes are represented by arrays of structures, each containing a node ID, node name, hierarchy, current state value, historical state sequence, and anomaly flag. Causal edges are represented by adjacency lists, with each node maintaining an edge list pointing to its successor node. The edge structure includes a target node pointer, propagation delay, propagation gain, and confidence level. After the graph is constructed, topological sorting is performed to verify the existence of loops. A normal causal chain in a brake-by-wire system should exhibit unidirectional propagation from the control layer to the execution layer and then to the vehicle layer. If a loop is detected, it indicates the presence of a feedback loop or data anomaly. For operating conditions involving ABS control, the slip rate signal from the vehicle layer is allowed to be fed back to the control layer to form a closed-loop regulation; such feedback edges are separately identified as control feedback types.

[0042] The graph update mechanism supports incremental construction. When new braking event data is input, it is first matched with the existing graph for pattern matching. If the similarity between the changed time sequence and the historical pattern exceeds 85%, the existing causal edges are reused and the moving average of the propagation delay is updated. If a new change pattern appears, a causal relationship re-search is triggered, and the newly identified nodes and causal edges are added to the graph. To prevent the graph from expanding indefinitely, the number of nodes is capped at 50. When the addition of new nodes exceeds this limit, the node with the lowest confidence score and that has not been activated in the last 500 braking events is deleted. The graph maintenance thread performs a global optimization every 100 braking events, recalculating the confidence of all causal edges and cleaning up low-confidence edges to ensure that the graph always reflects the true causal relationships of the current system.

[0043] The dynamic correlation graph exhibits differentiated characteristics across different vehicle speed ranges and road surface adhesion conditions. At low speeds, the braking response delay is shorter, with the boundary value of force transmission delay narrowing to 70ms to 110ms. At high speeds, due to hydraulic system pressure fluctuations and changes in tire dynamic characteristics, the boundary value expands to 80ms to 150ms. In low-adhesion conditions on icy and snowy roads, frequent ABS intervention significantly increases the activation frequency of the causal edge from the control layer to the execution layer. In this case, adding a dedicated causal edge from the ABS pressure regulating valve to the wheel cylinder pressure reduces the average transmission delay to approximately 30ms. Through condition-adaptive boundary value adjustment and causal edge activation strategies, the dynamic correlation graph can accurately characterize the system's inherent causal logic under different operating conditions, providing a reliable topological foundation for subsequent fault propagation simulations.

[0044] In one optional implementation, the step of calculating the rate of change of the time series data of each candidate node and identifying the moment exceeding the change criterion as the change moment; and comparing the change moments of other candidate nodes and target nodes within a preset time window by treating each candidate node as a target node includes: Collect time series data of each candidate node under normal historical conditions, calculate the statistical distribution characteristics of the historical change rate samples of each candidate node, and set adaptive change criteria for each candidate node based on the statistical distribution characteristics. The sliding window rate of change is calculated for the current time series data of each candidate node. When the rate of change exceeds the adaptive change criterion of the candidate node, it is recorded as a candidate change moment. The candidate change moment is continuously verified. When the duration of the change state exceeds the duration criterion, it is confirmed as a change moment. The size of the preset time window is determined based on the statistical characteristics of the signal delay and the force transmission delay. When the target node is a control layer variable, the upper bound of the time window is determined based on the signal delay. When the target node is an execution layer variable, the upper bound of the time window is determined by superimposing the signal delay and the force transmission delay. Within the upper bound of the time window, the change times of other candidate nodes are searched and compared with the change times of the target node.

[0045] For example, to accurately identify the causal relationships between variable nodes in a brake-by-wire system, a systematic change detection and time-series comparison method is needed to accurately capture the state changes of each node and determine their causal order. This implementation method uses adaptive change criteria setting, a continuous verification mechanism, and a time window determination strategy related to operating conditions to extract the key time-series information required for constructing a dynamic correlation graph.

[0046] When collecting time-series data under historical normal operating conditions, for each candidate node, including the brake pressure sensor, motor position sensor, pedal displacement sensor, wheel speed sensor, and braking torque sensor, at least 100 complete braking cycles of data samples are recorded at a sampling frequency of 200Hz. For each candidate node, its time-series data is extracted and the rate of change is calculated. The instantaneous rate of change is calculated using the forward differencing method. The rate of change of node i at time t is then calculated. The change in value at the node between time t and time t-Δt is calculated by dividing the time interval Δt, where Δt is taken as 5 ms. Statistical analysis is performed on the rate of change data from all historical samples, and the mean of the absolute values ​​of the rates of change is calculated. and standard deviation Based on these statistical distribution characteristics, adaptive change criteria are set for each candidate node. The calculation method is as follows Where k is the adjustment coefficient, which is set differently according to the node type. For control command type nodes, k=2.5 is set to improve sensitivity, and for measurement type nodes, k=3.0 is set to reduce noise interference.

[0047] During real-time operation, a sliding window rate of change is calculated for the current time series data of each candidate node. The window length is set to 10 ms, and the sliding step size is 5 ms. Within each window, the rate of change for that node is calculated, and the absolute value of the calculated rate of change is compared with the adaptive change criterion for that node. When a candidate node reaches time... When the absolute value of the rate of change first exceeds its adaptive change criterion, that moment is marked as a candidate change moment. However, due to potential false triggering caused by sensor noise, electromagnetic interference, or transient fluctuations, continuous verification of candidate change moments is required. Starting from the candidate change moment, the state changes of the node are continuously monitored, and the state difference of the node within the time period before and after the candidate change moment is calculated. For braking pressure nodes, the state difference is defined as the absolute value of the difference between the average value within 20 ms after the candidate change moment and the average value within 20 ms before the candidate change moment. A candidate change moment is only confirmed as a valid change moment when this state difference exceeds twice the normal fluctuation amplitude of the node, and the duration of this difference exceeds the continuity criterion. The continuity criterion is set according to the node type: 15 ms for fast-response control command nodes and 30 ms for physical measurement nodes with inertial characteristics.

[0048] Determining the size of the preset time window requires considering the statistical characteristics of signal delay and force transmission delay. By analyzing a large number of samples from historical data showing the timing of control command issuance and actuator response, the statistical distribution of signal delay is extracted, and the median of the signal delay is calculated. , 90 percentile and standard deviation Similarly, by analyzing the time difference between the moment the actuator actuates and the moment the vehicle's braking force is generated, the median force transmission delay can be obtained. , 90 percentile and standard deviation In practical systems, signal delay is typically between 8ms and 25ms, while force transmission delay varies between 15ms and 60ms depending on the type of brake.

[0049] When the target node is a control layer variable, such as the PWM duty cycle command or motor drive current command output by the brake controller, its preceding nodes are mainly the pedal displacement signal from the driver input layer or the brake request signal from the upper-level controller. In this case, the physical delay between the two nodes mainly consists of signal transmission and processing delays; therefore, the upper bound of the time window is determined based on the signal delay. The specific calculation method is as follows: This setting can cover more than 95% of normal signal transmission scenarios. The lower bound of the time window is set to the minimum physical delay of signal processing, typically taken as... To avoid misinterpreting random, non-causally related synchronous changes as causal relationships, the time window is set to 12ms to 35ms for scenarios where the brake controller receives pedal signals and outputs motor commands.

[0050] When the target node is an execution-layer variable, such as the actual clamping force of a brake caliper, brake disc pressure, or wheel deceleration, its preceding node may be a control-layer command variable. In this case, the time from the issuance of the control command to the manifestation of the execution effect involves two stages: signal transmission delay and force transmission delay. The signal is first transmitted from the controller to the actuator drive circuit, then the actuator performs physical actions to generate braking force, which is transmitted to the wheel through the transmission mechanism to produce the braking effect. Therefore, it is necessary to superimpose the signal delay and the force transmission delay to determine the upper bound of the time window. The calculation method is as follows: The lower bound of the time window is also determined based on the median of the two time delays minus the standard deviation, i.e. In a typical electromechanical braking system, the time window from the issuance of a control command to the establishment of wheel braking force is set to 30ms to 95ms.

[0051] After determining the time window, a comparative analysis is performed between the change times of other candidate nodes and the change times of the target node. Each candidate node is analyzed as a target node, and for each target node at time... The detected change occurred within its corresponding time window. The search within the node examines whether other candidate nodes exhibit a change at a specific time. If candidate node j exhibits a change at a specific time... If a node falls within this time window, it is determined that candidate node j may be a preceding causal node of the target node, and a candidate causal edge from node j to the target node is established in the dynamic association graph. To avoid random associations, frequency statistics are performed on each pair of candidate causal edges. Only when the causal relationship occurs more than 80% of the time in historical data is it confirmed as a stable causal edge and retained in the dynamic association graph.

[0052] To address the variations in time delay characteristics under different operating conditions, a condition correction factor is introduced during the time window determination process. The impact of operating parameters such as ambient temperature, brake pedal speed, and initial vehicle speed on signal and force transmission delays is analyzed from historical data to establish a condition correction model. When the ambient temperature is below zero degrees Celsius, the force transmission delay of the hydraulic system increases by 15% to 25%, and the upper bound of the time window is correspondingly expanded by this proportion. During emergency braking conditions where the brake pedal speed exceeds 200 mm / s, the control system enters a rapid response mode, reducing signal delay by approximately 20%, and the time window range is adjusted accordingly. This condition-adaptive time window setting strategy ensures accurate identification of causal relationships under various operating conditions, improves the robustness and accuracy of dynamic correlation graph construction, and provides a reliable topological foundation for subsequent fault propagation simulation and fault prediction.

[0053] In one optional implementation, when a deviation in the node state variable is detected in the dynamic correlation graph, the step of performing fault propagation simulation along the causal edge to obtain the fault propagation chain and the degradation time reference value of each node includes: When the deviation between a node's state variable and the statistical baseline value of the node's historical normal operating state exceeds the deviation criterion, it is marked as an initial abnormal node. The deviation criterion is determined by the difference between the fluctuation range of the node under the historical normal state and the distribution under the abnormal state. Obtain successor nodes from the initial abnormal node along its outgoing edges, and calculate the propagation influence of the state deviation of the initial abnormal node on each successor node based on the causal influence strength of the outgoing edges. For each successor node, a mapping function is fitted between the disturbance input and the degradation response data in the historical fault samples, and the degradation time baseline value of the successor node is calculated based on the mapping function. When the propagation impact of a subsequent node exceeds the disturbance propagation criterion, it continues to propagate along its outgoing edge, and the calculation of the propagation impact and the degradation time base value are repeated until the propagation impact does not exceed the disturbance propagation criterion. All nodes whose transmitted influence exceeds the disturbance propagation criterion and their causal edges are organized into a fault propagation chain according to the diffusion order, and the degradation time reference value of each node is recorded.

[0054] For example, during the operation of an online control braking system, it is necessary to monitor the state variables of each node in a dynamic correlation graph in real time, and identify anomalies by comparing them with historical normal operating states. Specifically, for each node in the graph, its mean and standard deviation are calculated in advance by statistically analyzing the historical data of that node during normal system operation, thus constructing a statistical benchmark value for that node. In actual operation, when the deviation between the real-time state variable of a node and its statistical benchmark value exceeds a preset deviation criterion, that node is marked as an initial abnormal node.

[0055] The deviation criterion is set by comprehensively considering two factors: first, the fluctuation range of the node under historical normal conditions, which is determined by analyzing the fluctuation amplitude of the node's state variables during normal operation to establish the upper and lower limits of normal fluctuation; second, the distribution difference of the node under abnormal conditions, which is determined by collecting the state distribution characteristics of the node when it is abnormal in historical failure cases and comparing them with the distribution under normal conditions. The deviation criterion is usually set as the upper limit of the normal fluctuation range plus a certain safety margin, which is determined based on the degree of separability between normal and abnormal distributions. When the real-time value of the node's state variable exceeds the range defined by the statistical benchmark plus or minus the deviation criterion, the anomaly detection mechanism is triggered.

[0056] Once the initial anomalous node is identified, the fault propagation simulation process is immediately initiated. Based on the causal edge relationships in the dynamic correlation graph, all subsequent nodes are obtained starting from the initial anomalous node and following the outgoing edges it points to. Outgoing edges represent causal influence paths, and each outgoing edge is associated with a causal influence strength coefficient. This coefficient quantifies the degree to which the state change of the preceding node affects the subsequent node. The causal influence strength coefficient is obtained by analyzing the correlation between changes in the preceding node and the response of the subsequent node in historical data. Specifically, partial correlation analysis or transfer entropy calculation methods can be used to extract the direct causal influence strength between two nodes after eliminating interference from other confounding factors.

[0057] The state deviation of the initial anomalous node needs to be quantified. The state deviation is defined as the difference between the real-time state variable value and the statistical baseline value; this difference can be positive or negative, representing values ​​above or below normal levels, respectively. Using the state deviation of the initial anomalous node as a disturbance input, the propagation impact on each successor node is calculated based on the causal influence strength coefficient of each outgoing edge. The propagation impact is calculated by multiplying the state deviation of the initial anomalous node by the causal influence strength coefficient of the corresponding outgoing edge. If the initial anomalous node has multiple outgoing edges, the propagation impact received by each successor node is calculated separately.

[0058] For each successor node, it is necessary to further evaluate the time progression of its degradation after receiving the transmitted impact. This evaluation relies on disturbance-response data accumulated from historical fault samples. Specifically, cases where the successor node has experienced similar disturbance inputs are retrieved from the historical fault database, and the amplitude of the disturbance input and the time interval from receiving the disturbance to significant degradation are extracted. The disturbance input amplitudes from multiple historical cases are used as independent variables, and the corresponding degradation time intervals are used as dependent variables. A mapping function is established through data fitting. For electrical nodes such as solenoid valve coils, the disturbance input is the overcurrent amplitude. (Unit A), Deterioration time (Unit: h) Using an exponential decay model Where a is the initial degradation time coefficient, b is the decay rate coefficient, and c is the residual life offset, these three parameters are determined by fitting historical samples using the least squares method. For mechanical wear nodes such as brake discs, the disturbance input is the braking torque. (Unit: N·m), using a power function model Where k is the wear coefficient and n is the wear exponent, typically n is between 0.5 and 1.5. For hydraulic seals, the disturbance input is the pressure fluctuation amplitude ΔP (in MPa), and a logarithmic function model is used. Where p is the baseline lifespan parameter and q is the pressure sensitivity coefficient. Model selection is based on the goodness of fit of historical data. Determine, when The model is used when the value is greater than 0.85.

[0059] After establishing the mapping function, the currently calculated propagated impact is substituted into the mapping function to calculate the baseline degradation time value of the successor node under the current disturbance. The baseline degradation time value represents the time required for the node to reach the failure threshold from receiving the disturbance under standard operating conditions. It should be noted that this is a baseline value; the actual degradation time will also be affected by the parameters of the current operating conditions and needs to be corrected in subsequent steps.

[0060] After calculating the propagation impact of the initial anomalous node's direct successor nodes and the degradation time baseline, it is necessary to determine whether the fault will continue to propagate to downstream nodes. The criterion is whether the propagation impact of each successor node exceeds a preset disturbance propagation criterion. Disturbance propagation criterion. The setting is based on the minimum detectable state change of the node, determined by analyzing the response relationship between node state changes and external disturbances in historical data. The specific calculation method is as follows: Calculate the standard deviation of the node under normal fluctuation conditions. The perturbation propagation criterion is set as follows: This ensures that downstream propagation is triggered only when the propagated impact exceeds three standard deviations of the normal fluctuation. For pressure sensor nodes, Approximately 0.02 MPa, then Set to 0.06MPa; for the motor speed node, At approximately 5 rpm, Set to 15 rpm. When the absolute value of the transmitted influence received by the successor node is greater than... If the node is significantly affected by the disturbance, it will continue to propagate downstream along its outgoing edge.

[0061] For a successor node whose propagated impact exceeds the disturbance propagation criterion, it is considered a new disturbance source, and its impact continues to propagate along its outgoing edges to more downstream nodes. At this point, the propagated impact received by the successor node is used as a new disturbance input. The next-level successor nodes pointed to by all its outgoing edges are then obtained, and the aforementioned propagated impact calculation and degradation time baseline value calculation process is repeated. When calculating the propagated impact of the next-level node, the propagated impact of the current node is also multiplied by the causal influence strength coefficient of the corresponding outgoing edge.

[0062] This diffusion simulation process recursively proceeds along the causal edges of the dynamic correlation graph. With each subsequent level of diffusion, the transmitted influence typically decreases gradually due to the causal influence strength coefficient being less than 1. When the transmitted influence received by all successor nodes at a certain level does not exceed the disturbance propagation criterion, it indicates that the impact of fault diffusion has decreased to a negligible level, and the diffusion simulation process terminates. Throughout the diffusion process, for each node whose transmitted influence exceeds the disturbance propagation criterion, the received transmitted influence value and the degradation time baseline value calculated using the mapping function are recorded.

[0063] After completing the fault propagation simulation, all nodes involved in the propagation path and their causal relationships need to be organized to form a structured fault propagation chain. The fault propagation chain arranges nodes according to the time sequence of propagation, with the starting node being the initial anomalous node, and subsequent nodes connected sequentially according to the order of fault impact transmission. Each node in the fault propagation chain is accompanied by two key data points: the amount of transmitted impact received by the node and the node's degradation time baseline value. The causal edges between nodes retain the causal influence strength information from the dynamic correlation graph, facilitating subsequent analysis of the strength of fault propagation.

[0064] When constructing a fault propagation chain, some nodes may receive propagated effects from multiple preceding nodes. For such nodes, it is necessary to comprehensively consider the propagated effects of all input paths and determine the total propagated effect of the node using either a summation or maximum value strategy. Based on this total propagated effect, the degradation time baseline value is calculated. The summation strategy is suitable for scenarios where the effects of multiple disturbance sources have a cumulative effect, while the maximum value strategy is suitable for scenarios where node degradation is mainly dominated by the strongest disturbance source. The specific strategy to choose depends on the node's physical characteristics and the statistical patterns of historical data.

[0065] The construction of the fault propagation chain provides key topology and time reference information for subsequent fault prediction. By identifying all nodes that may be affected by the fault and their degradation time reference values, it lays the foundation for further correction based on the current operating conditions, ensuring that fault prediction can cover the complete range of fault impact and avoid missing key affected nodes.

[0066] In one optional implementation, the steps of extracting current operating condition parameters, calculating the predicted degradation rate of each node in the fault propagation chain under the current operating condition through multiple regression analysis, calculating the corrected degradation time by comparing it with the degradation time baseline value, and determining the earliest node time to reach the failure threshold as the fault prediction value include: Extract the current braking request intensity, ambient temperature, and road surface adhesion status as operating parameters; For each node in the fault propagation chain, based on the node's hierarchical position in the fault propagation chain, identify the set of operating parameters and the degradation state variables of the preceding node that affect the node's degradation rate. Establish a ternary mapping relationship between the set of operating parameters, the degradation state variables of the preceding node, and the degradation rate of the node through multivariate regression analysis. Collect degradation rate samples of the node under different historical degradation accumulations, establish the mapping relationship between historical degradation accumulation and degradation rate gain coefficient through regression analysis, and calculate the degradation rate gain coefficient based on the current historical degradation accumulation of the node. Substituting the current operating condition parameter set and the degradation state variables of the preceding node into the ternary mapping relationship yields the basic degradation rate. Multiplying the basic degradation rate by the degradation rate gain coefficient yields the predicted degradation rate of the node. Divide the baseline degradation time of the node by the predicted degradation rate to obtain the corrected degradation time; determine the node with the shortest corrected degradation time and its time as the predicted fault arrival time.

[0067] For example, after obtaining the fault propagation chain and the baseline values ​​of the degradation time at each node, the degradation rate needs to be dynamically corrected based on actual operating conditions. The degradation rate of a brake-by-wire system varies significantly under different operating conditions. For instance, frequent solenoid valve actuation under high-frequency braking conditions accelerates coil insulation aging; increased brake fluid viscosity at low temperatures exacerbates hydraulic build-up response lag; and frequent triggering of anti-lock braking systems on slippery surfaces increases the workload of the pressure regulation module. Therefore, it is necessary to establish a quantitative mapping relationship between operating parameters and degradation rates.

[0068] The extraction of operating parameters needs to cover key factors that directly affect the degradation process of the braking system. Braking request intensity reflects the magnitude of the deceleration command issued by the driver's pedal input or the automatic driving system. This parameter is obtained through pedal sensors or controller area network messages, and its value typically ranges from 0 to 100%. Braking request intensity directly determines the working pressure level of the hydraulic actuator and the solenoid valve opening duration. Under high-intensity braking conditions, the mechanical stress and thermal load on the actuator increase significantly. Ambient temperature is collected in real time by onboard temperature sensors. The performance of components such as the electronic control unit, motor drive module, and hydraulic pump assembly in the brake-by-wire system is affected by temperature. Excessive temperature leads to increased leakage current and decreased insulation performance of electronic components, while excessively low temperature reduces brake fluid flow and hardens rubber seals. Road surface adhesion is estimated using wheel speed sensor signals combined with a vehicle dynamics model. The commonly used road surface adhesion coefficient is... Characterization of dry asphalt pavement The surface adhesion is approximately 0.8 to 0.9, dropping to 0.3 to 0.5 on wet and slippery surfaces, and potentially below 0.2 on icy and snowy surfaces. The road surface adhesion affects the intervention frequency of the anti-lock braking system (ABS). On low-adhesion surfaces, the pressure regulating valve needs to frequently perform pressure increase and decrease actions, accelerating valve body wear.

[0069] The hierarchical position in a fault propagation chain refers to the propagation distance of the node from the initial anomalous node. The initial anomalous node is defined as layer 0, the direct successor node of the initial anomalous node is layer 1, the successor node of the layer 1 node is layer 2, and so on. The degree to which nodes at different levels are affected by operating parameters varies. Layer 0 nodes are mainly affected by direct operating parameters, while nodes at layer 1 and above also need to consider the transmission effect of the degradation state of preceding nodes. For example, when the brake pressure sensor is the initial anomalous node at layer 0, its degradation rate is mainly affected by ambient temperature and pressure fluctuation amplitude; while the degradation rate of the pressure regulating valve node at layer 1 is affected not only by temperature but also by the number of abnormal valve body actions caused by pressure sensor output deviation.

[0070] For each node in the fault propagation chain, it is necessary to identify the set of operating condition parameters that directly affect the degradation rate of that node. Not all operating condition parameters have a significant impact on the degradation rate of each node. For example, the degradation rate of the brake fluid temperature sensor is mainly affected by ambient temperature and has a weak correlation with road surface adhesion; while the degradation rate of the pressure regulating valve is strongly correlated with both braking request intensity and road surface adhesion. Correlation analysis is performed using historical operating data. When the absolute value of the Pearson correlation coefficient between a certain operating condition parameter and the node's degradation rate is greater than 0.3, that parameter is included in the set of operating condition parameters for that node. In addition to operating condition parameters, the degradation state of preceding nodes in the fault propagation chain also affects the degradation rate of the current node. For example, pressure sensor signal deviation can cause overshoot or oscillation in the pressure closed-loop control, increasing the number of pressure regulating valve actions and thus accelerating valve body degradation. The degradation state variables of preceding nodes are quantified by the deviation of the node's output in the dynamic correlation graph from the normal baseline value.

[0071] A ternary mapping relationship between the set of operating condition parameters, the degradation state variables of the preceding node, and the degradation rate of the current node is established using a multiple linear regression method. Let the node degradation rate be... The working condition parameter set contains m parameters, denoted as m. The preceding node's deteriorated state variables contain n variables, denoted as... Then the ternary mapping relation is expressed as ,in For the intercept term, The regression coefficients of the operating condition parameters are... These are the regression coefficients for the preceding node degradation state variable. The regression coefficients are solved using the least squares method, constructing a training sample set using the operating parameters recorded in historical operating data, the degradation state of the preceding node, and the measured values ​​of the node degradation rate within the corresponding time period. To avoid overfitting, ridge regression is used to introduce a regularization term when the sample size is insufficient.

[0072] The degradation rate of a node is influenced not only by its current operating conditions but also by its historical cumulative degradation. As the cumulative degradation increases, microscopic damage within the component gradually accumulates, leading to an accelerated degradation rate. For example, motor carbon brushes exhibit a slow degradation rate in the early stages of wear; however, once the wear depth exceeds a certain critical value, the contact resistance increases sharply, and the degradation rate accelerates significantly. Historical cumulative degradation is defined as the amount of degradation accumulated by the node from its initial normal state to the current moment, and can be obtained by integrating the degradation rate over time. Degradation rate samples of the node at different historical cumulative degradation levels are collected, derived from records of the node's degradation evolution process in historical operating data. The historical cumulative degradation is used as the basis for further analysis. The degradation rate gain coefficient is the independent variable. Regression analysis was performed with the degradation rate as the dependent variable. The degradation rate gain coefficient was defined as the ratio of the current degradation rate to the initial degradation rate under standard operating conditions. The regression model adopted an exponential form. ,in The parameters to be fitted reflect the intensity of the degradation acceleration effect. The parameters are determined by fitting historical samples using a nonlinear least squares method. .

[0073] When calculating the predicted degradation rate of this node, the current operating condition parameter set and the degradation state variables of the preceding node are first substituted into the ternary mapping relationship to obtain the basic degradation rate. The base degradation rate reflects the node degradation speed under current operating conditions, without considering the cumulative effect of historical degradation. Subsequently, the current cumulative historical degradation of the node is used as the basis for further analysis. Substitute mapping relationship Calculate the degradation rate gain factor. Multiply the base degradation rate by the degradation rate gain factor. The predicted degradation rate of the node is obtained. This predicted value takes into account both the immediate impact of the current operating conditions and the cumulative effect of historical degradation, and can more accurately reflect the actual degradation rate of the node at the current moment.

[0074] After obtaining the predicted degradation rate, the corrected degradation time is calculated by combining it with the baseline degradation time value. The baseline degradation time value represents the time required for the node to degrade from its current state to the failure threshold under standard operating conditions and without historical degradation accumulation; this value has been determined during the failure propagation simulation phase. The corrected degradation time is obtained by dividing the baseline degradation time value by the predicted degradation rate value, i.e. ,in To correct the degradation time, This is the baseline value for degradation time. When the predicted degradation rate is greater than 1, it indicates that the current operating conditions or historical cumulative effects have accelerated the degradation process, and the corrected degradation time will be shorter than the baseline value; when the predicted degradation rate is less than 1, the corrected degradation time will be longer than the baseline value.

[0075] The fault propagation chain contains multiple nodes, each with its own corrective degradation time calculated according to the above process. The node with the shortest corrective degradation time and its time are determined as the predicted fault arrival time. This node represents the weakest link in the fault propagation chain that reaches the failure threshold earliest, and its failure time is the expected time for the entire system to exhibit fault characteristics. For example, in a fault propagation chain containing three nodes—a pressure sensor, a hydraulic pump, and a pressure regulating valve—the calculated corrective degradation times are 120 hours, 95 hours, and 150 hours, respectively. The hydraulic pump has the shortest corrective degradation time, and the fault is predicted to manifest 95 hours later through hydraulic pump failure. This prediction provides a time window for maintenance decisions, enabling maintenance personnel to complete preventative maintenance or component replacement before the actual failure occurs, avoiding safety risks and economic losses caused by sudden system failure.

[0076] In one optional implementation, the step of updating the boundary values ​​and multiple regression analysis parameters based on the actual fault deviation after the fault actually occurs includes: Collect a snapshot of the system state at the actual moment the fault occurs, and extract the actual fault propagation path and the actual degradation time of each node; The actual fault propagation path is compared with the predicted fault propagation chain. If the paths are inconsistent, the mean and standard deviation of the corresponding propagation path are recalculated to update the boundary values. If the paths are consistent but the deviation between the actual degradation time and the baseline degradation time exceeds the baseline deviation threshold, the mapping function between the disturbance input and degradation response of the node is refitted. If the paths are consistent and the baseline degradation time is accurate but the deviation between the predicted fault arrival time and the actual occurrence time exceeds the prediction deviation threshold, multiple regression analysis is re-executed based on the newly added fault samples to update the ternary mapping relationship and the gain coefficient mapping relationship. For parameters that need to be updated, the accuracy of the parameter in historical predictions is extracted as the confidence weight, and the historical value of the parameter is combined with the updated value calculated based on the new fault samples using a weighted average method. When the operating parameters at the time of the fault exceed the historical operating parameter distribution range, the fault sample is marked as an abnormal operating condition sample and will not participate in parameter updates.

[0077] For example, in the closed-loop optimization phase of an online control braking system, a systematic parameter update process is executed based on the data feedback after each actual fault occurs to continuously improve prediction accuracy. This update mechanism includes four core steps: fault data acquisition, differential diagnosis, parameter classification and updating, and abnormal operating condition screening.

[0078] When a fault actually occurs, a system state snapshot is immediately triggered. This snapshot records the complete state vector of each actuator in the brake-by-wire system at the current moment, including the pressure measurement value of the brake pressure sensor, motor current and position encoder output, hydraulic cylinder piston displacement, brake caliper clamping force feedback, wheel speed sensor data, and vehicle accelerometer output. By performing timestamp alignment and causal relationship analysis on the multi-time sequence data in the snapshot, the actual fault propagation path is extracted. This path is represented as a node sequence, recording the propagation trajectory from the initial fault source node to the final failure node. Each node is accompanied by the timestamp of the node's initial degradation, the measured degradation rate, and the time when the failure threshold is reached. The actual degradation time of each node is calculated, that is, the time elapsed from the node's initial performance deviation to reaching the preset failure threshold. For example, when brake pressure sensor drift triggers a fault chain, the entire time span from the sensor output deviation exceeding ±2% to the braking force response delay exceeding 80ms is recorded, while the time information of intermediate nodes such as the hydraulic regulating valve response node and pressure establishment node is extracted.

[0079] The actual fault propagation paths collected are compared in detail with the fault propagation chains predicted by the system. First, a path consistency check is performed to compare whether the node sequences of the two paths match completely. If the actual path contains nodes not included in the predicted chain, or if the node order changes, the paths are considered inconsistent. In this case, for newly appearing causal edges in the actual propagation path, the corresponding time delay samples are re-counted. The time difference between the change time of the preceding node and the response time of the succeeding node of this causal edge is extracted. The time delay data of this edge in the most recent 50 operations are included in the sample set, and a new mean and standard deviation of the time delay are calculated. The boundary values ​​are updated to the endpoint values ​​of the interval equal to the mean plus or minus three times the standard deviation. These boundary values ​​are used to subsequently determine whether a causal edge is valid. For example, if a fault reveals that the brake pedal signal directly affects the rear wheel braking force but the predicted chain does not include this direct path, the time delay from the change in pedal angle to the rear wheel braking force response is counted. If the mean is 15ms and the standard deviation is 3ms, then the boundary values ​​are set to [6ms, 24ms].

[0080] After the path consistency check passes, a degradation time accuracy assessment is performed. The deviation between the actual degradation time and the predicted degradation time baseline value is calculated. If the absolute value of the deviation exceeds the baseline deviation threshold, the mapping function refit is triggered. The baseline deviation threshold is dynamically set according to the node type. For time-sensitive nodes such as electronic control units, the response threshold is set to 10ms, and for mechanically inertial nodes such as hydraulic pump start-up threshold, it is set to 50ms. When refitting is triggered, the disturbance input sequence and degradation response sequence of the node are extracted. The disturbance input includes the amplitude, frequency characteristics, and duration of abnormal signals transmitted by preceding nodes to the node. The degradation response characterizes the decay curve of the node's performance parameters over time. The least squares method is used to refit the mapping function between the disturbance input and the degradation time. The form of this function is selected according to the node's physical characteristics. An exponential decay model is used for electrical nodes, and a power function model is used for mechanical wear nodes. For example, for a brake disc wear node, if it is predicted to fail after 50 hours under the current disturbance but actually fails after 35 hours, the recent wear measurement data of the node and disturbance parameters such as braking frequency and braking torque are extracted, and the relationship function between the wear rate and the disturbance input is refitted.

[0081] For cases where the path and degradation time baseline values ​​are accurate, but the predicted arrival time of the fault has a large deviation, the diagnosis is inaccurate operating condition influencing parameters. In this case, the deviation between the predicted arrival time of the fault and the actual occurrence time is calculated. If the absolute value exceeds the prediction deviation threshold, a multivariate regression analysis parameter update is initiated. The prediction deviation threshold is set according to safety margin requirements, set at 5% of the predicted time for critical faults in the braking system, and relaxed to 15% for non-critical components. Operating condition parameters at the time of fault occurrence are extracted, including ambient temperature, vehicle load, road gradient, braking frequency, and vehicle speed distribution statistics. New fault samples are added to the training set, and the multivariate regression analysis is re-executed. This analysis establishes the mapping relationship between the triplet of operating condition parameters and the degradation rate adjustment coefficient, and updates the gain coefficients of each operating condition parameter on the degradation rate at different nodes. Stepwise regression is used to screen for significant influencing factors, eliminating operating condition parameters with correlation coefficients below 0.3, and retaining factors that significantly affect the degradation rate to construct a simplified regression model. For example, when the ambient temperature drops from 25℃ to -15℃, the increased viscosity of the hydraulic oil leads to a longer pressure build-up time. By updating the gain coefficient of temperature on the deterioration rate of hydraulic nodes through regression analysis, the prediction of the fault arrival time under low temperature conditions can be corrected.

[0082] During the parameter update execution phase, a confidence-weighted mechanism is introduced to avoid drastic parameter fluctuations caused by a single anomalous sample. For the parameter to be updated, the accuracy of that parameter in all historical predictions is used as the confidence weight. Accuracy is defined as the proportion of actual failure moments falling within the prediction time window in the predictions involving that parameter, with the time window width set to ±10% of the predicted value. The confidence weight is mapped to the interval [0.3, 0.9] using the logistic function of the accuracy, ensuring that parameters with excellent historical performance retain more prior information during updates. The final updated parameter value is calculated using a weighted average method, as shown in the formula: ,in This indicates the historical values ​​of the parameter. This represents the updated value calculated based on the new fault sample, where w is the confidence weight. For example, if the historical prediction accuracy of a certain time delay boundary value is 85%, corresponding to a confidence weight of 0.82, when the new fault sample calculates that the boundary value should be adjusted from 20ms to 25ms, the actual updated value is 0.82×20+0.18×25=20.9ms, achieving a smooth transition.

[0083] To prevent extreme operating condition samples from contaminating model parameters, anomaly screening is performed. The operating condition parameter vector at the moment of failure is extracted and its distribution is compared with a historical operating condition parameter database. Mahalanobis distance is used to measure the deviation of the current operating condition parameters from the historical distribution center; the distance calculation formula is as follows: Where x is the current operating condition parameter vector, This is a vector of historical operating condition parameter mean values. This is the covariance matrix. When the Mahalanobis distance exceeds a threshold... When an abnormal operating condition is identified, the threshold is determined based on the 95th percentile of the chi-square distribution of historical samples. Abnormal operating condition samples are separately marked and archived in an anomaly database, and are not included in regular parameter updates. For example, if a vehicle malfunctions in an extreme high-temperature environment of 60°C, while the historical operating temperature range is [-20°C, 45°C], the Mahalanobis distance of this sample significantly exceeds the threshold, and it is marked as an abnormal operating condition sample to prevent its extreme characteristics from affecting the prediction model under normal operating conditions. After a certain number of abnormal operating condition samples are accumulated, a dedicated abnormal operating condition prediction branch can be trained to expand the system's adaptability to extreme conditions.

[0084] Through the aforementioned closed-loop update mechanism, actual fault feedback information is continuously absorbed, and causal boundary values, node degradation mapping functions, and operating condition influence parameters are dynamically optimized, so that the fault prediction accuracy steadily improves over time. At the same time, the robustness and reliability of parameter updates are ensured through confidence weighting and anomaly screening.

[0085] A second aspect of this invention provides an AI-based fault prediction system for brake-by-wire systems, comprising: The data association unit is used to acquire the measurement data, control command data and vehicle response data of the actuator of the brake-by-wire system, extract the signal delay and force transmission delay between the data, identify the variable transmission path that conforms to the physical causal order and construct a dynamic association graph. Causal edges are established only when the change time of the previous node is earlier than the response time of the subsequent node and the time difference is within the boundary value of the physical reachable range. The boundary value is determined by the mean and standard deviation of the statistical historical delay samples. The fault propagation unit is used to simulate the fault propagation chain and the degradation time base value of each node by performing fault propagation along the causal edge when a deviation of the node state variable is detected in the dynamic correlation graph. The fault prediction unit is used to extract current operating condition parameters, calculate the predicted value of the deterioration rate of each node in the fault propagation chain under the current operating condition through multiple regression analysis, calculate the corrected deterioration time with the deterioration time reference value, and determine the earliest node time to reach the failure threshold as the fault prediction value. The parameter update unit is used to update the boundary values ​​and multiple regression analysis parameters according to the actual fault deviation after the fault actually occurs.

[0086] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0087] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0088] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI-based fault prediction method for brake-by-wire systems, characterized in that, include: Acquire the measurement data, control command data, and vehicle response data of the actuator of the brake-by-wire system, extract the signal delay and force transmission delay between the data, identify the variable transmission path that conforms to the physical causal order, construct a dynamic correlation graph, and establish a causal edge only when the change time of the previous node is earlier than the response time of the subsequent node and the time difference is within the boundary value of the physical reachable range. The boundary value is determined by the mean and standard deviation of the statistical historical delay samples. When a deviation of the node state variable is detected in the dynamic correlation graph, a fault propagation simulation is performed along the causal edge to obtain the fault propagation chain and the baseline value of the deterioration time of each node. Extract current operating condition parameters, calculate the predicted deterioration rate of each node in the fault propagation chain under the current operating condition through multiple regression analysis, calculate the corrected deterioration time with the baseline value of the deterioration time, and determine the node time that reaches the failure threshold earliest as the fault prediction value. After the fault actually occurs, the boundary values ​​and multivariate regression analysis parameters are updated based on the actual fault deviation. The step of obtaining the fault propagation chain and the degradation time base value of each node by performing fault propagation simulation along the causal edge when a deviation of the node state variable is detected in the dynamic correlation graph includes: When the deviation between a node's state variable and the statistical baseline value of the node's historical normal operating state exceeds the deviation criterion, it is marked as an initial abnormal node. The deviation criterion is determined by the difference between the fluctuation range of the node under the historical normal state and the distribution under the abnormal state. Obtain successor nodes from the initial abnormal node along its outgoing edges, and calculate the propagation influence of the state deviation of the initial abnormal node on each successor node based on the causal influence strength of the outgoing edges. For each successor node, a mapping function is fitted between the disturbance input and the degradation response data in the historical fault samples, and the degradation time baseline value of the successor node is calculated based on the mapping function. When the propagation impact of a subsequent node exceeds the disturbance propagation criterion, it continues to propagate along its outgoing edge, and the calculation of the propagation impact and the degradation time base value are repeated until the propagation impact does not exceed the disturbance propagation criterion. All nodes whose transmitted influence exceeds the disturbance propagation criterion and their causal edges are organized into a fault propagation chain according to the diffusion order, and the degradation time reference value of each node is recorded.

2. The method according to claim 1, characterized in that, The steps for extracting signal delay and force transmission delay between data include: The moment when the actuator response signal in the measurement data of the execution unit begins to rise is marked as the execution response moment, and the moment when the amplitude of the control signal in the control command data exceeds the activation threshold is marked as the command issuance moment. The time difference between the execution response moment and the command issuance moment is calculated as the signal delay. The pressure rise rate is identified by performing differential operation on the pressure signal in the measurement data of the execution unit and the moment when it exceeds the establishment criterion is identified as the pressure establishment moment. The deceleration signal in the vehicle response data is filtered and the deceleration is identified by the moment when it exceeds the response criterion as the deceleration response moment. The time difference between the pressure establishment moment and the deceleration response moment is calculated as the force transmission delay.

3. The method according to claim 1, characterized in that, The steps for constructing a dynamic association graph include: Extract control layer variables, execution layer variables, and vehicle layer variables as candidate nodes from execution unit measurement data, control command data, and vehicle response data; By calculating the rate of change of the time series data of each candidate node, the moment that exceeds the change criterion is identified as the moment of change. Each candidate node is treated as a target node. Within a preset time window, the change times of other candidate nodes and the target node are compared. Candidate nodes whose change times are earlier than the target node are marked as potential predecessor nodes, and candidate nodes whose change times are later than the target node are marked as potential successor nodes. The transmission path type is determined based on the variable hierarchy of potential predecessor and successor nodes. When the transmission path is from the control layer to the execution layer, the boundary value is associated with the signal delay. When the transmission path is from the execution layer to the vehicle layer, the boundary value is associated with the force transmission delay. A causal edge is established when the change time of the current sequence node is earlier than the response time of the subsequent sequence node, the time difference falls within the boundary value, and the association verification is passed. Construct a dynamic association graph using the confirmed nodes and causal edges.

4. The method according to claim 3, characterized in that, By calculating the rate of change of the time series data of each candidate node, the moment that exceeds the change criterion is identified as the moment of change. The steps of treating each candidate node as the target node and comparing the changes of other candidate nodes with those of the target node within a preset time window include: Collect time series data of each candidate node under normal historical conditions, calculate the statistical distribution characteristics of the historical change rate samples of each candidate node, and set adaptive change criteria for each candidate node based on the statistical distribution characteristics. The sliding window rate of change is calculated for the current time series data of each candidate node. When the rate of change exceeds the adaptive change criterion of the candidate node, it is recorded as a candidate change moment. The candidate change moment is continuously verified. When the duration of the change state exceeds the duration criterion, it is confirmed as a change moment. The size of the preset time window is determined based on the statistical characteristics of the signal delay and the force transmission delay. When the target node is a control layer variable, the upper bound of the time window is determined based on the signal delay. When the target node is an execution layer variable, the upper bound of the time window is determined by superimposing the signal delay and the force transmission delay. Within the upper bound of the time window, the change times of other candidate nodes are searched and compared with the change times of the target node.

5. The method according to claim 1, characterized in that, The steps of extracting current operating condition parameters, calculating the predicted degradation rate of each node in the fault propagation chain under the current operating condition through multiple regression analysis, calculating the corrected degradation time by comparing it with the baseline degradation time value, and determining the earliest node time to reach the failure threshold as the fault prediction value include: For each node in the fault propagation chain, based on the node's hierarchical position in the fault propagation chain, identify the set of operating parameters and the degradation state variables of the preceding node that affect the node's degradation rate. Establish a ternary mapping relationship between the set of operating parameters, the degradation state variables of the preceding node, and the degradation rate of the node through multivariate regression analysis. Collect degradation rate samples of the node under different historical degradation accumulations, establish the mapping relationship between historical degradation accumulation and degradation rate gain coefficient through regression analysis, and calculate the degradation rate gain coefficient based on the current historical degradation accumulation of the node. Substituting the current operating condition parameter set and the degradation state variables of the preceding node into the ternary mapping relationship yields the basic degradation rate. Multiplying the basic degradation rate by the degradation rate gain coefficient yields the predicted degradation rate of the node. The modified degradation time is obtained by dividing the baseline degradation time of the node by the predicted degradation rate. The node with the shortest correction degradation time and its time are determined as the predicted fault arrival time.

6. The method according to claim 5, characterized in that, After a fault actually occurs, the steps for updating the boundary values ​​and multiple regression analysis parameters based on the actual fault deviation include: Collect a snapshot of the system state at the actual moment the fault occurs, and extract the actual fault propagation path and the actual degradation time of each node; The actual fault propagation path is compared with the predicted fault propagation chain. When the paths are inconsistent, the mean and standard deviation of the delay of the corresponding propagation path are recalculated to update the boundary value. When the paths are consistent but the deviation between the actual degradation time and the degradation time benchmark value exceeds the benchmark deviation threshold, the mapping function between the perturbation input and degradation response of the node is refitted. When the path is consistent and the degradation time baseline value is accurate, but the deviation between the predicted value of the fault arrival time and the actual occurrence time exceeds the prediction deviation threshold, the multivariate regression analysis is re-executed based on the newly added fault samples to update the ternary mapping relationship and the gain coefficient mapping relationship. For parameters that need to be updated, the accuracy of the parameter in historical predictions is extracted as the confidence weight, and the historical value of the parameter is combined with the updated value calculated based on the new fault samples using a weighted average method. When the operating parameters at the time of the fault exceed the historical operating parameter distribution range, the fault sample is marked as an abnormal operating condition sample and will not participate in parameter updates.

7. An AI-based fault prediction system for brake-by-wire systems, used to implement the method as described in any one of claims 1-6, characterized in that, include: The data association unit is used to acquire the measurement data, control command data and vehicle response data of the actuator of the brake-by-wire system, extract the signal delay and force transmission delay between the data, identify the variable transmission path that conforms to the physical causal order and construct a dynamic association graph. Causal edges are established only when the change time of the previous node is earlier than the response time of the subsequent node and the time difference is within the boundary value of the physical reachable range. The boundary value is determined by the mean and standard deviation of the statistical historical delay samples. The fault propagation unit is used to simulate the fault propagation chain and the degradation time base value of each node by performing fault propagation along the causal edge when a deviation of the node state variable is detected in the dynamic correlation graph. The fault prediction unit is used to extract current operating condition parameters, calculate the predicted value of the deterioration rate of each node in the fault propagation chain under the current operating condition through multiple regression analysis, calculate the corrected deterioration time with the deterioration time reference value, and determine the earliest node time to reach the failure threshold as the fault prediction value. The parameter update unit is used to update the boundary values ​​and multiple regression analysis parameters according to the actual fault deviation after the fault actually occurs.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images