Electronic vacuum pump control method and system for vehicle based on braking intention recognition

Abstract

This invention discloses a method and system for controlling an automotive electronic vacuum pump based on braking intention recognition, belonging to the field of intention recognition. The method includes: acquiring vehicle status information, driver operation information, and environmental perception information from an advanced driver assistance module; calculating environmental risk levels based on environmental perception information and vehicle status information, including a forward collision risk level based on collision time and a downhill risk level based on road gradient; extracting current pedal feature parameters based on operation information, and inputting them along with the environmental risk level and historical braking tendency quantification values ​​into a Bayesian network to calculate the posterior probability distribution of braking intention at the current moment through probabilistic inference; generating pre-scheduled control commands for the electronic vacuum pump based on the posterior probability distribution of braking intention and a preset probability threshold, and sending them to the electronic vacuum pump controller for execution. This application solves the technical problem of existing automotive electronic vacuum pumps having slow response and difficulty in handling complex operating conditions, resulting in insufficient braking safety.

Description

Technical Field

[0001] This invention relates to the field of intent recognition, and more specifically to a method and system for controlling an automotive electronic vacuum pump based on braking intent recognition. Background Technology

[0002] New energy vehicles have eliminated the vacuum source in the engine intake manifold of traditional fuel vehicles, and instead use electronic vacuum pumps to provide negative pressure for the brake vacuum booster system. Existing electronic vacuum pumps generally only start to replenish pressure after the vacuum level in the pipeline drops to a preset lower limit, which has obvious response lag. At the same time, the existing control schemes cannot predict the braking demand of different levels of urgency. Under high-risk conditions, there is a risk of insufficient vacuum reserve leading to insufficient braking safety. In addition, frequent passive start-stop will increase the energy consumption and mechanical wear of the electronic vacuum pump, shortening its service life. Summary of the Invention

[0003] This application provides a method and system for controlling an automotive electronic vacuum pump based on braking intention recognition, aiming to solve the technical problems of insufficient braking safety caused by the slow response and inability to cope with complex working conditions in existing automotive electronic vacuum pumps.

[0004] In view of the above problems, this application provides a method and system for controlling an automotive electronic vacuum pump based on braking intention recognition.

[0005] The first aspect disclosed in this application provides a method for controlling an automotive electronic vacuum pump based on braking intention recognition, the method comprising: The system acquires vehicle status information, driver operation information, and environmental perception information from the advanced driver assistance module; calculates environmental risk levels based on the environmental perception information and vehicle status information, including forward collision risk levels based on collision time and downhill risk levels based on road gradient; constructs a Bayesian network, extracts current pedal feature parameters based on the driver operation information, and inputs them into the Bayesian network along with the environmental risk level and the driver's historical braking tendency quantification value, and calculates the posterior probability distribution of braking intention at the current moment through probabilistic inference; generates pre-scheduled control commands for the electronic vacuum pump based on the posterior probability distribution of braking intention and a preset probability threshold, and sends them to the electronic vacuum pump controller for execution.

[0006] Another aspect of this application discloses a vehicle electronic vacuum pump control system based on braking intention recognition, the system comprising: The system comprises the following modules: a data acquisition module for acquiring vehicle status information, driver operation information, and environmental perception information from the advanced driver assistance module; a calculation module for calculating environmental risk levels based on the environmental perception information and vehicle status information, including a forward collision risk level based on collision time and a downhill risk level based on road gradient; an inference module for constructing a Bayesian network, extracting current pedal feature parameters from the driver operation information, inputting them along with the environmental risk level and the driver's historical braking tendency quantification value into the Bayesian network, and calculating the posterior probability distribution of braking intention at the current moment through probabilistic inference; and an execution module for generating pre-scheduled control commands for the electronic vacuum pump based on the posterior probability distribution of braking intention and a preset probability threshold, and sending them to the electronic vacuum pump controller for execution.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: By integrating vehicle status, driver operation, and ADAS environmental perception information, a multi-dimensional environmental risk quantification system encompassing forward collision risk and downhill risk is constructed. Combining a driver's historical braking tendency quantification method with a four-node Bayesian network probabilistic inference model, the system achieves advanced recognition of the driver's braking intention and generates pre-scheduled control commands for the electronic vacuum pump. This allows for vacuum reserve to be completed in advance before the braking action takes effect, effectively solving the response lag problem of passive control and significantly improving braking safety in emergency braking and long downhill conditions. At the same time, the hierarchical pre-scheduled control avoids ineffective start-stop of the vacuum pump, reducing its operating energy consumption and mechanical wear, and extending the service life of components.

[0008] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0009] Figure 1 A flowchart illustrating a vehicle electronic vacuum pump control method based on braking intent recognition is provided for embodiments of this application; Figure 2 A schematic diagram of the structure of an automotive electronic vacuum pump control system based on braking intention recognition is provided for embodiments of this application.

[0010] Figure labeling: Acquisition module 11, Calculation module 12, Inference module 13, Execution module 14. Detailed Implementation

[0011] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0012] The overall concept of the technical solution provided in this application is as follows: This application provides a method and system for controlling an automotive electronic vacuum pump based on braking intention recognition. By integrating environmental perception information from an advanced driver assistance module, vehicle status information, and driver operation information, a comprehensive environmental risk level is calculated, combining the risks of forward collisions and downhill slopes. A Bayesian network is constructed by combining the driver's historical braking tendency quantification value and real-time pedal feature parameters. Through probabilistic reasoning, the posterior probability distribution of braking intention is accurately predicted, and a hierarchical pre-scheduling control command is generated to pre-regulate the operation of the electronic vacuum pump. This effectively solves the response lag problem of traditional control. It can not only establish sufficient braking vacuum reserves in advance under high-risk conditions, improving the response speed of the braking system and driving safety, but also dynamically adjust the vacuum pump operating state according to actual braking needs, reducing ineffective start-stop operations and lowering system energy consumption and component wear.

[0013] After introducing the basic principles of this application, various non-limiting embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0014] Example 1, as Figure 1 As shown in the embodiment of this application, a method for controlling an automotive electronic vacuum pump based on braking intention recognition is provided. The method includes: S100: Acquire vehicle status information, driver operation information, and environmental perception information from advanced driver assistance modules.

[0015] Specifically, the vehicle braking system main control unit, through the vehicle real-time bus network, supporting high-precision sensing sensors, and the communication interface with the vehicle intelligent control unit, synchronously completes the high real-time and high-synchronization acquisition and preprocessing of three types of core information. Among them, the vehicle status information first collected by the main control unit refers to a set of parameters that can comprehensively reflect the vehicle's own driving conditions and the operating status of core braking-related components. Specifically, it covers basic parameters such as the vehicle's real-time driving speed, vehicle braking deceleration, braking system pipeline vacuum, vehicle longitudinal acceleration, vehicle battery working voltage, and the current operating status of the electronic vacuum pump. These parameters are mainly collected synchronously at a high frequency of milliseconds from the vehicle's core control units and sensing elements, such as the body controller, power system controller, wheel speed sensors, and vacuum sensors, via the vehicle CAN / LIN bus. The sampling frequency matches the real-time requirements of vehicle braking safety control, and all collected parameters are configured with a unified timestamp.

[0016] Synchronously, the master control unit collects driver operation information, which refers to a set of characteristic parameters that can directly and accurately reflect the driver's real-time operation behavior on the vehicle braking system. The core includes key data such as the real-time displacement of the brake pedal, the pedal force applied by the driver, the instantaneous speed of the pedal movement, the pedal stroke change rate, and the on / off state of the brake pedal switch. Such parameters are collected in real time by components such as high-precision displacement sensors, pressure sensors, and angle sensors equipped at the brake pedal assembly, and are transmitted to the master control unit in real time through hard wires or in-vehicle buses.

[0017] At the same time, the master control unit completes real-time data interaction with the advanced driving assistance module. The advanced driving assistance module is the ADAS intelligent control unit installed in the vehicle, which integrates multi-source environmental perception hardware such as millimeter-wave radars, front-view cameras, and optionally lidar. It has core functions such as front target recognition, road alignment detection, obstacle perception, and relative motion parameter calculation. It is the core carrier for this solution to obtain information on the external environment of the vehicle's driving. The master control unit obtains environmental perception information from this module in real time through the in-vehicle Ethernet or dedicated CAN bus interface. This information is a set of parameters that comprehensively reflect the state of the external road environment and traffic participants during the vehicle's driving after the advanced driving assistance module completes multi-source raw data fusion, target recognition, and parameter calculation. The core includes key data such as the relative distance, relative driving speed, relative acceleration between the vehicle and the vehicle / obstacle ahead, the slope parameter of the road ahead, the length of the downhill section, lane line information, and the position and motion state of surrounding traffic participants.

[0018] S200: Calculate the environmental risk level based on the environmental perception information and the vehicle state information, including the pre-collision risk level calculated based on the time to collision and the downhill risk level based on the road slope.

[0019] Specifically, calculating the environmental risk level refers to a core operation process of multi-dimensionally quantifying and grading the braking safety risks brought by the external traffic and road environment during the vehicle's driving based on the environmental perception information and vehicle state information collected in the previous steps, and outputting standardized risk indicators. The overall operation is divided into two parallel calculation branches: forward collision risk and braking risk on the downhill section. In the first calculation branch, the master control unit first extracts core parameters such as the relative distance and relative driving speed between the vehicle and the vehicle / obstacle ahead from the environmental perception information, and completes the calculation of the time to collision. The time to collision refers to the remaining time required for the two vehicles to collide on the premise that the relative motion trend between the vehicle and the front target remains unchanged, and its core calculation formula is: TTC = d rel / v rel ; where, d rel is the relative distance between the vehicle and the vehicle ahead, and v relLet v be the approach speed of this vehicle relative to the vehicle in front, and when v rel When the distance between the vehicle and the vehicle in front is continuously increasing (≤0), the collision time is considered infinite, corresponding to no collision risk. After calculating the collision time, the main control unit calls the preset mapping relationship between collision time and collision probability level obtained by analyzing historical collision data to complete the classification and output of the forward collision risk level. The forward collision risk level refers to a classification index based on the collision time, which represents the probability of a collision accident occurring in front of the vehicle. Specifically, it can be divided into three levels: high, medium, and low. For example, when the collision time is ≤1.5s, it is set as high risk; when 1.5s < collision time ≤3.0s, it is set as medium risk; and when the collision time >3.0s, it is set as low risk. This achieves standardized quantification of collision risk in scenarios such as following other vehicles and sudden obstacles. In the parallel second calculation branch, the main control unit simultaneously completes the quantification calculation of downhill braking risk. First, it extracts downhill feature parameters composed of road slope and corresponding downhill section length from environmental perception information. The road slope refers to the value obtained through the advanced driver assistance module. The longitudinal slope value of the current driving segment obtained by the multi-source sensing hardware is the core parameter characterizing the increase in braking load on downhill sections. Simultaneously, the real-time driving speed of the vehicle is extracted from the vehicle's status information. These parameters are then input into a pre-constructed downhill risk model to calculate the downhill risk level. The downhill risk level is a grading index quantified by road slope, downhill length, and vehicle speed, characterizing the increased risk of brake fade and brake failure on long or steep downhill sections. The downhill risk model used is a pre-trained random forest classification model. This model undergoes supervised training using a large amount of historical vehicle speed data, historical downhill feature parameters, and historical downhill risk level data assigned based on braking accident probability. It can accurately output a risk level matching the current downhill conditions. After completing the calculations of the two branches, the main control unit uses the output forward collision risk level and downhill risk level together to constitute the final environmental risk level result. This result will serve as core observational evidence input into the subsequent Bayesian network.

[0020] S300: Construct a Bayesian network, extract the current pedal feature parameters based on the driver's operation information, and input them into the Bayesian network along with the environmental risk level and the driver's historical braking tendency quantification value. Calculate the posterior probability distribution of the braking intention at the current moment through probabilistic inference.

[0021] Specifically, the first step is to construct and initialize a Bayesian network. A Bayesian network is a mathematical model based on a probabilistic graphical model that enables uncertain causal reasoning. It can integrate multi-source heterogeneous observational evidence to calculate the posterior probability of a target event, naturally adapting to the complex and uncertain braking intention recognition requirements in driving scenarios. The Bayesian network constructed in this solution pre-defines four types of core nodes: environmental risk level nodes, driver's historical braking tendency nodes, current pedal feature nodes, and braking intention posterior probability nodes. Simultaneously, the causal relationships between nodes are pre-defined: the environmental risk level node and the driver's historical braking tendency node act as parent nodes pointing to the current pedal feature node; the environmental risk level node, the driver's historical braking tendency node, and the current pedal feature node together act as parent nodes pointing to the braking intention posterior probability node. Furthermore, each node in the Bayesian network is pre-configured with a conditional probability table obtained by offline collection of real driving data and training using the expectation-maximization algorithm. This conditional probability table fully records the probability values ​​of any node taking each state under different combinations of parent nodes, providing a basic probability calculation basis for subsequent probabilistic reasoning.

[0022] After completing the construction and initialization of the Bayesian network, the main control unit extracts the current pedal feature parameters from the previously collected driver operation information. These current pedal feature parameters refer to a set of core quantitative features that can accurately characterize the driver's current brake pedal operation behavior in real time. Specifically, these include key parameters such as current brake pedal displacement, pedal applied force, instantaneous pedal speed, and pedal travel rate of change. Simultaneously, the main control unit calls the pre-calculated driver's historical braking tendency quantification value. This value is a standardized quantitative indicator used to characterize the driver's fixed braking operation habits and style formed over a long period of driving. It is obtained by pre-statistically analyzing the average maximum pedal speed and average braking deceleration of all braking events performed by the driver within a preset time period. These parameters are input into a pre-trained clustering model, which outputs a mild, conventional, or aggressive driver type label. This driver type label is then mapped to a continuous quantization value between 0 and 1. A higher value indicates a more aggressive braking style. This quantization value provides personalized historical behavioral prior information about the driver for braking intent recognition. Subsequently, the main control unit extracts... The three core data types—current pedal feature parameters, environmental risk level calculated in the preceding steps, and driver's historical braking tendency quantification—are synchronously input into a pre-constructed Bayesian network as observational evidence. Probabilistic inference is then performed, whereby the posterior probability of a target node is calculated based on the causal structure of the Bayesian network and a pre-configured conditional probability table, combined with the input observational evidence. This scheme specifically employs variable elimination to perform the inference. First, the three types of input observational evidence are mapped to the corresponding definite states of nodes in the Bayesian network. Then, based on the conditional probability tables of all nodes and the input observational evidence, non-evidence nodes are eliminated by probability summation. Finally, the core operation of calculating the posterior probability distribution of the braking intention at the current moment is completed. The posterior probability distribution of the braking intention at the current moment refers to the set of probability values ​​output by the Bayesian network inference, given the current environmental risk, the driver's historical braking habits, and real-time pedal operation—probability values ​​indicating the driver's current braking intention is in different preset states, such as light braking, normal braking, and emergency braking. This distribution represents the probability distribution of the type and intensity of the driver's braking intention at the current moment.

[0023] S400: Based on the posterior probability distribution of the braking intention and the preset probability threshold, generate a pre-scheduling control command for the electronic vacuum pump and send it to the electronic vacuum pump controller for execution.

[0024] Specifically, firstly, the vehicle braking system's main control unit analyzes the posterior probability distribution of the input braking intent, extracting real-time probability values ​​corresponding to different braking intent levels, including mild braking, normal braking, and emergency braking. These levels are then matched against preset probability thresholds. These preset probability thresholds are critical probability values ​​used for tiered triggering of braking intent probability results, calibrated using extensive real-vehicle driving data and set in conjunction with vehicle braking system safety design requirements and electronic vacuum pump operating characteristics. Differentiated trigger thresholds are set for braking intents of different urgency levels; for example, a high-probability trigger threshold is set for the highest-priority emergency braking intent, while a suitable medium-probability trigger threshold is set for normal braking intents. Simultaneously, a corresponding probability hysteresis range is set to avoid frequent vacuum pump start-stop cycles. All thresholds can be optimized and adjusted through offline calibration. After completing probability matching and tiered determination, the main control unit generates corresponding pre-scheduled control commands based on the matched braking intent level and probability intensity. These pre-scheduled control commands differ from the passive triggering based on the lower limit of brake line vacuum by traditional electronic vacuum pumps. The responsive control commands, generated based on the predicted braking intent, are proactive control commands designed to preemptively regulate the vacuum source before the driver's braking action fully takes effect and the pipeline vacuum level drops. The commands fall into four main categories: vacuum pump pre-start commands, target speed control commands, continuous pressure-maintaining operation commands, and standby shutdown commands. Each category includes clearly defined start times, target operating speeds, target pipeline vacuum levels, maximum operating time, and fault protection trigger conditions. For example, when the probability of an emergency braking intent exceeds a preset high-probability threshold, a high-priority vacuum pump full-speed pre-start command is generated, requiring the electronic vacuum pump to operate at its rated maximum speed to raise the brake pipeline vacuum level to the highest safety reserve threshold in the shortest possible time, providing sufficient vacuum assistance for the upcoming emergency braking. When the probability of a normal braking intent exceeds the corresponding threshold, a low-speed pre-start or pressure-maintaining operation command is generated to keep the pipeline vacuum level within the optimal operating range. When only a slight braking intent probability is detected and all core probability values ​​have not reached the trigger threshold, a command to maintain the current operating state or standby is generated to prevent the vacuum pump from ineffectively starting or stopping.

[0025] Subsequently, the main control unit sends the generated pre-scheduled control commands to the electronic vacuum pump controller in real time via the vehicle's CAN bus in the form of high-priority messages. The electronic vacuum pump controller is the underlying execution control unit of the electronic vacuum pump, which integrates a power drive module, a bus communication module, an operating condition acquisition module, and a fault protection module. It establishes real-time bidirectional communication with the vehicle's main control unit. On the one hand, it is responsible for receiving control commands issued by the main control unit, and on the other hand, it transmits back data such as the operating status, operating current, pipeline vacuum feedback value, and fault codes of the electronic vacuum pump in real time.

[0026] Upon receiving the instruction, the electronic vacuum pump controller parses and verifies the validity of the pre-scheduled control instruction. Once the verification is successful, it drives the electronic vacuum pump to complete the corresponding execution action. The electronic vacuum pump is the power supply component of the brake vacuum booster system in new energy vehicles, used to replace the vacuum supply capability of the intake manifold of traditional fuel vehicle engines, providing a stable negative pressure vacuum source for the brake vacuum booster. Its start / stop status, operating speed, and running time directly determine the speed at which the vacuum level is established and maintained in the brake line. Execution refers to the electronic vacuum pump controller completing the entire process of controlling the electronic vacuum pump's operating status according to the requirements of the pre-scheduled control instruction. Specifically, this includes precisely controlling the vacuum pump's start-up timing, drive power, and operating speed, while simultaneously acquiring the vacuum level feedback signal of the brake line in real time, dynamically correcting the vacuum pump's operating status in a closed loop to ensure that the target vacuum level is reached within the time required by the instruction. At the same time, it monitors the vacuum pump's operating temperature, operating current, and supply voltage throughout the process, and immediately executes protective actions when overload, overheating, or undervoltage fault conditions are triggered to prevent component damage.

[0027] Furthermore, in the method provided in the application embodiment, the nodes of the Bayesian network include an environmental risk level node, a driver's historical braking tendency node, a current pedal feature node, and a braking intention posterior probability node; wherein, the environmental risk level node and the driver's historical braking tendency node serve as parent nodes pointing to the current pedal feature node, and the environmental risk level node, the driver's historical braking tendency node, and the current pedal feature node all point to the braking intention posterior probability node.

[0028] Specifically, this involves the entire process of defining Bayesian network nodes, partitioning the state space, and constructing directed causal connections. Bayesian network nodes are the basic units constituting the Bayesian probabilistic graphical model. Each node corresponds to a random variable, representing an observable or inferable core feature dimension in the entire braking intention recognition process. This solution predefines four logically closed-loop, interconnected core nodes for the Bayesian network: environmental risk level node, driver's historical braking tendency node, current pedal feature node, and posterior probability node of braking intention. In the specific implementation process, the standardization of these four core nodes is first completed. Definition and State Space Calibration: Firstly, the environmental risk level node, an observable discrete random variable node, corresponds to the comprehensive environmental risk level result obtained from the preceding steps, integrating forward collision risk and downhill risk. Its state space is pre-calibrated using a large amount of real driving accident data, dividing it into three mutually exclusive discrete states: low risk, medium risk, and high risk. Each state corresponds to a clear risk quantification interval and is the core input node characterizing the external objective safety risks of vehicle driving. Secondly, the driver's historical braking tendency node, an observable discrete and continuous compatible random variable node, corresponds to the environmental risk level result obtained from the preceding steps. The obtained quantitative value of the driver's historical braking tendency can be pre-divided into three discrete driving style labels: mild, normal, and aggressive. It can also be mapped to a continuous quantization interval between 0 and 1. The higher the value, the more aggressive the driver's braking operation style. It is a priori input node that represents the driver's long-term braking operation habits and personalized driving style. The third is the current pedal feature node, which is an observable discrete random variable node. It corresponds to the current pedal feature parameters extracted from the driver's real-time operation information. Its state space is based on the pedal displacement, pedal applied force, pedal instantaneous speed, and pedal travel rate of change, etc. The quantification interval of the heart feature is divided into four mutually exclusive discrete states: no operation, slight operation, normal operation, and rapid braking operation. It is the real-time core observation node that represents the driver's braking operation behavior at the current moment. The fourth is the braking intention posterior probability node, which is the target inference node of the entire Bayesian network, that is, the hidden variable node to be solved. It corresponds to the driver's true braking intention type at the current moment. Its state space is pre-divided into four mutually exclusive discrete states: no braking intention, slight braking intention, normal braking intention, and emergency braking intention, based on the actual needs of vehicle braking control. It is the final output target of the entire Bayesian probabilistic inference.

[0029] After defining the four nodes and partitioning their state spaces, a directed connection relationship is further established between the nodes to conform to the causal logic of real driving scenarios. The parent node, in the directed acyclic graph of the Bayesian network, is the upstream node that directly influences the probability distribution of its downstream child nodes. The corresponding child nodes are downstream nodes whose probability distribution is constrained by the parent node's state and determined based on different combinations of the parent node's states. First, a core causal relationship is established: the environmental risk level node and the driver's historical braking tendency node serve as parent nodes pointing to the current pedal feature node. This design restores the underlying causal logic in real driving scenarios, namely, the level of risk in the external driving environment and the driver's inherent braking habits directly determine the driver's braking behavior. For example, in a high-collision-risk environment, an aggressive driver is more likely to rapidly depress the brake pedal, while a mild-mannered driver in a low-risk, stable driving environment has a very low probability of making large pedal movements. The design provides a clear parent node dependency relationship for constructing the conditional probability table of the current pedal feature node, avoiding the problem of pedal features being isolated from the environment and driving habits in traditional recognition schemes. Subsequently, a second set of core causal pointing relationships is set: the environmental risk level node, the driver's historical braking tendency node, and the current pedal feature node together serve as parent nodes pointing to the posterior probability node of braking intention. This design fully covers the three core dimensions that determine the driver's braking intention: the risk constraints of the external objective environment, the driver's subjective historical braking habit tendency, and the driver's real-time pedal operation behavior. The three parent nodes provide complete causal input for the reasoning of braking intention from the three levels of objective environment, prior habits, and real-time behavior, without missing any core information. At the same time, this directed acyclic topology fully conforms to the probabilistic reasoning rules of Bayesian networks, completely avoiding the circular dependency problem, and providing a logically consistent and compliant topological foundation for the subsequent offline training of the conditional probability table and the real-time probabilistic reasoning of variable elimination.

[0030] Furthermore, in the method provided in the application embodiment, calculating the environmental risk level based on the environmental perception information and the vehicle status information includes: extracting the relative distance, relative speed, and downhill feature parameters between the vehicle and the vehicle in front from the environmental perception information; calculating the collision time based on the relative distance and relative speed, and determining the forward collision risk level by calling a preset collision time-level mapping relationship; extracting the vehicle speed from the vehicle status information, and calculating the downhill risk level by calling a pre-constructed downhill risk model in combination with the downhill feature parameters; and using the forward collision risk level and the downhill risk level together to constitute the environmental risk level.

[0031] Specifically, the vehicle braking system's main control unit first completes the parsing and core feature extraction of the pre-collected data. It extracts three types of core parameters from the environmental perception information. This environmental perception information refers to the structured dataset output by the vehicle's advanced driver assistance module, which uses millimeter-wave radar, a forward-facing binocular camera, and optional lidar to perform surrounding environment modeling, target recognition, and motion parameter calculation. The dataset includes the status of surrounding traffic participants and the alignment parameters of the road ahead. The three extracted parameters are the relative distance between the vehicle and the vehicle in front, relative speed, and downhill feature parameters. The relative distance, measured in meters, is the longitudinal straight-line distance between the front bumper of the vehicle and the rear of the nearest target vehicle in the same lane. This distance is obtained in real-time by millimeter-wave radar or lidar with a sampling frequency of at least 50Hz to ensure data real-time performance. The relative speed, measured in km / h, is the longitudinal speed difference between the vehicle's speed and the target vehicle's speed. A positive relative speed indicates that the vehicle and the vehicle in front are in a continuous approach. The status is negative, indicating that the longitudinal distance between the two vehicles is continuously increasing. The downhill feature parameters refer to the core parameter set that characterizes the geometric features of the downhill road section in the current and forward-looking range of the vehicle. Specifically, it includes the longitudinal slope value of the road, in percentage. Positive values ​​represent the downhill direction and the effective length of the downhill section, in meters. It is calculated by the forward-looking camera and the vehicle-mounted high-precision map data. The forward-looking range covers the road area in front of the vehicle for no less than 500 meters, avoiding the lack of risk prediction in long downhill scenarios. Simultaneously, the main control unit extracts the vehicle speed parameter from the vehicle status information. The vehicle status information refers to the parameter set that characterizes the vehicle's own driving conditions and the operating status of core components, which is collected in real time through the vehicle's CAN / LIN bus. The vehicle speed refers to the longitudinal driving speed of the vehicle body, in km / h, which is calculated by the joint calculation of the vehicle's four-wheel wheel speed sensors. Its sampling frequency is consistent with the sampling frequency of the environmental perception information. All extracted parameters are configured with a unified high-precision timestamp to ensure the time sequence alignment of multi-source data and avoid calculation errors caused by time sequence deviations.

[0032] After parameter extraction, the main control unit executes the quantitative calculations of two independent branches: forward collision risk and downhill risk, in parallel. In the first branch, the forward collision risk calculation, the main control unit first calculates the collision time based on the extracted relative distance and relative speed. The collision time, abbreviated as TTC, refers to the remaining time required for a longitudinal collision between the two vehicles under ideal conditions where the current relative motion states of the vehicle and the target vehicle ahead remain unchanged. Its core calculation formula is TTC=d rel / v rel In the formula d re v represents the relative distance between the vehicle and the vehicle in front, as extracted above. rel For the longitudinal relative velocity of this vehicle relative to the vehicle in front, which was extracted earlier, and to avoid invalid calculations and logical anomalies, boundary processing rules are pre-set: when vrel When d ≤ 0, meaning there is no tendency for the vehicle to approach the vehicle in front, the TTC is assigned an infinite value, corresponding to no collision risk; when d rel When the distance exceeds the maximum effective detection range of the vehicle radar, the default calibration is 150 meters, indicating no effective approaching target, and the TTC is also assigned to infinity. After calculating the collision time, the main control unit calls the preset collision time-level mapping relationship to determine and output the forward collision risk level. The preset collision time-level mapping relationship refers to the corresponding classification rule between collision time and collision probability obtained by statistically analyzing publicly available real vehicle collision accident datasets and millions of kilometers of natural driving data collected from real vehicle road tests. This mapping relationship divides the collision time into three mutually exclusive categories. The risk range corresponds to three levels of forward collision risk. The specific classification rules are as follows: when TTC ≤ 1.5s, the probability of a collision is extremely high, and the forward collision risk level is determined to be high risk; when 1.5s < TTC ≤ 3.0s, the probability of a collision is moderate, and the forward collision risk level is determined to be medium risk; when TTC > 3.0s or TTC is infinite, the probability of a collision is extremely low, and the forward collision risk level is determined to be low risk. At the same time, this mapping rule can be calibrated and adjusted offline according to the braking performance, body size, and braking system response characteristics of different vehicle models to ensure compatibility with different vehicle models.

[0033] In the parallel second downhill risk calculation branch, the main control unit inputs the extracted vehicle speed and the synchronously extracted downhill feature parameters into the pre-built downhill risk model to complete the calculation of the downhill risk level. The pre-built downhill risk model is a machine learning classification model specifically used to quantify the risk of brake fade and the probability of continuous braking failure on downhill sections. This solution specifically adopts a random forest classification model with strong generalization ability and excellent anti-interference performance. To ensure the feasibility and full disclosure of the solution, the model's construction and training process is as follows: First, the model's input features are determined to have three dimensions: the vehicle's real-time driving speed, the road's longitudinal slope value, and the effective length of the downhill section. The model's output is the downhill risk level, which is also divided into three mutually exclusive discrete levels: low risk, medium risk, and high risk, consistent with the classification system of the front collision risk level. Then, an offline training dataset is constructed, collecting real-vehicle driving data covering different terrains such as plains, mountains, and hills, and different load conditions. The system includes full-condition braking data under different vehicle speeds, gradients, and downhill lengths. Based on braking system temperature rise test data and braking accident probability statistics under the same conditions, each valid sample is labeled with its downhill risk level. A labeled training dataset containing no fewer than 100,000 valid samples is constructed. A five-fold cross-validation method is used for supervised training of the random forest model, optimizing core hyperparameters such as the number of decision trees, maximum tree depth, and minimum number of samples for node splits. After training, the model's risk level classification accuracy is no less than 95%. Generalization is verified through multi-condition real-vehicle road tests. Finally, the trained and verified model is stored in the vehicle's main control unit's automotive-grade storage chip for real-time use. During real-time calculation, the main control unit inputs three input features—vehicle speed, road gradient, and downhill length—after timestamp alignment, into the pre-constructed downhill risk model. After real-time inference, the model outputs the downhill risk level corresponding to the current downhill condition.

[0034] After calculating the risk levels of the two branches, the main control unit uses the aforementioned forward collision risk level and downhill risk level to construct the environmental risk level. Specifically, the principle of choosing the highest risk level, common to automotive-grade electronic control systems, is adopted to fuse the two risk levels. The fusion rules are as follows: when either of the two branches outputs a high risk level, the final comprehensive environmental risk level is directly determined to be high risk; when neither branch outputs a high risk level, but either outputs a medium risk level, the final comprehensive environmental risk level is determined to be medium risk; when both branches output a low risk level, the final comprehensive environmental risk level is determined to be low risk. The fused environmental risk level is a standardized three-level discrete quantization result, which can be directly mapped to the corresponding state of the environmental risk level node in the subsequent Bayesian network, providing standardized observational evidence for probabilistic inference.

[0035] Furthermore, in the method provided in the application embodiment, the collision time-level mapping relationship is obtained by analyzing the mapping relationship between collision time and collision probability level obtained from historical collision data; the downhill risk model is a random forest classification model, which is trained by historical vehicle speed data, historical downhill feature parameters and historical downhill risk level data, wherein the historical downhill risk level data is assigned based on the probability of braking accidents.

[0036] Specifically, the first step is to construct the collision time-level mapping relationship. This mapping relationship refers to a one-to-one correspondence rule, pre-defined through statistical data analysis, used to convert continuous collision time values ​​calculated in real time into standardized discrete levels of pre-collision risk. The construction process begins with the comprehensive collection and preprocessing of historical collision data. This historical collision data refers to multi-source authoritative real-vehicle test datasets used to statistically analyze collision occurrence patterns. Specifically, this includes publicly available forward collision accident samples from the CIDAS database for in-depth investigations of domestic traffic accidents, full-condition real-vehicle collision test data released by the National Motor Vehicle Product Quality Supervision and Inspection Center, and millions of kilometers of natural driving data collected from vehicle manufacturers' own fleets. The data is categorized into four main types: near-collision accident retrospective data, forward collision test data from the C-NCAP new car evaluation procedure, and more. The total number of valid samples is no less than 100,000. After data collection, a standardized preprocessing procedure is performed to remove redundant samples with missing data fields, abnormal sensor signals, or invalid test conditions. Valid samples containing complete collision time series, vehicle motion parameters, and the final collision result are selected. Then, with a time step of 0.5 seconds, the collision time is divided into continuous intervals from 0 to 10 seconds. The percentage of valid samples that ultimately collide within each collision time interval is calculated to obtain the collision probability corresponding to different collision time intervals. Subsequently, based on the statistical results, the collision probability level is classified and the mapping relationship is calibrated. The collision probability level refers to a standardized risk classification based on the probability of a collision, taking into account vehicle braking safety control requirements. This scheme divides it into three mutually exclusive and fully covered levels: high, medium, and low, corresponding to clearly defined probability ranges: a collision probability ≥ 80% is considered high risk; 20% ≤ collision probability < 80% is considered medium risk; and a collision probability < 20% is considered low risk. Based on this classification rule, the statistically obtained collision time intervals are linked to the collision probability levels to form a final collision time-level mapping relationship. Specifically, when the collision time ≤ 1.5s, the corresponding collision probability exceeds 85%, mapping to a high risk level; when 1.5s < collision time ≤ 3.0s... When the collision time is greater than 3.0s, the probability of a collision is between 25% and 78%, which is mapped to a medium-risk level. When the collision time is greater than 3.0s, the probability of a collision is less than 18%, which is mapped to a low-risk level. At the same time, the boundary condition calibration rules are supplemented: when the relative speed between the vehicle and the vehicle in front is less than or equal to 0 (i.e., there is no tendency to approach), and the relative distance exceeds the maximum effective detection range of the vehicle radar of 150m, the collision time is regarded as infinite and uniformly mapped to a low-risk level. This mapping relationship can be adjusted offline according to the performance parameters such as the curb weight, braking system response time, and maximum braking deceleration of different vehicle models. It also needs to be verified through actual vehicle track following and emergency braking condition tests to ensure that the accuracy of risk level determination is not less than 95% and to avoid misjudgment and omission.

[0037] The entire process of building and training a downhill risk model was completed simultaneously. This model is a machine learning classification model used to quantify the braking risk level of downhill sections based on vehicle driving status and downhill road geometry. This solution explicitly adopts a random forest classification model as the model architecture. The random forest classification model is a supervised machine learning model based on an ensemble learning framework, composed of multiple independent decision trees operating in parallel. It completes the classification output through multi-dimensional splitting decisions on input features, exhibiting strong anti-overfitting ability, good generalization, and high robustness to noisy data. It is fully adaptable to the requirements of complex automotive-grade real-world operating environments. The model is trained using three major historical data types. Based on historical datasets, the first category is historical vehicle speed data, which is real-time longitudinal driving speed data of vehicles during downhill driving, collected by actual vehicles, with a sampling frequency of no less than 10Hz, covering the entire vehicle speed range from 0km / h to 120km / h, including vehicle speed samples under empty, half-loaded, and fully loaded conditions; the second category is historical downhill feature parameters, which is core geometric feature data of downhill sections collected by actual vehicles, specifically including road longitudinal slope value, effective length of downhill section, slope covering the entire slope range from 0% to 15%, and downhill length covering the entire length range from 50m to 20km, including full-scene samples such as continuous long downhill in mountainous areas, short downhill in hilly areas, and downhill on urban ramps.The third category is historical downhill risk level data, which is the labeled data used for model supervised training. This data is strictly labeled according to the rules based on the probability of braking accidents. The probability of braking accidents refers to the probability of a traffic accident occurring under specific downhill conditions due to brake system thermal fade, continuous decay of brake line vacuum leading to decreased braking performance, brake failure, and ultimately, brake failure. The specific assignment rules are as follows: First, through brake system bench durability tests, the brake disc temperature rise curve, brake line vacuum decay curve, and braking performance decay rate are tested during continuous braking under different gradients, downhill lengths, and vehicle speeds. This is combined with domestic traffic accident data. The database contains braking accident statistics for different downhill conditions. The probability of braking accidents for each condition is determined, and risk levels are assigned based on these probabilities: a braking accident probability ≥70% is assigned a high-risk level, 20% ≤ probability <70% is assigned a medium-risk level, and a probability <20% is assigned a low-risk level. After completing the data collection and labeling for these three categories, a valid training dataset of no less than 150,000 records is constructed. The ratio of high, medium, and low-risk samples is controlled at 1:3:6 to avoid sample imbalance. Simultaneously, the dataset is divided into training, validation, and test sets in a 7:2:1 ratio to ensure model stability. To assess the effectiveness of training and validation, supervised training of the random forest classification model was subsequently completed. First, the model input features were defined as three dimensions: real-time vehicle speed, road longitudinal slope, and effective length of downhill sections. The output was a downhill risk level categorized into high, medium, and low. Then, the core hyperparameters were set: 100 decision trees, a maximum tree depth of 12, a minimum number of samples for node splits of 5, and a minimum number of samples for leaf nodes of 2. The Gini coefficient was used as the evaluation metric for node splits. Iterative training was performed using the training set data, minimizing the classification loss through optimization. Simultaneously, the model's generalization ability was monitored in real-time using the validation set data. Training is stopped early when the accuracy on the validation set no longer improves after 10 consecutive iterations to avoid overfitting. After training, the model performance is validated using test set data. The overall classification accuracy of the model is required to be no less than 96%, and the recall rate for high-risk operating conditions must be no less than 99%, eliminating any missed cases in high-risk scenarios. After successful validation, the trained model is converted into solidified code that can run in the vehicle's MCU using automotive-grade code generation tools and stored in the automotive-grade storage chip of the braking system's main control unit for real-time operation. An OTA online update interface is also reserved to optimize and upgrade model parameters based on road condition data from actual usage scenarios.

[0038] Furthermore, in the method provided in the application embodiment, the step of calculating the driver's historical braking tendency quantification value includes: statistically analyzing the average maximum pedal speed and average braking deceleration of all braking events completed by the driver within a preset time period; inputting the average maximum pedal speed and the average braking deceleration into a pre-trained clustering model, outputting a driver type label, and mapping the driver type label to a quantization value between 0 and 1 to obtain the driver's historical braking tendency quantification value.

[0039] Specifically, the first step is to screen and statistically analyze the basic data of braking behavior. A standardized definition of a braking event is established: a complete driver braking operation begins when the brake pedal is switched on and off, the pedal travel exceeds the preset minimum trigger threshold of 2% of the full travel, and ends when the brake pedal fully returns to its original position, the pedal travel returns to zero, and the vehicle's longitudinal braking deceleration drops to zero. This serves as the smallest statistical unit. A preset time period is defined, which is a valid time and mileage window determined in advance through vehicle calibration for retrospectively analyzing the driver's historical braking behavior characteristics. The default calibration is the vehicle's most recent 1000 kilometers or the natural driving cycle of the last 3 months. The minimum effective sample threshold is set to no less than 50 effective braking events. Invalid braking event samples are removed due to pedal mis-touch, extreme conditions triggering ABS, abnormal sensor signals, and low-adhesion icy or snowy roads. All raw data of effective braking events are stored in the vehicle's main control unit's automotive-grade non-volatile memory chip, which is not lost when power is off and is updated every time the vehicle is powered off.

[0040] Subsequently, two statistical characteristics were calculated. The first was the average maximum pedal speed. The maximum pedal speed for a single braking event refers to the maximum instantaneous rate of change of pedal displacement, calculated differentially from the pedal displacement data collected by a brake pedal displacement sensor with a sampling frequency of no less than 100Hz throughout the entire braking event, measured in mm / s. This directly reflects the speed and urgency of the driver's braking. The average maximum pedal speed is the arithmetic mean of the maximum pedal speeds corresponding to all effective braking events within a preset time period. The second was the average braking deceleration. The braking deceleration for a single braking event refers to the average absolute value of the vehicle's longitudinal deceleration, calculated jointly by the vehicle's longitudinal acceleration sensor and the four-wheel wheel speed sensors throughout the entire braking event, measured in m / s. 2 It directly reflects the intensity of the driver's braking operation and the strength of the braking demand, while the average braking deceleration is the arithmetic mean of the single average braking deceleration corresponding to all effective braking events within a preset time period.

[0041] After completing the statistical calculations of the two core features, they are input into a pre-trained clustering model to complete the driving style classification. The pre-trained clustering model refers to a machine learning model that has been pre-trained unsupervised using massive amounts of natural driving data, embedded in the vehicle's main control unit, and used to automatically classify driver braking style types based on braking behavior characteristics. This solution adopts the K-Means unsupervised clustering algorithm, which has strong generalization, excellent anti-interference capabilities, and is compatible with automotive-grade low-computing-power platforms. To avoid insufficient disclosure issues under patent law, the complete construction and training process of this model is clearly defined as follows: First, constructing a training dataset. The data sources include publicly available natural driving datasets released by the China Automotive Engineering Research Institute and millions of kilometers of real-vehicle driving data collected by OEMs covering drivers of different ages, driving experience, and regions. Two core features are extracted for each effective braking event: the maximum pedal speed and the average braking deceleration. A feature dataset with no less than 200,000 valid samples is constructed, fully covering all types of driving styles: mild, conventional, and aggressive. Second, data preprocessing: Z-score standardization is performed on the two feature dimensions. The standardization formula is: ; in This represents the mean of the entire dataset for the corresponding feature. To eliminate the dimensional differences between the two features, the standard deviation of the entire dataset is used to ensure consistent weighting of their contributions to the clustering results. The third step determines the optimal number of clusters. Through joint verification using the elbow rule and silhouette coefficient method, the optimal number of clusters is determined to be 3, corresponding to three different driver braking styles. The fourth step involves model training. The K-Means++ algorithm is used to initialize cluster centers, avoiding local optima caused by random initial centers. The maximum number of iterations is set to 300, and the convergence threshold is 1e-5. Unsupervised clustering training is performed on the standardized training dataset, ultimately yielding three stable cluster centers corresponding to three driver type labels: mild, conventional, and aggressive. The aggressive type... The braking style corresponds to high maximum pedal speed and high average braking deceleration; the normal type corresponds to a medium-level braking style; and the mild type corresponds to a low maximum pedal speed and low average braking deceleration. The fifth step is model performance verification. The silhouette coefficient is used to verify the cluster discrimination, requiring an average silhouette coefficient of no less than 0.75. Simultaneously, the classification accuracy is verified using 10,000 manually labeled driving style samples, requiring an overall classification accuracy of no less than 92%. After successful verification, the cluster center coordinates, standardized parameters, and classification rules of the trained clustering model are all solidified into the automotive-grade storage chip of the vehicle braking system's main control unit. An OTA online update interface is also reserved to adapt to changes in driver driving styles. The model parameters are optimized after periodic changes. During real-time classification calculations, the average pedal speed and average braking deceleration obtained from the aforementioned statistics are first standardized using Z-scores according to the standardized parameters determined in the pre-training phase. Then, they are input into the pre-trained clustering model. The model calculates the Euclidean distance from the standardized sample to the three cluster centers, assigns the sample to the nearest cluster center, and outputs the corresponding driver type label. After classification, the driver type label is mapped to a continuous quantized value between 0 and 1, ultimately obtaining the driver's historical braking tendency quantization value. This quantization value refers to the value used to standardize and characterize the driver's long-term aggressive braking behavior, and is a Bayesian network. A continuous index provides prior probability. The closer the value is to 1, the more aggressive the driver's braking style and the higher the probability of emergency braking. The closer the value is to 0, the more moderate the driver's braking style and the higher the probability of smooth braking. The specific mapping rules are as follows: the basic mapping value for the moderate label is 0.2, the basic mapping value for the normal label is 0.5, and the basic mapping value for the aggressive label is 0.8. At the same time, it is linearly fine-tuned within the range of ±0.1 based on the relative distance from the sample to the corresponding cluster center to ensure the continuity of the quantification value and avoid the problem of abrupt changes in discrete mapping. In addition, a cold start adaptation rule is set. When the vehicle is in a new state and the number of effective braking event samples is less than 50, the default value of 0 corresponding to the normal type is used.A value of 5 is used as the initial quantization value to avoid identification errors during the cold start phase. The quantization value is updated every 100 kilometers or weekly, re-statistically analyzing feature parameters and updating the quantization value to ensure accurate adaptation to the gradual changes in the driver's driving style.

[0042] Furthermore, in the method provided in the application embodiment, a Bayesian network is constructed. Current pedal feature parameters are extracted based on the driver's operation information and input into the Bayesian network along with the environmental risk level and the driver's historical braking tendency quantification value. The posterior probability distribution of braking intention at the current moment is calculated through probabilistic inference. This includes: pre-configuring a conditional probability table for each node in the Bayesian network; mapping the environmental risk level, the driver's historical braking tendency quantification value, and the current pedal feature parameters as observation evidence to the determined states of the corresponding nodes in the Bayesian network; and using variable elimination, based on the conditional probability tables of all nodes in the Bayesian network and the observation evidence, probabilistic summation is performed to eliminate non-evidence nodes, calculating the probability distribution of each state taken by the posterior probability node of braking intention under given observation evidence, which is then used as the posterior probability distribution of braking intention.

[0043] Specifically, the first step is to define and partition the discrete state space of four core nodes, ensuring that the state of each node is quantifiable, mappable, and usable for probability calculations. The four nodes are: First, the environmental risk level node, which is an observable discrete evidence node corresponding to the comprehensive environmental risk level calculated in the preceding steps, integrating pre-collision and downhill risks. Its state space is pre-divided into three mutually exclusive and fully covered discrete states: low risk, medium risk, and high risk, perfectly matching the output of the preceding environmental risk level calculation step; Second, the driver's historical braking tendency node, which is an observable discrete evidence node corresponding to the driver's historical braking tendency calculated in the preceding steps. The quantified value, a continuous quantified index between 0 and 1 used to standardize and characterize the aggressiveness of the driver's long-term braking operation, is pre-divided into three mutually exclusive discrete states: mild, normal, and aggressive, corresponding to the quantified value intervals [0, 0.35), [0.35, 0.65], and (0.65, 1], respectively. Third, the current pedal feature node is an observable discrete evidence node, corresponding to the current pedal feature parameters extracted from the driver's operation information. These current pedal feature parameters refer to the core quantified feature set that can characterize the driver's current braking operation behavior in real time, specifically including the real-time displacement of the brake pedal and the force applied by the pedal. The four core dimensions—instantaneous pedal movement speed, pedal travel rate of change, and other parameters—are collected in real-time by brake pedal displacement and pressure sensors with a sampling frequency of at least 100Hz. After preprocessing with sliding window filtering to remove noise interference and extract effective features, the state space of this node is pre-divided into four mutually exclusive discrete states: no operation, slight operation, normal operation, and rapid pressing operation. These correspond to quantization intervals of pedal travel < 2% of full travel, 2% ≤ pedal travel < 15% of full travel, 15% ≤ pedal travel < 50% of full travel, and pedal travel ≥ 50% of full travel with pedal speed ≥ 300mm / s, respectively. Fourth, the posterior probability node for braking intent is the entire Bayesian network. The target inference node, i.e. the hidden variable node to be solved, corresponds to the driver's true braking intention at the current moment. Its state space is pre-divided into four mutually exclusive discrete states: no braking intention, slight braking intention, normal braking intention, and emergency braking intention. This is completely matched with the hierarchical strategy of the back-end electronic vacuum pump pre-scheduling control. After the node definition is completed, the directed acyclic topology is built strictly according to the preset causal logic. Specifically, the environmental risk level node and the driver's historical braking tendency node serve as parent nodes pointing to the current pedal feature node. The environmental risk level node, the driver's historical braking tendency node, and the current pedal feature node together serve as parent nodes pointing to the posterior probability node of the braking intention.

[0044] Subsequently, the step of pre-configuring conditional probability tables for each node in the Bayesian network is executed. The conditional probability table is a core component of the Bayesian network, used to record the conditional probability values ​​of any child node in the network taking each discrete state under the condition that all its parent nodes take different state combinations. It is the core foundational data for probabilistic inference. In this scheme, the conditional probability table of each node is obtained by offline collection of real driving data and training using the Expectation-Maximization (EM) algorithm. Its complete implementation is as follows: First, construct an offline training dataset. The data source covers publicly available domestic natural driving datasets and millions of kilometers of real vehicle driving data collected by vehicle manufacturers, covering different ages, driving experience, driving styles, road conditions, and risk scenarios. Extract all feature data corresponding to the four nodes of the Bayesian network, including environmental risk level, driver braking tendency type, pedal feature state, and manually labeled real braking intention labels. Construct a valid labeled sample set with a total of no less than 300,000 samples, covering all state combinations of all nodes without sample blind spots. Second, for scenarios with hidden variables, use the Expectation-Maximization algorithm to complete... The training of the conditional probability table involves an iterative optimization algorithm for estimating parameters of probabilistic models with latent variables. This algorithm converges to the optimal conditional probability parameters even when some data is unobservable. The training process consists of two iterative steps: the E-step (expectation step), which calculates the posterior probability expectation of the latent variables and braking intention nodes based on the initial values ​​of the conditional probability table in the current iteration and the observed data from the training dataset; and the M-step (maximization step), which updates the conditional probability table parameters of each node using maximum likelihood estimation based on the expected results calculated in the E-step. The E-step and M-step iterations are repeated until the change in the parameters of the conditional probability table is less than a preset convergence threshold of 1e-6, at which point the iteration stops, resulting in the final conditional probability table. The third step is the validation and solidification of the conditional probability table. The trained conditional probability table is validated using 50,000 independent real-vehicle test samples. The overall accuracy of braking intention recognition is required to be no less than 95%, and the recall rate of emergency braking intention recognition is required to be no less than 99%. After successful validation, the conditional probability tables of all nodes are solidified into the automotive-grade non-volatile memory chip of the vehicle braking system's main control unit for real-time inference.

[0045] After completing the construction of the Bayesian network and configuring the conditional probability table, the current pedal feature parameters are extracted in real time. The extracted current pedal feature parameters, the environmental risk level output from the preceding steps, and the driver's historical braking tendency quantification value are then input into the constructed Bayesian network. Subsequently, the first step of real-time inference is executed: the environmental risk level, the driver's historical braking tendency quantification value, and the current pedal feature parameters are used as observational evidence and mapped to the determined states of the corresponding nodes in the Bayesian network. Observational evidence refers to directly observable and determined states during the Bayesian network inference process. The known variables are used to provide known conditions for calculating the posterior probability of the latent variables. The specific mapping rules are as follows: First, the comprehensive environmental risk level output by the preceding steps is directly mapped to the corresponding discrete state of the environmental risk level node. High risk, medium risk, and low risk correspond to the three definite states of the node, respectively. Second, the driver's historical braking tendency quantization value between 0 and 1 output by the preceding steps is mapped to the corresponding discrete state of the driver's historical braking tendency node according to the preset interval division rules. Quantization value < 0.35 is mapped to the mild type, 0.35 ≤ quantization value ≤ 0.65 is mapped to the normal type, and quantization value > 0.65 is mapped to the aggressive type; thirdly, the extracted current pedal feature parameters are mapped to the corresponding discrete states of the current pedal feature nodes according to the preset pedal travel and pedal speed range division rules, completing the state locking of the three evidence nodes. The update frequency of all observed evidence is not less than 50Hz to ensure the real-time and continuous nature of the reasoning; after completing the mapping and locking of the observed evidence, the variable elimination method is executed. Based on the conditional probability table of all nodes in the Bayesian network and the observed evidence, the non-evidence nodes are summed and eliminated to calculate the probability distribution of the braking intention posterior probability node under given observed evidence. This is the core probabilistic reasoning step for the braking intention posterior probability distribution at the current moment. Among them, the variable elimination method is the Bayesian network's... The most commonly used exact inference algorithm in Yeesian network inference is based on the principle of factoring the joint probability distribution and eliminating non-evidence and non-target hidden variables by summing and integrating according to a preset elimination order. This significantly reduces the complexity of probability calculation and is suitable for the low computing power and high real-time requirements of automotive-grade MCUs. Probabilistic inference refers to the entire process of solving the posterior probability of the target hidden variable based on known observation evidence. The steps of the variable elimination method in this scheme are as follows: First, based on the topology and conditional probability table of the Bayesian network, construct the factorization formula of the joint probability distribution of all nodes. Combining the topology of this scheme, the joint probability distribution is decomposed as: P(environmental risk level, driver's historical braking tendency, current pedal characteristics, braking intention) = P(Environmental Risk Level) • P(Driver's Historical Braking Tendency) • P(Current Pedal Characteristics | Environmental Risk Level, Driver's Historical Braking Tendency) • P(Brake Intent | Environmental Risk Level, Driver's Historical Braking Tendency, Current Pedal Characteristics), where P(A|B) represents the conditional probability of A occurring given that B occurs, corresponding to the pre-trained parameters in the conditional probability table; The second step is to determine the elimination order and target variable. In this scheme, the target variable is the posterior probability node of braking intent. The states of the three evidence nodes are already locked to definite values, and there are no other non-evidence or non-target variables. Therefore, the posterior probability of the target variable is calculated directly based on Bayes' theorem and combined with the definite states of the observed evidence; The third step is to substitute the definite states of the observed evidence and combine... The pre-trained parameters in the conditional probability table calculate the conditional probability values ​​for the posterior probability nodes of braking intention in four states: no braking intention, slight braking intention, normal braking intention, and emergency braking intention. After normalization, the posterior probability distribution of braking intention at the current moment is obtained, where the sum of the probability values ​​of the four states is 1. This posterior probability distribution of braking intention refers to the set of probability values ​​for the driver's current braking intention in one of the four preset states, given three types of observational evidence: current environmental risk, driver's historical braking habits, and real-time pedal operation. It fully characterizes the type and intensity of the driver's braking demand at the current moment. The execution time of a single iteration of the entire inference process does not exceed 20ms, fully meeting the high real-time requirements of vehicle braking control.

[0046] Furthermore, in the method provided in the application embodiment, the conditional probability table records the probability values ​​of any node taking each state under different combinations of all parent nodes; the conditional probability table is obtained by offline collection of real driving data and training using the expectation-maximization algorithm.

[0047] Specifically, the first step is to standardize and design the structure of the conditional probability table. The conditional probability table (CPT) is a core parameter matrix in a Bayesian directed acyclic graph (DAG) model used to quantify causal dependencies between nodes. It is defined as recording the probability of any node taking any state given all combinations of its parent nodes. A parent node is an upstream node in the causal topology of a Bayesian network that directly causally constrains the probability distribution of its downstream child nodes. This scheme, combined with a pre-constructed four-node Bayesian network topology, designs a conditional probability table structure for each node that perfectly matches the causal logic. Specifically, for the two root nodes without parent nodes—the environmental risk level node and the driver's historical braking tendency node—the conditional probability table is a prior probability table of the corresponding node's discrete states. The environmental risk level node includes three discrete states: low risk, medium risk, and high risk, and its prior probability table is a 1x3 parameter matrix. The driver's historical braking tendency node includes three discrete states: mild, normal, and aggressive, and its prior probability table is also a 1x3 parameter matrix. For nodes with parent nodes, the conditional probability table is further defined as follows: The current pedal feature node, which shares a common parent node, has a total of 3×3=9 possible state combinations with its parent nodes. This node itself contains four discrete states: no operation, slight operation, normal operation, and rapid pedal press. Therefore, its conditional probability table is a 9x4 parameter matrix, where each row corresponds to a set of parent node state combinations, and each column corresponds to a discrete state of the node itself. Each parameter value in the matrix represents the conditional probability value of the node taking the corresponding state under the given parent node state combination. This is relevant for nodes based on environmental risk level, driver's historical braking tendency, and the current pedal feature node. The posterior probability node of braking intention, which shares a common parent node, has a total of 3×3×4=36 possible combinations of parent node states. This node itself contains four discrete states: no braking intention, slight braking intention, normal braking intention, and emergency braking intention. Therefore, its conditional probability table is a 36-row, 4-column parameter matrix that completely covers the probability distribution of braking intention under all parent node state combinations. All parameter values ​​in the conditional probability table are constrained to be between 0 and 1. Furthermore, under the same parent node state combination, the sum of the probability values ​​of all discrete states of the child node is strictly equal to 1, ensuring the mathematical validity of the probability calculation.

[0048] After completing the structural design of the conditional probability table, the entire process of offline collection of real driving data was executed. This provided sufficient, compliant, and real-world driving scenario-appropriate data sources for the training of the conditional probability table. Offline collection of real driving data refers to the process of collecting real-vehicle driving behavior data covering all scenarios, all operating conditions, and all types of driving styles through multiple channels before deploying the model in the actual vehicle, and completing standardized preprocessing and annotation. Specifically, this involved: first, determining multi-source compliant data sources, specifically covering four categories: publicly available forward collision and braking condition samples from the China In-Depth Accident Investigation System (CIDAS) database; a million-kilometer-level publicly available natural driving dataset released by the China Automotive Engineering Research Institute; over 3 million kilometers of natural driving data collected from real-vehicle road tests by OEMs, covering different regions and road conditions across the country; and full-type braking condition test data from the C-NCAP and C-IASI new car evaluation procedures. This ensured the authority and scenario coverage of the data sources. Subsequently, data screening and preprocessing were completed, extracting feature fields corresponding to the four nodes of the Bayesian network from the raw data, including environmental risk level sequences, driver braking behavior feature sequences, etc. The brake pedal operation feature sequence and vehicle braking deceleration sequence were analyzed. Invalid samples such as abnormal sensor signals, extreme extreme conditions triggered by ABS, accidental brake pedal touch, and low-adhesion roads on icy or snowy surfaces were removed. At the same time, timestamp alignment was performed on all valid samples to ensure that the temporal error of multi-source feature data within the same sample does not exceed 10ms, avoiding causal logic errors caused by temporal deviations. Subsequently, sample annotation was completed. For each valid braking event sample, professional engineers with experience in vehicle braking system development completed the manual annotation of the true braking intention by combining the pedal operation, vehicle deceleration, and environmental conditions throughout the braking process. The annotation results corresponded completely with the four discrete states of the posterior probability node of the braking intention. Finally, a valid annotated sample set with a total of no less than 300,000 samples was constructed. The ratio of high, medium, and low environmental risk samples was 1:2:7, the ratio of mild, normal, and aggressive driving style samples was 2.5:6:1.5, and the ratio of no braking, slight braking, normal braking, and emergency braking intention samples was 3:2:4:1. This ensured that the sample set had no scene blind spots and no sample imbalance, providing a reliable data foundation for subsequent model training.

[0049] After completing the dataset construction, the expectation-maximization algorithm is used to iteratively train the conditional probability tables of all nodes. This algorithm can converge to the globally optimal parameters of the conditional probability tables even in scenarios where some samples contain latent variables, meaning braking intent cannot be directly observed. This fully meets the training requirements of the Bayesian network in this scheme. The training process is as follows: First, parameter initialization: initial values ​​that conform to real driving logic are set for the conditional probability tables of all nodes. The initial values ​​of the prior probability tables of the two root nodes are set based on the statistical distribution of the sample set. The initial prior values ​​for the environmental risk level node are low risk 70%, medium risk 20%, and high risk 10%. The initial prior values ​​for the driver's historical braking tendency node are mild 25%, normal 60%, and aggressive 15%. The initial values ​​of the conditional probability tables of the two child nodes are set by combining the experience of braking system development experts with small sample statistical results to ensure that the initial parameters conform to the causal logic of real driving and avoid getting trapped in local optima during the iteration process. Second, the E-step (expectation step): based on the conditional probability table parameters of the current iteration, combined with observable features from offline collected real driving data, namely environmental risk level, driver braking tendency type, and pedal characteristic state, the parameters are calculated using Bayesian formula. The first step is to calculate the latent variables, specifically the posterior probability distribution of braking intention nodes under the current observed features. This involves calculating the expected probability of braking intention taking each discrete state in each sample, filling in the missing information in the incomplete data, and solving for the log-likelihood expectation of the complete data. The second step, the M-step, is the maximization step. Based on the log-likelihood expectation of the complete data calculated in the E-step, all parameters of the conditional probability tables of all nodes are updated using the maximum likelihood estimation method. This involves recalculating the conditional probability values ​​of each child node taking each discrete state under all combinations of its parent node's states, ensuring that the updated parameters maximize the complete data. The log-likelihood function of the data; the fourth step is convergence judgment and iteration termination. Repeat the E-step and M-step iteration process. After each iteration, calculate the maximum absolute change of all conditional probability table parameters after the current update and the parameters of the previous round. When the maximum absolute change of all parameters is less than the preset convergence threshold 1e-6, or the number of iterations reaches the preset maximum number of iterations of 1000, stop the iteration. At the same time, in order to further avoid the local optimum problem, five different initial parameter values ​​are used for training. Finally, the set of convergent parameters with the largest log-likelihood value is selected as the final conditional probability table.After training, the performance of the conditional probability table is verified and solidified to automotive-grade standards. 50,000 real-vehicle test samples, completely independent of the training set, are used to verify the generalization performance of the trained conditional probability table. The verification requirements are: overall braking intention recognition accuracy no less than 95%, emergency braking intention recognition recall no less than 99%, regular braking intention recognition accuracy no less than 93%, and slight braking intention recognition accuracy no less than 90%. Simultaneously, the accuracy fluctuation should not exceed 5% in sub-sample sets under different road conditions and driving styles to ensure the robustness and generalization ability of the conditional probability table. After successful verification, the conditional probability tables of all nodes are converted into a two-dimensional array format that can be directly called by an automotive-grade MCU and solidified into the automotive-grade non-volatile storage chip of the vehicle braking system's main control unit to ensure data integrity even when power is lost. An OTA online update interface is also reserved, allowing for incremental optimization of the conditional probability table based on subsequently collected real-vehicle operational data to adapt to differences in driving habits across different regions and usage scenarios.

[0050] In summary, the vehicle electronic vacuum pump control method based on braking intention recognition provided in this application has the following technical effects: By integrating vehicle status, driver operation, and ADAS environmental perception information, a multi-dimensional environmental risk quantification system encompassing forward collision risk and downhill risk is constructed. Combining a driver's historical braking tendency quantification method with a four-node Bayesian network probabilistic inference model, the system achieves advanced recognition of the driver's braking intention and generates pre-scheduled control commands for the electronic vacuum pump. This allows for vacuum reserve to be completed in advance before the braking action takes effect, effectively solving the response lag problem of passive control and significantly improving braking safety in emergency braking and long downhill conditions. At the same time, the hierarchical pre-scheduled control avoids ineffective start-stop of the vacuum pump, reducing its operating energy consumption and mechanical wear, and extending the service life of components.

[0051] Example 2 is based on the same inventive concept as the vehicle electronic vacuum pump control method based on braking intention recognition in the previous examples, such as... Figure 2 As shown in the embodiment of this application, a vehicle electronic vacuum pump control system based on braking intention recognition is provided. The system includes: The acquisition module 11 is used to acquire vehicle status information, driver operation information, and environmental perception information from the advanced driver assistance module; the calculation module 12 is used to calculate the environmental risk level based on the environmental perception information and the vehicle status information, including the forward collision risk level calculated based on collision time and the downhill risk level based on road slope; the inference module 13 is used to construct a Bayesian network, extract the current pedal feature parameters based on the driver operation information, and input them into the Bayesian network along with the environmental risk level and the driver's historical braking tendency quantification value, and calculate the posterior probability distribution of braking intention at the current moment through probabilistic inference; the execution module 14 is used to generate a pre-scheduled control command for the electronic vacuum pump based on the posterior probability distribution of braking intention and a preset probability threshold, and send it to the electronic vacuum pump controller for execution.

[0052] Furthermore, the inference module 13 is also used to perform the following steps: the nodes of the Bayesian network include an environmental risk level node, a driver's historical braking tendency node, a current pedal feature node, and a braking intention posterior probability node; wherein, the environmental risk level node and the driver's historical braking tendency node serve as parent nodes pointing to the current pedal feature node, and the environmental risk level node, the driver's historical braking tendency node, and the current pedal feature node all point to the braking intention posterior probability node.

[0053] Furthermore, the calculation module 12 is also used to perform the following steps: extracting the relative distance, relative speed, and downhill feature parameters between the vehicle and the vehicle in front from the environmental perception information; calculating the collision time based on the relative distance and relative speed, and calling a preset collision time-level mapping relationship to determine the forward collision risk level; extracting the vehicle speed from the vehicle status information, and combining the downhill feature parameters to call a pre-built downhill risk model to calculate the downhill risk level; and using the forward collision risk level and the downhill risk level together to constitute the environmental risk level.

[0054] Furthermore, the calculation module 12 is also used to perform the following steps: the collision time-level mapping relationship is obtained by analyzing the mapping relationship between collision time and collision probability level from historical collision data; the downhill risk model is a random forest classification model, which is trained by historical vehicle speed data, historical downhill feature parameters and historical downhill risk level data, wherein the historical downhill risk level data is assigned based on the probability of braking accidents.

[0055] Furthermore, the inference module 13 is also used to perform the following steps: statistically analyze the average maximum pedal speed and average braking deceleration of all braking events completed by the driver in the past preset time period; input the average maximum pedal speed and the average braking deceleration into a pre-trained clustering model, output a driver type label, and map the driver type label to a quantized value between 0 and 1 to obtain the driver's historical braking tendency quantization value.

[0056] Furthermore, the inference module 13 is also used to perform the following steps: pre-configure a conditional probability table for each node in the Bayesian network; map the environmental risk level, the driver's historical braking tendency quantification value, and the current pedal feature parameters as observation evidence to the determined state of the corresponding node in the Bayesian network; and use variable elimination to sum and eliminate non-evidence nodes according to the conditional probability table of all nodes in the Bayesian network and the observation evidence, and calculate the probability distribution of the braking intention posterior probability node taking each state under given observation evidence, as the braking intention posterior probability distribution.

[0057] Furthermore, the inference module 13 is also used to perform the following steps: the conditional probability table records the probability values ​​of any node taking each state under different combinations of all parent nodes; the conditional probability table is obtained by offline collection of real driving data and training using the expectation-maximization algorithm.

[0058] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for controlling an automotive electronic vacuum pump based on braking intent recognition, characterized in that, include: Acquire vehicle status information, driver operation information, and environmental perception information from advanced driver assistance modules; The environmental risk level is calculated based on the environmental perception information and the vehicle status information, including the forward collision risk level calculated based on the collision time and the downhill risk level calculated based on the road slope. A Bayesian network is constructed, and the current pedal feature parameters are extracted based on the driver's operation information. These parameters, along with the environmental risk level and the driver's historical braking tendency quantification value, are input into the Bayesian network. The posterior probability distribution of the braking intention at the current moment is calculated through probabilistic inference. Based on the posterior probability distribution of the braking intention and the preset probability threshold, a pre-scheduling control command for the electronic vacuum pump is generated and sent to the electronic vacuum pump controller for execution.

2. The vehicle electronic vacuum pump control method based on braking intention recognition as described in claim 1, characterized in that, The nodes of the Bayesian network include environmental risk level nodes, driver's historical braking tendency nodes, current pedal feature nodes, and braking intention posterior probability nodes. The environmental risk level node and the driver's historical braking tendency node serve as parent nodes pointing to the current pedal feature node. The environmental risk level node, the driver's historical braking tendency node, and the current pedal feature node all point to the posterior probability node of braking intention.

3. The vehicle electronic vacuum pump control method based on braking intention recognition as described in claim 1, characterized in that, Calculating the environmental risk level based on the environmental perception information and the vehicle status information includes: The relative distance, relative speed, and downhill characteristic parameters between the vehicle and the vehicle in front are extracted from the environmental perception information. The collision time is calculated based on the relative distance and relative speed, and the forward collision risk level is determined by calling the preset collision time-level mapping relationship; The vehicle speed is extracted from the vehicle status information, and the downhill risk level is calculated by calling a pre-built downhill risk model in combination with the downhill feature parameters. The environmental risk level is composed of the forward collision risk level and the downhill risk level.

4. The vehicle electronic vacuum pump control method based on braking intention recognition as described in claim 3, characterized in that, The collision time-level mapping relationship is obtained by analyzing historical collision data to determine the mapping relationship between collision time and collision probability level. The downhill risk model is a random forest classification model, which is trained using historical vehicle speed data, historical downhill feature parameters, and historical downhill risk level data. The historical downhill risk level data is assigned based on the probability of braking accidents.

5. The vehicle electronic vacuum pump control method based on braking intention recognition as described in claim 1, characterized in that, The calculation steps for the driver's historical braking tendency quantification value include: The system calculates the average maximum pedal speed and average braking deceleration for all braking events performed by the driver within a preset time period. The maximum average pedal speed and the average braking deceleration are input into a pre-trained clustering model, which outputs a driver type label. The driver type label is then mapped to a quantized value between 0 and 1 to obtain the driver's historical braking tendency quantization value.

6. The vehicle electronic vacuum pump control method based on braking intention recognition as described in claim 2, characterized in that, A Bayesian network is constructed, and the current pedal feature parameters are extracted based on the driver's operation information. These parameters, along with the environmental risk level and the driver's historical braking tendency quantification value, are input into the Bayesian network. Probabilistic inference is used to calculate the posterior probability distribution of the braking intention at the current moment, including: A conditional probability table is pre-configured for each node in the Bayesian network; The environmental risk level, the driver's historical braking tendency quantification value, and the current pedal characteristic parameters are used as observation evidence and mapped to the determination state of the corresponding node in the Bayesian network, respectively. Using the variable elimination method, based on the conditional probability table of all nodes in the Bayesian network and the observation evidence, non-evidence nodes are eliminated by probability summation, and the probability distribution of braking intention posterior probability nodes taking each state under given observation evidence is calculated, which is used as the braking intention posterior probability distribution.

7. The vehicle electronic vacuum pump control method based on braking intention recognition as described in claim 6, characterized in that, The conditional probability table records the probability values ​​of any node taking each state under different combinations of all parent nodes taking states. The conditional probability table was obtained by collecting real driving data offline and training it using the expectation-maximization algorithm.

8. A vehicle electronic vacuum pump control system based on braking intention recognition, characterized in that, The system is used to execute the vehicle electronic vacuum pump control method based on braking intention recognition as described in any one of claims 1-7, the system comprising: The data acquisition module is used to acquire vehicle status information, driver operation information, and environmental perception information from the advanced driver assistance module. The calculation module is used to calculate the environmental risk level based on the environmental perception information and the vehicle status information, including the forward collision risk level calculated based on the collision time and the downhill risk level calculated based on the road slope. The inference module is used to construct a Bayesian network, extract the current pedal feature parameters based on the driver's operation information, and input them into the Bayesian network along with the environmental risk level and the driver's historical braking tendency quantification value. The posterior probability distribution of the braking intention at the current moment is calculated through probabilistic inference. The execution module is used to generate a pre-scheduled control command for the electronic vacuum pump based on the posterior probability distribution of the braking intention and a preset probability threshold, and send it to the electronic vacuum pump controller for execution.

Images