Electronic vacuum pump pre-starting control method and system based on pedal behavior recognition

Abstract

This invention discloses a pre-start control method and system for an electronic vacuum pump based on pedal behavior recognition, belonging to the field of electronic control technology for new energy vehicles. The method includes: collecting multi-dimensional pedal behavior data; preprocessing the multi-dimensional pedal behavior data and synchronizing a first pedal behavior feature set to a first intent recognition model; entering a first pre-start mode when preset pre-start conditions are met; obtaining a second pedal behavior feature set and synchronizing it to a second intent recognition model for braking intent confirmation; generating a second pre-start control command when the braking intent confirmation result is a braking intent, switching the electronic vacuum pump to the second pre-start mode; stopping the pre-start operation of the electronic vacuum pump when the confirmation result is no braking intent. This application solves the technical problems in the prior art where the electronic vacuum pump adopts a post-pump start strategy, and a single intent recognition model cannot simultaneously consider response speed and accuracy, resulting in long braking assist delays and high false start rates.

Description

Technical Field

[0001] This invention relates to the field of electronic control technology for new energy vehicles, specifically to an electronic vacuum pump pre-start control method and system based on pedal behavior recognition. Background Technology

[0002] With the rapid development of the new energy vehicle industry, vacuum-assisted braking systems remain the mainstream braking solution for passenger cars. Unlike traditional fuel vehicles that can obtain a vacuum source from the engine intake manifold, new energy vehicles lack engine vacuum supply and rely entirely on electronic vacuum pumps to provide assistance to the braking system. The control performance of these pumps directly determines braking response speed and driving safety. In complex scenarios such as high-speed driving and autonomous driving, every 100ms increase in braking delay significantly increases the risk of collision, placing higher demands on the pre-start control of electronic vacuum pumps. Currently, mainstream electronic vacuum pumps generally adopt a passive control strategy of starting after being pressed, which results in a delay in assist build-up and makes it difficult to meet the safety requirements of emergency braking. Existing pre-start solutions mostly use a single intent recognition model, which cannot balance response speed and recognition accuracy. A high false start rate leads to unnecessary energy consumption and noise interference, seriously affecting driving comfort and system lifespan. Therefore, an efficient and reliable pre-start control method is urgently needed. Summary of the Invention

[0003] This application provides a pre-start control method and system for electronic vacuum pumps based on pedal behavior recognition, aiming to solve the technical problems of long braking assist delay and high false start rate caused by the use of a post-pump start strategy in existing electronic vacuum pumps and the inability of a single intent recognition model to balance response speed and accuracy.

[0004] In view of the above problems, this application provides a method and system for pre-start control of electronic vacuum pumps based on pedal behavior recognition.

[0005] The first aspect disclosed in this application provides a pre-start control method for an electronic vacuum pump based on pedal behavior recognition. The method includes: real-time acquisition of multi-dimensional pedal behavior data of the brake pedal via a pedal behavior sensor within a preset time window; preprocessing the multi-dimensional pedal behavior data to obtain a first pedal behavior feature set; synchronizing the first pedal behavior feature set to a first intention recognition model to obtain a braking intention prediction result; when the braking intention prediction result meets preset pre-start conditions, acquiring the real-time vacuum level of the electronic vacuum pump; generating a first pre-start control command based on the real-time vacuum level and the braking intention prediction result to start the electronic vacuum pump into a first pre-start mode; executing the first pre-start mode to collect data from the brake pedal in real time, obtaining a second pedal behavior feature set, synchronizing it to a second intention recognition model for braking intention confirmation, and obtaining a braking intention confirmation result; when the braking intention confirmation result is confirmed, generating a second pre-start control command to switch the electronic vacuum pump to a second pre-start mode; when the braking intention confirmation result is no braking intention, generating an exit control command to stop the pre-start operation of the electronic vacuum pump.

[0006] Another aspect of this application discloses an electronic vacuum pump pre-start control system based on pedal behavior recognition. This system includes: a behavior data acquisition module, used to acquire multi-dimensional pedal behavior data of the brake pedal in real time through a pedal behavior sensor within a preset time window; a prediction result acquisition module, used to preprocess the multi-dimensional pedal behavior data to obtain a first pedal behavior feature set, synchronize the first pedal behavior feature set to a first intention recognition model, and obtain a braking intention prediction result; and a control command generation module, used to acquire the real-time vacuum level of the electronic vacuum pump when the braking intention prediction result meets preset pre-start conditions, and based on the real-time vacuum level... The system generates a first pre-start control command based on the braking intention prediction result, and starts the electronic vacuum pump to enter the first pre-start mode; the confirmation result acquisition module is used to perform real-time acquisition of the brake pedal in the first pre-start mode, obtain a second pedal behavior feature set, synchronize it to the second intention recognition model for braking intention confirmation, and obtain a braking intention confirmation result; the braking intention judgment module is used to generate a second pre-start control command when the braking intention confirmation result is braking intention confirmation, and switch the electronic vacuum pump to the second pre-start mode, and generate an exit control command when the braking intention confirmation result is no braking intention, and stop the pre-start operation of the electronic vacuum pump.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: By employing a technical solution that uses a first intention recognition model to quickly predict braking intention and trigger the electronic vacuum pump to pre-start in stages, and a second intention recognition model to perform high-precision secondary confirmation of braking intention, and then executes staged power control based on the confirmation result, this solution solves the technical problems in the prior art where the electronic vacuum pump starts after being pressed, resulting in long braking assist delays and a high false start rate due to the inability of a single intention recognition model to balance response speed and accuracy. This achieves the technical effects of shortening braking assist delays, reducing false start rates, and effectively reducing system energy consumption and operating noise.

[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 the electronic vacuum pump pre-start control method based on pedal behavior recognition is provided for embodiments of this application; Figure 2 A schematic diagram of the structure of an electronic vacuum pump pre-start control system based on pedal behavior recognition is provided for embodiments of this application.

[0010] Explanation of reference numerals in the attached diagram: Behavior data acquisition module 11, prediction result acquisition module 12, control command generation module 13, confirmation result acquisition module 14, braking intention judgment module 15. 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 an electronic vacuum pump pre-start control method and system based on pedal behavior recognition. It collects brake pedal behavior data in real time, quickly predicts braking intention using a first intention recognition model, and triggers graded pre-start of the electronic vacuum pump. A second intention recognition model performs high-precision secondary confirmation, and based on the confirmation result, switches between full-power operation and stops pre-start, balancing braking response speed and control accuracy.

[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 1As shown in the embodiment of this application, a pre-start control method for an electronic vacuum pump based on pedal behavior recognition is provided. The method includes: Step S100: Collect multi-dimensional pedal behavior data of the brake pedal in real time through the pedal behavior sensor within a preset time window.

[0015] Specifically, a preset time window refers to a fixed-length sliding time window, a pre-defined time interval used to capture continuous time-series data. The window slides forward in fixed steps to achieve uninterrupted data acquisition. For example, a 200ms window with a 50ms step size means that a sample containing complete data from the most recent 200ms is output every 50ms. Pedal behavior sensors are dedicated sensors installed at different positions on the brake pedal to collect the pedal's physical motion parameters. These include four types: cable-operated displacement sensors, thin-film pressure sensors, triaxial accelerometers, and angle sensors. Multidimensional pedal behavior data includes both time-domain and physical quantity dimensions of pedal motion time-series data. Core dimensions include absolute pedal displacement, instantaneous velocity, angular acceleration, pedal force, rate of change of stroke, pedal angle, and the first and second derivatives of each parameter.

[0016] Specifically, after the new energy vehicle is powered on, the pedal data acquisition module is initialized, a preset time window with a fixed length of 200ms and a sliding step of 50ms is set, and the sampling frequency of all pedal behavior sensors is uniformly calibrated to 1kHz to ensure complete synchronization of data timestamps. A cable-type displacement sensor installed in the middle of the brake pedal arm collects the absolute displacement value of the pedal in real time. Four distributed thin-film pressure sensors installed under the pedal surface synchronously collect the distribution of the driver's foot force and the total pressure. A triaxial accelerometer installed at the pedal pivot collects the angular acceleration and direction of motion of the pedal, and an angle sensor installed at the hinge point collects the rotation angle of the pedal. The raw analog signals output by all sensors are converted into 16-bit digital signals by the AD conversion module of the vehicle ECU and then stored in a circular buffer according to the timestamp. When the amount of data in the buffer reaches 200 sampling points, corresponding to a 200ms time window, all multi-dimensional time-series data in the window are immediately extracted and packaged into the input samples for the first-level intent recognition. At the same time, the sliding window advances forward by 50ms and continues to collect the next set of data. This cycle repeats to achieve millisecond-level real-time monitoring of brake pedal behavior.

[0017] This step, through a fixed sliding time window acquisition mechanism, ensures real-time braking intent recognition while keeping the computational load of a single set of data within a reasonable range, avoiding the waste of computing power of the vehicle ECU caused by full-time acquisition; the multi-dimensional pedal behavior data acquired synchronously by multiple sensors can comprehensively capture the dynamic characteristics of the driver's braking actions. Compared with traditional single pedal displacement data, it can effectively distinguish between different behaviors such as emergency braking, normal braking, and accidental touch, improving the accuracy and robustness of the subsequent two-level intent recognition models from the data source.

[0018] Step S200: Preprocess the multi-dimensional pedal behavior data to obtain a first pedal behavior feature set, synchronize the first pedal behavior feature set to the first intention recognition model, and obtain the braking intention prediction result.

[0019] Furthermore, the multi-dimensional pedal behavior data is preprocessed to obtain a first pedal behavior feature set. The first pedal behavior feature set is then synchronized to a first intention recognition model to obtain a braking intention prediction result. The method includes: performing denoising and smoothing processing on the multi-dimensional pedal behavior data through a sliding window to construct a first time-domain feature subset; extracting the pedal displacement time series and performing variational mode decomposition, calculating the sample entropy of multiple components and extracting the maximum value to construct a first frequency-domain feature subset; concatenating the first time-domain feature subset and the first frequency-domain feature subset to form the first pedal behavior feature set; constructing a first intention recognition model, synchronizing the first pedal behavior feature set to the first intention recognition model for braking recognition, and generating a braking intention prediction result.

[0020] Specifically, the first time-domain feature subset refers to the set of statistical features of pedal motion extracted from the time dimension, used to intuitively characterize the amplitude, rate, force, and time-domain correlation characteristics of pedal action. Variational mode decomposition is an adaptive signal decomposition method that decomposes non-stationary pedal displacement sequences into multiple intrinsic mode components, avoiding mode aliasing and highlighting the frequency domain characteristics of the signal. Sample entropy is a nonlinear index used to measure the complexity and irregularity of the time series; its value reflects the abrupt changes in braking action. The first frequency-domain feature subset refers to the extreme values ​​of sample entropy based on the decomposed components of the displacement sequence, characterizing the frequency-domain nonlinear characteristics of pedal motion. The first intent recognition model is a support vector machine classification model optimized by the whale optimization algorithm; its lightweight structure ensures fast inference and is used for preliminary prediction of braking intent.

[0021] Specifically, after acquiring multi-dimensional pedal behavior data, the displacement, velocity, acceleration, and pedaling force data are first denoised and smoothed using a 5-point sliding window to eliminate abnormal jump points. Then, 12 types of features are extracted from the processed data: maximum pedal displacement, average pedal displacement, change in pedal displacement, maximum instantaneous pedal velocity, average instantaneous pedal velocity, maximum pedal acceleration, average pedal acceleration, peak pedaling force, average pedaling force, rate of change of pedaling force, duration of pedal action, and temporal correlation coefficient between pedal displacement and pedaling force. This constructs a first temporal feature subset. The pedal displacement time series is extracted and variational mode decomposition is performed to obtain multiple modal components. The sample entropy of each component is calculated, and the maximum value is extracted to form a first frequency domain feature subset. The first temporal feature subset and the first frequency domain feature subset are concatenated dimensionally to obtain a first pedal behavior feature set. Finally, this feature set is input into the trained first intent recognition model to complete the braking intent classification and recognition, and output the prediction results and corresponding confidence scores for emergency braking, normal braking, and no braking.

[0022] Preferably, variational mode decomposition achieves signal decomposition by constructing and solving a constrained variational problem. Specifically, the center frequency, Lagrange multipliers, and iteration count of each modal component are initialized. The alternating direction multiplier method iteratively updates each modal component and its center frequency, decomposing the original signal into several intrinsic mode functions with finite bandwidth and non-overlapping center frequencies. Decomposition stops when the iteration error is less than a preset threshold or when the maximum number of iterations is reached. This application targets a pedal displacement time series within a 200ms window, setting the number of decomposed modes K=4, the penalty factor α=2000, and the maximum number of iterations to 100. This results in four intrinsic mode components with sequentially increasing center frequencies, corresponding to the low-frequency stationary component and the high-frequency abrupt component of the pedal motion, effectively suppressing the mode aliasing problem of traditional empirical mode decomposition and providing a clean frequency domain feature foundation for subsequent sample entropy extraction.

[0023] The core of feature concatenation, which merges time-domain and frequency-domain feature vectors, is alignment along the same dimension and sequential concatenation. Specifically, the method involves: first, normalizing all feature values ​​of the first time-domain and first frequency-domain feature subsets, mapping them to the [0, 1] interval to eliminate dimensional differences between different feature dimensions and prevent excessively high weights of certain features from interfering with model recognition. The fixed concatenation order is "time-domain feature vector first, frequency-domain feature vector second," meaning all components of the time-domain features are arranged first, followed by all components of the frequency-domain features. The standardized time-domain feature vectors, such as [X1, X2, ..., X12], where X1 is the maximum pedal displacement, X2 is the average pedal displacement, and so on, are concatenated sequentially with the frequency-domain feature vectors, such as [Y1, Y2], where Y1 and Y2 are the maximum sample entropy values ​​of the VMD decomposition components, to form a unified feature vector, [X1, X2, ..., X12, Y1, Y2]. This vector represents the first pedal behavior feature set. It is important to note that when splicing, the sample timestamps of the two feature subsets must correspond completely to avoid feature distortion caused by cross-window feature splicing, and to ensure that the spliced ​​feature set can truly reflect the pedal behavior characteristics within the same time period.

[0024] This step improves the signal-to-noise ratio and reduces noise interference by using a sliding window for denoising. Multidimensional time-domain features comprehensively depict the temporal variation of pedal motion. Combined with frequency-domain features constructed from variational mode decomposition and sample entropy, the nonlinearity and abrupt changes of braking action can be accurately captured. Feature concatenation achieves the fusion of time-domain and frequency-domain features. With the optimized support vector machine model, while ensuring millisecond-level fast inference, the accuracy of the initial prediction of braking intention is significantly improved, providing a reliable basis for the pre-start of the electronic vacuum pump.

[0025] Furthermore, the process of constructing the first intent recognition model includes: retrieving historical driving scenario data; performing pedal behavior analysis based on the historical driving scenario data to extract multiple sets of intelligent pedal behavior data; identifying braking intent based on the multiple sets of intelligent pedal behavior data to generate multiple braking intent labels; associating and matching the multiple sets of intelligent pedal behavior data with the multiple braking intent labels to construct a training dataset; traversing the training dataset to extract features and obtain multiple pedal behavior feature vectors; constructing a support vector machine (SVM) with the multiple pedal behavior feature vectors as input and the braking intent labels as output, and using the whale optimization algorithm to globally optimize the SVM to obtain the optimal parameter combination; and performing classification training on the SVM based on the optimal parameter combination to construct the first intent recognition model.

[0026] Specifically, historical driving scenario data covers raw real-world driving data for different road conditions, vehicle models, and driver habits, including multi-dimensional time-series data such as pedal behavior, vehicle speed, braking deceleration, and vehicle status. Intelligent pedal behavior data refers to valid pedal behavior segments containing complete braking action cycles selected from historical data, excluding invalid data such as vehicle stalling or sensor malfunctions.

[0027] Braking Intent Labels: These labels are used to categorize the intentions behind pedal actions. This scheme sets three categories: no braking intent, normal braking intent, and emergency braking intent. Support Vector Machine (SVM): This is a binary classification algorithm based on statistical learning theory. It maps low-dimensional data to a high-dimensional space using kernel functions to achieve linear separability. This scheme extends it to a three-class classification model. Whale Optimization Algorithm: This algorithm simulates the global optimization behavior of whales. It finds the optimal solution through three stages: surrounding prey, bubble net attack, and searching for prey. It has a fast convergence speed and is not easily trapped in local optima.

[0028] The first intent recognition model employs a lightweight Gaussian radial basis function (RBF) kernel support vector machine (SVM) structure based on a one-to-one strategy, without complex hidden layers, designed for real-time inference in automotive embedded systems. The model input layer receives a 14-dimensional first pedal behavior feature vector, consisting of 12-dimensional time-domain features and 2-dimensional frequency-domain features. The kernel mapping layer maps low-dimensional linearly inseparable feature vectors to a high-dimensional feature space using a Gaussian radial basis function (RBF) kernel, achieving linear separability between different braking intents. The penalty coefficient C and width γ of the kernel function are globally optimized using a whale optimization algorithm, with typical optimal values ​​of C=12.5 and γ=0.32. The classification decision layer uses a one-to-one multi-class strategy, constructing three independent binary SVM sub-models to perform binary classification tasks for no braking / regular braking, no braking / emergency braking, and regular braking / emergency braking, respectively. The final predicted intent type is determined through a majority voting mechanism. The confidence output layer uses a Platt scaling algorithm to convert the decision function output of the SVM sub-models into probability values ​​in the 0-1 range, serving as the predicted confidence for the corresponding intent.

[0029] Specifically, historical driving scenario data from multiple scenarios collected from real vehicles were retrieved. Through data cleaning to remove outliers and invalid samples, multiple sets of intelligent pedal behavior data containing the complete 200ms process before and after braking were extracted. Annotators, combining the vehicle's actual braking state and dashcam video, labeled each set of data with braking intent, generating three categories of braking intent labels: no braking, normal braking, and emergency braking. Each set of intelligent pedal behavior data was associated and matched with its corresponding label, and divided into training, validation, and test sets in a 7:2:1 ratio to construct a complete training dataset. All samples in the training set were traversed to extract the corresponding labels for each set of samples. The pedal behavior feature vector is obtained; a support vector machine model based on Gaussian kernel function is constructed, and a three-class classification task is set with the feature vector as input and the braking intention label as output. The whale optimization algorithm is used to globally optimize the penalty coefficient C of SVM with a search range of [0.1, 100] and the width of Gaussian kernel function γ with a search range of [0.01, 10]. The population size is set to 30 and the maximum number of iterations is 50. The optimization stops when the classification accuracy of the validation set converges, and the optimal parameter combination is obtained. The support vector machine is trained based on this parameter combination. When the classification accuracy of the test set reaches more than 92%, the construction of the first intention recognition model is completed.

[0030] Preferably, the first intent recognition model training data annotation method adopts a dual-dimensional quantitative annotation method of objective braking results and subjective action intent. All annotation work is completed by more than 3 annotators with professional knowledge of automotive braking systems. The annotation results are cross-validated and reviewed by experts to ensure that the annotation accuracy is ≥98%. First, the original historical driving data is sliced. Taking the starting point of each pedal action as the center, complete data segments of 400ms are extracted before and after each pedal action, and matched with the 200ms input window during model inference to cover the entire action process. Invalid segments with sensor failure, missing data, or duration less than 100ms are removed to finally obtain the effective annotated sample set. The annotators first reviewed the dashcam video and vehicle CAN bus data for the corresponding time period to confirm the vehicle's driving environment and actual braking results. They then assigned a corresponding braking intent label to each data segment based on the aforementioned quantitative standards. All samples were initially annotated independently by two annotators; samples with consistent annotation results were directly entered into the database. Ambiguous samples with inconsistent annotation results were submitted to a third senior annotator for arbitration. A random sample of 10% of the annotated samples was then reviewed by experts; if the review accuracy rate was below 98%, all samples in that batch were re-annotated.

[0031] This step ensures the model's generalization ability through training with real-vehicle data in multiple scenarios. The whale optimization algorithm is used instead of the traditional grid search method, which effectively solves the problems of difficult SVM parameter tuning and easy getting trapped in local optima, and improves the classification accuracy of the first-level intent recognition. At the same time, the optimized SVM model has a lightweight structure and reduced inference latency, which fully meets the real-time requirements of the vehicle embedded system and provides a fast and reliable preliminary judgment basis for the early pre-start of the electronic vacuum pump.

[0032] Step S300: When the braking intention prediction result meets the preset pre-start conditions, the real-time vacuum degree of the electronic vacuum pump is obtained, and a first pre-start control command is generated according to the real-time vacuum degree and the braking intention prediction result to start the electronic vacuum pump and enter the first pre-start mode.

[0033] Furthermore, when the braking intention prediction result meets the preset pre-start conditions, the real-time vacuum level of the electronic vacuum pump is obtained, and a first pre-start control command is generated based on the real-time vacuum level and the braking intention prediction result to start the electronic vacuum pump into a first pre-start mode. The method includes: parsing the braking intention prediction result to obtain the predicted intention type and prediction confidence level, wherein the predicted intention type includes emergency braking intention and normal braking intention; introducing the real-time vacuum level data of the electronic vacuum pump, and when the predicted intention type is the emergency braking intention and the prediction confidence level is greater than or equal to a first prediction confidence level threshold, generating the first pre-start control command to start the electronic vacuum pump into an emergency pre-start mode; when the predicted intention type is the normal braking intention and the prediction confidence level is greater than or equal to a second prediction confidence level threshold, generating the first pre-start control command to start the electronic vacuum pump into a normal pre-start mode; wherein the first prediction confidence level threshold is less than the second prediction confidence level threshold.

[0034] Specifically, the preset pre-start condition refers to the combined triggering condition where the real-time vacuum level of the electronic vacuum pump is lower than the minimum safe operating threshold, and the type and confidence level of the braking intent prediction result match the corresponding start threshold. The predicted intent type refers to the driver intent classification result output by the first intent recognition model, divided into three categories: emergency braking intent, normal braking intent, and no braking intent. The prediction confidence level refers to the model's probability assessment of the correctness of the prediction result, ranging from 0 to 1; a higher value indicates a more reliable prediction result. The first / second prediction confidence threshold refers to the minimum confidence level requirement for triggering the corresponding pre-start mode, with typical values ​​of 80% for the first threshold and 85% for the second threshold. The real-time vacuum level refers to the absolute pressure value inside the vacuum tank (unit: kPa); a smaller value indicates a higher vacuum level and stronger braking assistance, with a minimum safe operating vacuum level of -60 kPa. The emergency pre-start mode refers to the rapid pressure-building mode where the electronic vacuum pump operates at 70% of its rated power, used to quickly establish vacuum assistance in emergency braking scenarios. The conventional pre-start mode refers to the low-speed pressure-building mode in which the electronic vacuum pump operates at 30% of its rated power, used to smoothly establish vacuum assistance in conventional braking scenarios.

[0035] Specifically, after receiving the braking intent prediction result output by the first intent recognition model, it performs structured analysis to extract the predicted intent type and the corresponding prediction confidence value; simultaneously, it acquires the current real-time vacuum level data through a high-precision vacuum sensor installed on the electronic vacuum pump's vacuum tank; it then sequentially checks whether the preset pre-start conditions are met and verifies whether the real-time vacuum level is below the minimum safe operating threshold of -60kPa. If the vacuum level is sufficient, it skips the pre-start process and returns to the data acquisition stage; if the vacuum level is insufficient, it continues to match the intent and confidence conditions; if the predicted intent type is an emergency braking intent... If the predicted confidence level is greater than or equal to the first predicted confidence threshold of 80%, an emergency pre-start control command is generated, which controls the electronic vacuum pump to start at 70% of its rated power and enter the emergency pre-start mode via a PWM signal. If the predicted intention type is a normal braking intention and the predicted confidence level is greater than or equal to the second predicted confidence threshold of 85%, a normal pre-start control command is generated, which controls the electronic vacuum pump to start at 30% of its rated power and enter the normal pre-start mode. If the predicted intention is no braking or the confidence level does not reach the corresponding threshold, no pre-start command is generated, and the process returns to step S100 to continue collecting pedal behavior data in a loop.

[0036] This step achieves precise and differentiated pre-start of the electronic vacuum pump through a triple control logic of pre-vacuum degree verification, intent type classification, and confidence degree differentiation threshold. The design of the first prediction confidence threshold being lower than the second threshold ensures priority response in emergency braking scenarios, tolerating an extremely low probability of false start rather than missing a start, thus minimizing the brake assist build-up time. In normal braking scenarios, a higher threshold is used to effectively filter out interference behaviors such as driver accidental touch or foot on the pedal, significantly reducing the false start rate. At the same time, dynamic judgment based on real-time vacuum degree avoids ineffective operation when the vacuum degree is sufficient. While ensuring the safety of vehicle braking, it reduces the average energy consumption of the electronic vacuum pump, and also reduces noise interference and component wear caused by frequent start-stop.

[0037] Step S400: Execute the first pre-start mode to collect data on the brake pedal in real time, obtain the second pedal behavior feature set, synchronize it to the second intention recognition model to confirm the braking intention, and obtain the braking intention confirmation result.

[0038] Furthermore, the method involves executing the first pre-start mode to collect data on the brake pedal in real time, obtaining a second pedal behavior feature set, and synchronizing this set to the second intent recognition model for brake intent confirmation, thereby obtaining a brake intent confirmation result. The method includes: continuously recording data in the first pre-start mode to obtain the start duration; dynamically adjusting the time window length according to the start duration and setting a target time window; supplementing the collection of pedal behavior data based on the target time window to obtain supplementary pedal behavior data; performing feature analysis based on the supplementary pedal behavior data to construct a second pedal behavior feature set; constructing a second intent recognition model based on an attention-enhanced convolutional neural network; synchronizing the second pedal behavior feature set to the second intent recognition model; and generating the brake intent confirmation result.

[0039] Specifically, the startup duration refers to the running time of the electronic vacuum pump from entering the first pre-start mode to the current moment, used to dynamically adjust the length of the subsequent data acquisition time window. The target time window refers to the final data acquisition window determined after dynamic adjustment. Supplementary pedal behavior data refers to the pedal behavior data additionally collected based on the target time window after the first pre-start mode is started, used to supplement the data information during the first stage of prediction and improve the accuracy of intent recognition. The second pedal behavior feature set refers to the feature set constructed based on the supplementary pedal behavior data. Compared with the first feature set, it contains more complete pedal action temporal information, with the same dimensions as the first feature set, namely 12-dimensional time domain and 2-dimensional frequency domain, but the feature values ​​cover a longer / more precise action process. Attention-enhanced convolutional neural network refers to a deep learning model that introduces a channel attention module on the basis of a traditional convolutional neural network. It can automatically assign higher weights to key features and suppress interference from irrelevant features, used for the second-level high-precision intent confirmation. The second intent recognition model is a high-precision classification model built based on the attention-enhanced convolutional neural network, prioritizing recognition accuracy, used for secondary confirmation of the first-level prediction results.

[0040] Specifically, after the electronic vacuum pump is activated and enters the first pre-start mode, the on-board timer starts recording the start duration. Simultaneously, the time window length is dynamically adjusted according to the type of the current first pre-start mode: for emergency pre-start mode, to ensure rapid response, the target time window is set to 100ms; for regular pre-start mode, to ensure data integrity, the target time window is set to 300ms. Based on this target time window, the brake pedal is supplemented with data collected by a pedal behavior sensor to obtain supplementary pedal behavior data covering the target time window. Using the same feature extraction method as in step S200, the supplementary pedal behavior data is denoised, smoothed, and subjected to time-domain and frequency-domain feature extraction and concatenation to construct a second pedal behavior feature set. This feature set is reshaped into a two-dimensional feature matrix of time step × feature dimension, and input into a pre-trained second intent recognition model based on an attention-enhanced convolutional neural network. The model extracts high-dimensional features through multiple convolutional layers, assigns higher weights to key features via a channel attention module, and completes inference through a fully connected classification layer, generating a confirmation result containing either brake intent confirmation or no brake intent and the corresponding confidence level.

[0041] The preferred second intent recognition model employs a lightweight convolutional neural network structure based on a channel attention mechanism. The model input is a two-dimensional feature matrix consisting of a time step and a feature dimension. Hierarchical feature extraction is performed through three convolutional layers, gradually mapping low-dimensional temporal features to a high-dimensional abstract feature map. Subsequently, a channel attention module is connected to compress the feature space of the high-dimensional feature map using global average pooling. Attention weights for each feature channel are generated through two fully connected layers and a sigmoid activation function. These weights, in the 0-1 range, are multiplied channel-by-channel with the original high-dimensional feature map, giving higher weights to key braking features such as peak pedal acceleration and displacement change rate, while suppressing irrelevant noise features. Finally, through two fully connected classification layers and a softmax activation function, the model outputs probability values ​​for two categories: confirmed braking intent and no braking intent. The category corresponding to the highest probability is taken as the final confirmation result, and its probability value represents the confirmation confidence level.

[0042] The training of the second intent recognition model employs a two-stage training approach, supplemented by basic data and error-prone samples, prioritizing recognition accuracy. The first stage involves basic training, extracting samples covering the complete braking process from historical driving scenario data. For emergency braking scenarios, samples are extracted 100ms after the start of the action, and for regular braking scenarios, 300ms. The same two-dimensional quantization annotation method as the first intent recognition model is used for annotation. A basic training dataset is constructed, divided into training, validation, and test sets in a 7:2:1 ratio. During training, the cross-entropy loss function and Adam optimizer are used, with an initial learning rate of 0.001, a batch size of 32, and a maximum of 100 iterations. Model performance is monitored using the validation set; training stops when the validation set loss does not decrease for 10 consecutive iterations to prevent overfitting. The second stage involves supplementing the training with error-prone samples. This stage involves collecting misclassified samples from the first intent recognition model, including missed emergency braking and misclassified regular braking, and adding them to the training set to fine-tune the model, further improving its ability to recognize complex scenes and error-prone samples. Finally, the trained model must achieve a binary classification accuracy of over 98% on the test set and a recall rate of 100% for emergency braking intents to ensure no missed cases.

[0043] This step achieves a balance between rapid confirmation in emergency scenarios and accurate recognition in normal scenarios by dynamically adjusting the time window length. The second pedal behavior feature set, constructed based on supplementary data, contains more complete pedal action information than the first-stage feature set, effectively avoiding misjudgments caused by incomplete data. The application of attention-enhanced convolutional neural networks can automatically focus on key features that distinguish braking intentions, improving the accuracy of second-level intention recognition. The cooperation of the two-level intention recognition architecture ensures millisecond-level response in emergency braking scenarios and significantly reduces the false start rate through high-precision secondary confirmation, providing a reliable basis for subsequent hierarchical control.

[0044] Furthermore, the second pedal behavior feature set is synchronized to the second intent recognition model to generate a braking intent confirmation result. The method includes: reshaping the second pedal behavior feature set into a two-dimensional feature matrix, wherein the behavior time step dimension and columns of the two-dimensional feature matrix are feature dimensions; inputting the two-dimensional feature matrix into the multi-layer convolutional layer of the second intent recognition model for feature extraction to obtain a high-dimensional feature map; inputting the high-dimensional feature map into the channel attention module of the second intent recognition model for pooling and compressing the feature space, and performing fully connected shared fusion based on the pooled compressed data to generate channel attention weights; multiplying the channel attention weights with the high-dimensional feature map channel by channel to obtain an attention-enhanced feature map; and inputting the attention-enhanced feature map into the fully connected classification layer of the second intent recognition model to generate the braking intent confirmation result.

[0045] Specifically, a two-dimensional feature matrix refers to a matrix formed by rearranging a one-dimensional temporal feature vector according to the time step and feature dimension. The rows represent the time step dimension, corresponding to sampling points, and the columns represent the feature dimension, corresponding to 14-dimensional stepper behavior features. This is the standard input format for convolutional neural networks. A high-dimensional feature map is a multi-channel feature matrix output by the convolutional layer. Each channel corresponds to an abstract feature, and the third convolutional layer outputs a 128-channel high-dimensional feature map. The channel attention module is a feature enhancement mechanism that generates weights by learning the importance of features in each channel, automatically amplifying key features and suppressing irrelevant noise features. Global average pooling compresses the two-dimensional feature map of each channel into a scalar value, preserving global feature information while significantly reducing computational cost. A fully connected classification layer is a network layer that maps high-dimensional features to a classification space, outputting probability values ​​for each category through a softmax activation function.

[0046] Specifically, after obtaining the second pedal behavior feature set, it is first reshaped into a two-dimensional feature matrix, which is 100 rows × 14 columns in the emergency pre-start mode, corresponding to a 100ms window and 1kHz. The sampling rate and 14-dimensional features are set to 300 rows × 14 columns in the standard pre-launch mode. This two-dimensional feature matrix is ​​input into the first convolutional layer of the second intent recognition model. Feature extraction is performed using 32 3×3 convolutional kernels with a stride of 1 and padding of 1, outputting a 100×14×32 feature map. This is then processed by a second layer with 64 3×3 convolutional kernels to obtain a 100×14×64 feature map. Finally, a third layer with 128 2×2 convolutional kernels and a stride of 2 extracts a 50×7×128 high-dimensional feature map. This high-dimensional feature map is then input into the channel attention module. Global average pooling is used to compress each 50×7 channel feature map into a single scalar, resulting in a 1×1×128 compressed feature vector. This compressed feature vector is then shared and fused through two fully connected layers. The first layer uses ReLU activation for 64 neurons, and the second layer uses Sigma activation for 128 neurons. The OID activation function generates 128 channel attention weights in the 0-1 range. For example, in an emergency braking scenario, the channel weight corresponding to the peak pedal acceleration is 0.95, the channel weight corresponding to the rate of change of pedal force is 0.92, and the channel weight corresponding to the average pedal displacement is only 0.3. Then, the feature map of each channel is multiplied by the corresponding attention weight channel by channel to obtain the attention-enhanced feature map. Finally, the attention-enhanced feature map is flattened into a one-dimensional vector and input into two fully connected classification layers. The first layer has 256 neurons using the ReLU activation function, and the second layer has 2 neurons using the softmax activation function. The output is the probability value of two classes: braking intention confirmation and no braking intention. For example, if the output is [0.98, 0.02], then the class corresponding to the maximum probability is taken as the final braking intention confirmation result, and its probability value of 0.98 is the confirmation confidence.

[0047] This step reshapes one-dimensional features into a two-dimensional matrix, fully utilizing the local feature extraction capabilities of convolutional neural networks to effectively capture the temporal correlation characteristics of pedal behavior. The introduction of the channel attention module enables the model to automatically focus on the most critical features for distinguishing braking intentions, suppressing the influence of irrelevant noise such as vehicle electromagnetic interference and driver foot vibration. Compared with traditional convolutional neural networks, this improves the binary classification accuracy. At the same time, the model adopts a lightweight design, with a single-sample inference latency stable within 10ms, fully meeting the real-time requirements of vehicles. This provides a high-precision judgment basis for the subsequent graded control of the electronic vacuum pump, avoiding energy consumption and noise problems caused by false starts, and ensuring zero missed judgments in emergency braking scenarios.

[0048] Step S500: When the braking intention confirmation result is braking intention confirmed, a second pre-start control command is generated to switch the electronic vacuum pump to the second pre-start mode. When the braking intention confirmation result is no braking intention, an exit control command is generated to stop the pre-start operation of the electronic vacuum pump.

[0049] Furthermore, when the braking intent confirmation result is confirmed, a second pre-start control command is generated to switch the electronic vacuum pump to the second pre-start mode. When the braking intent confirmation result is no braking intent, an exit control command is generated to stop the pre-start operation of the electronic vacuum pump. The method includes: performing confidence analysis based on the braking intent confirmation result to extract a braking intent confirmation confidence score; when the braking intent confirmation confidence score is greater than or equal to a first confirmation threshold, a second pre-start control command is generated to switch the electronic vacuum pump from the first pre-start mode to a full-power pre-start mode; when the braking intent confirmation confidence score is greater than or equal to the second confirmation threshold and less than the first confirmation threshold, the electronic vacuum pump is maintained in the first pre-start mode, and supplementary pedal behavior data is collected simultaneously for braking intent... Figure 2 The first confirmation is then confirmed; if the confidence score of the braking intention confirmation is less than the second confirmation threshold, the exit control command is generated to stop the pre-start operation of the electronic vacuum pump; wherein, the first confirmation threshold is greater than the second confirmation threshold.

[0050] Specifically, the braking intent confirmation confidence score refers to the 0-1 interval probability value output by the second intent recognition model, representing the reliability of the model's judgment of the existence of braking intent; the higher the value, the more accurate the judgment. The first confirmation threshold refers to the minimum confidence requirement for triggering full-power pre-start, typically 95%, which is the highest level of confirmation threshold. The second confirmation threshold refers to the minimum confidence requirement for maintaining pre-start and performing secondary confirmation, typically 85%; below this threshold, it is determined that there is no braking intent. The second pre-start mode is the full-power pre-start mode, a rapid pressure-building mode where the electronic vacuum pump operates at 100% rated power, capable of establishing a vacuum level sufficient for braking in the shortest possible time. Figure 2 The second confirmation is a loop mechanism that, for results with a confidence level in the ambiguity range, continues to collect pedal data and re-inputs it into the second intent recognition model for inference, avoiding the randomness of a single judgment. The exit control command is used to cut off the control signal supplying the electronic vacuum pump, immediately stopping the pre-start operation and avoiding unnecessary energy consumption and noise.

[0051] Specifically, after receiving the braking intent confirmation result output by the second intent recognition model, confidence analysis is performed to extract the braking intent confirmation confidence score in the 0-1 interval. This score is compared with the preset two-level confirmation thresholds, and a three-level hierarchical control strategy is executed: when the confirmation confidence score is greater than or equal to the first confirmation threshold of 95%, regardless of whether the current mode is emergency pre-start mode (70% power) or normal pre-start mode (30% power), a second pre-start control command is immediately generated. The duty cycle of the electronic vacuum pump is adjusted to 100% via the PWM signal, switching to the full-power pre-start mode to quickly establish vacuum assistance; when the confirmation confidence score is greater than or equal to the second confirmation threshold of 85% but less than 95%, the intent is determined to be ambiguous. The mode is neither switched nor stopped, and the current first pre-start mode continues to operate. At the same time, supplementary pedal behavior data for the next 100ms is collected synchronously and input into the second intent recognition model again for braking intent confirmation. Figure 2 If the confidence level remains in the fuzzy range after the second confirmation, the second confirmation process will be repeated a maximum of three times. Exceeding this number will automatically generate an exit command. When the confirmation confidence score is less than the 85% second confirmation threshold, an exit control command will be immediately generated, cutting off the power supply circuit to the electronic vacuum pump and stopping the pre-start operation. For example, if the confirmation confidence level is 98%, it will directly switch from 70% emergency pre-start to 100% full power; if the confidence level is 90%, it will maintain 30% normal pre-start and perform a second confirmation; if the confidence level is 72%, all pre-start operations will be immediately stopped.

[0052] Preferably, the first and second confirmation thresholds are jointly determined through ROC curve analysis combined with business objective constraints and real-vehicle verification iterations. The core is to achieve the optimal balance between braking safety and energy consumption noise. On a test set of the second intent recognition model containing a large number of samples, all possible thresholds in the 0-1 interval are traversed, and the true positive rate corresponding to each threshold is calculated, i.e., the correct recognition rate of braking intent and the false positive rate, i.e., the proportion of non-braking being mistaken for braking, and an ROC curve is plotted. Then, constraints are set in combination with business priorities: the first confirmation threshold must meet the requirement that the false positive rate is ≤1% to avoid significant noise and energy consumption caused by full-power false starts. Under this constraint, the threshold corresponding to the maximum true positive rate is taken, and it is finally determined to be 95%; the second confirmation threshold must meet the requirement that the true positive rate is ≥99.9% to absolutely eliminate missed braking intents. Under this constraint, the threshold corresponding to the minimum false positive rate is taken, and it is finally determined to be 85%. Finally, through 100,000 kilometers of real-vehicle road testing, the false start rate and missed detection rate are fine-tuned within ±2% for different scenarios such as urban congestion, highway cruising, and mountain roads to ensure the robustness of the thresholds in all scenarios.

[0053] This step, through a three-level hierarchical control logic of full-power switching, secondary confirmation, and immediate exit, overcomes the limitations of the traditional binary control scheme of "on or off" in pre-start solutions, achieving precise dynamic adjustment of the electronic vacuum pump power. The design of the first confirmation threshold being higher than the second confirmation threshold ensures that only braking intentions with extremely high confidence will trigger full-power operation, minimizing energy consumption and noise caused by false starts. The secondary confirmation mechanism effectively solves the ambiguity problem of single intention recognition, avoiding both braking delays caused by false exits and energy waste caused by continuous operation due to false starts. At the same time, hierarchical control reduces the time of full-power operation of the electronic vacuum pump, extending the service life of the motor and pump body, and significantly improving driving comfort and system reliability.

[0054] In summary, the electronic vacuum pump pre-start control method based on pedal behavior recognition provided in this application has the following technical effects: 1. Employing a logic of prediction, pre-start, confirmation, and hierarchical control, combined with a two-level intent recognition model and the real-time vacuum level of the electronic vacuum pump, the vacuum pump is started in stages and with varying power. This solves the braking delay problem of traditional pedal-assisted start, balancing response speed and recognition accuracy, shortening the braking assist delay, and significantly reducing energy consumption and noise from false starts.

[0055] 2. Denoising and smoothing of multi-dimensional pedal data is performed using a sliding window, extracting temporal statistical features, variational mode decomposition, and sample entropy frequency domain features, which are then concatenated. This effectively eliminates sensor noise and vehicle electromagnetic interference, comprehensively captures the nonlinear and abrupt characteristics of pedal behavior, provides high-quality feature input for first-level rapid intent recognition, and improves prediction robustness.

[0056] 3. The time window length is dynamically adjusted based on the duration of the first pre-start mode, and supplementary pedal behavior data is collected to construct a second feature set. This achieves a balance between rapid confirmation in emergency scenarios and accurate recognition in regular scenarios, supplementing more complete action timing information and effectively avoiding misjudgments and omissions caused by incomplete data.

[0057] Example 2 is based on the same inventive concept as the electronic vacuum pump pre-start control method based on pedal behavior recognition in the previous examples, such as... Figure 2 As shown in the figure, this application provides an electronic vacuum pump pre-start control system based on pedal behavior recognition. The system includes: The system comprises: a behavior data acquisition module 11, used to acquire multi-dimensional pedal behavior data of the brake pedal in real time through a pedal behavior sensor within a preset time window; a prediction result acquisition module 12, used to preprocess the multi-dimensional pedal behavior data to obtain a first pedal behavior feature set, synchronize the first pedal behavior feature set to a first intention recognition model, and obtain a braking intention prediction result; a control command generation module 13, used to acquire the real-time vacuum degree of the electronic vacuum pump when the braking intention prediction result meets the preset pre-start conditions, generate a first pre-start control command based on the real-time vacuum degree and the braking intention prediction result, and start the electronic vacuum pump to enter the first pre-start mode; a confirmation result acquisition module 14, used to perform real-time acquisition of the brake pedal in the first pre-start mode, acquire a second pedal behavior feature set, synchronize it to a second intention recognition model for braking intention confirmation, and obtain a braking intention confirmation result; and a braking intention judgment module 15, used to generate a second pre-start control command when the braking intention confirmation result is braking intention confirmation, switch the electronic vacuum pump to the second pre-start mode, and generate an exit control command when the braking intention confirmation result is no braking intention, stopping the pre-start operation of the electronic vacuum pump.

[0058] Furthermore, the prediction result acquisition module 12 is also used to perform the following steps: denoising and smoothing the multi-dimensional pedal behavior data through a sliding window to construct a first time-domain feature subset; extracting the pedal displacement time series for variational mode decomposition, calculating the sample entropy of multiple components for maximum extraction, and constructing a first frequency-domain feature subset; concatenating the first time-domain feature subset and the first frequency-domain feature subset to form the first pedal behavior feature set; constructing a first intent recognition model, synchronizing the first pedal behavior feature set to the first intent recognition model for braking recognition, and generating a braking intent prediction result.

[0059] Furthermore, the prediction result acquisition module 12 is also used to perform the following steps: retrieve historical driving scenario data, perform pedal behavior analysis based on the historical driving scenario data, and extract multiple sets of intelligent pedal behavior data; identify braking intentions based on the multiple sets of intelligent pedal behavior data to generate multiple braking intention labels; associate and match the multiple sets of intelligent pedal behavior data with the multiple braking intention labels to construct a training dataset; traverse the training dataset to extract features and obtain multiple pedal behavior feature vectors; construct a support vector machine with the multiple pedal behavior feature vectors as input and the braking intention labels as output, and perform global optimization on the support vector machine using the whale optimization algorithm to obtain the optimal parameter combination; perform classification training on the support vector machine based on the optimal parameter combination to construct the first intention recognition model.

[0060] Furthermore, the control command generation module 13 is also used to perform the following steps: parse the braking intention prediction result to obtain the predicted intention type and prediction confidence level, wherein the predicted intention type includes emergency braking intention and normal braking intention; introduce the real-time vacuum level data of the electronic vacuum pump, and when the predicted intention type is the emergency braking intention and the prediction confidence level is greater than or equal to the first prediction confidence level threshold, generate the first pre-start control command to start the electronic vacuum pump into the emergency pre-start mode; when the predicted intention type is the normal braking intention and the prediction confidence level is greater than or equal to the second prediction confidence level threshold, generate the first pre-start control command to start the electronic vacuum pump into the normal pre-start mode; wherein, the first prediction confidence level threshold is less than the second prediction confidence level threshold.

[0061] Furthermore, the confirmation result acquisition module 14 is also used to perform the following steps: continuously record the first pre-start mode to obtain the start duration; dynamically adjust the time window length according to the start duration and set a target time window; supplement the pedal behavior based on the target time window to obtain supplementary pedal behavior data; perform feature analysis based on the supplementary pedal behavior data to construct a second pedal behavior feature set; construct a second intent recognition model based on an attention-enhanced convolutional neural network, synchronize the second pedal behavior feature set to the second intent recognition model, and generate the braking intent confirmation result.

[0062] Furthermore, the confirmation result acquisition module 14 is also used to perform the following steps: reshaping the second pedal behavior feature set into a two-dimensional feature matrix, wherein the behavior time step dimension and columns of the two-dimensional feature matrix are feature dimensions; inputting the two-dimensional feature matrix into the multi-layer convolutional layer of the second intent recognition model for feature extraction to obtain a high-dimensional feature map; inputting the high-dimensional feature map into the channel attention module of the second intent recognition model for pooling and compressing the feature space, and performing fully connected shared fusion based on the pooled compressed data to generate channel attention weights; multiplying the channel attention weights with the high-dimensional feature map channel by channel to obtain an attention-enhanced feature map; inputting the attention-enhanced feature map into the fully connected classification layer of the second intent recognition model to generate the braking intent confirmation result.

[0063] Furthermore, the braking intention determination module 15 is also used to perform the following steps: performing confidence analysis based on the braking intention confirmation result and extracting a braking intention confirmation confidence score; when the braking intention confirmation confidence score is greater than or equal to a first confirmation threshold, generating a second pre-start control command to switch the electronic vacuum pump from the first pre-start mode to a full-power pre-start mode; when the braking intention confirmation confidence score is greater than or equal to the second confirmation threshold and less than the first confirmation threshold, maintaining the electronic vacuum pump in the first pre-start mode and simultaneously collecting supplementary pedal behavior data for braking intention determination. Figure 2 The first confirmation is then confirmed; if the confidence score of the braking intention confirmation is less than the second confirmation threshold, the exit control command is generated to stop the pre-start operation of the electronic vacuum pump; wherein, the first confirmation threshold is greater than the second confirmation threshold.

[0064] 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 pre-start control method for an electronic vacuum pump based on pedal behavior recognition, characterized in that, The method includes: Multi-dimensional pedal behavior data of the brake pedal are collected in real time through a pedal behavior sensor within a preset time window. The multi-dimensional pedal behavior data is preprocessed to obtain a first pedal behavior feature set, and the first pedal behavior feature set is synchronized to a first intention recognition model to obtain a braking intention prediction result. When the braking intention prediction result meets the preset pre-start condition, the real-time vacuum degree of the electronic vacuum pump is obtained, and a first pre-start control command is generated based on the real-time vacuum degree and the braking intention prediction result to start the electronic vacuum pump and enter the first pre-start mode. The first pre-start mode is executed to collect data on the brake pedal in real time, and the second pedal behavior feature set is obtained and synchronized to the second intention recognition model to confirm the braking intention and obtain the braking intention confirmation result. If the braking intent confirmation result is "braking intent confirmed", a second pre-start control command is generated to switch the electronic vacuum pump to the second pre-start mode. If the braking intent confirmation result is "no braking intent", an exit control command is generated to stop the pre-start operation of the electronic vacuum pump.

2. The electronic vacuum pump pre-start control method based on pedal behavior recognition as described in claim 1, characterized in that, The method includes preprocessing the multi-dimensional pedal behavior data to obtain a first pedal behavior feature set, synchronizing the first pedal behavior feature set to a first intention recognition model, and obtaining a braking intention prediction result. The multidimensional pedal behavior data is denoised and smoothed using a sliding window to construct a first temporal feature subset. Extract the pedal displacement time series and perform variational mode decomposition. Calculate the sample entropy of multiple components and extract the maximum value to construct the first frequency domain feature subset. The first time-domain feature subset and the first frequency-domain feature subset are concatenated to form the first pedal behavior feature set; A first intent recognition model is constructed, and the first pedal behavior feature set is synchronized to the first intent recognition model for braking recognition, generating a braking intent prediction result.

3. The electronic vacuum pump pre-start control method based on pedal behavior recognition as described in claim 2, characterized in that, The process of building a first intent recognition model includes the following methods: Retrieve historical driving scenario data, perform pedal behavior analysis based on the historical driving scenario data, and extract multiple sets of intelligent pedal behavior data; Based on the multiple sets of intelligent pedal behavior data, braking intention is identified, and multiple braking intention tags are generated. The multiple sets of intelligent pedal behavior data are associated and matched with the multiple braking intention labels to construct a training dataset; The training dataset is traversed to extract features and obtain multiple pedal behavior feature vectors. A support vector machine is constructed with the multiple pedal behavior feature vectors as input and the braking intention label as output. The optimal parameter combination is obtained by global optimization of the support vector machine through the whale optimization algorithm. The support vector machine is trained for classification based on the optimal parameter combination to construct the first intent recognition model.

4. The electronic vacuum pump pre-start control method based on pedal behavior recognition as described in claim 1, characterized in that, When the braking intent prediction result meets the preset pre-start condition, the real-time vacuum level of the electronic vacuum pump is obtained, and a first pre-start control command is generated based on the real-time vacuum level and the braking intent prediction result to start the electronic vacuum pump and enter the first pre-start mode. The method includes: The braking intention prediction results are analyzed to obtain the predicted intention type and prediction confidence level. The predicted intention type includes emergency braking intention and normal braking intention. Real-time vacuum data of the electronic vacuum pump is introduced. When the predicted intention type is the emergency braking intention and the prediction confidence level is greater than or equal to the first prediction confidence level threshold, the first pre-start control command is generated to start the electronic vacuum pump and enter the emergency pre-start mode. When the predicted intention type is the conventional braking intention and the predicted confidence level is greater than or equal to the second predicted confidence threshold, a first pre-start control command is generated to start the electronic vacuum pump and enter the conventional pre-start mode. Wherein, the first prediction confidence threshold is less than the second prediction confidence threshold.

5. The electronic vacuum pump pre-start control method based on pedal behavior recognition as described in claim 1, characterized in that, The method involves executing the first pre-start mode to collect real-time data on the brake pedal, obtaining a second pedal behavior feature set, synchronizing it to the second intent recognition model for brake intent confirmation, and obtaining a brake intent confirmation result. The method includes: The first pre-boot mode is executed to continuously record and obtain the boot duration; The time window length is dynamically adjusted according to the startup duration, and a target time window is set. Based on the target time window, supplementary pedal behavior data is collected to obtain supplementary pedal behavior data. Based on the supplementary pedal behavior data, feature analysis is performed to construct a second pedal behavior feature set; A second intent recognition model is constructed based on an attention-enhanced convolutional neural network. The second pedal behavior feature set is synchronized to the second intent recognition model to generate the braking intent confirmation result.

6. The electronic vacuum pump pre-start control method based on pedal behavior recognition as described in claim 5, characterized in that, The method involves synchronizing the second pedal behavior feature set to the second intent recognition model to generate a braking intent confirmation result, including: The second pedal behavior feature set is reshaped into a two-dimensional feature matrix, wherein the behavior time step dimension and columns of the two-dimensional feature matrix are feature dimensions. The two-dimensional feature matrix is ​​input into the multi-layer convolutional layer of the second intent recognition model for feature extraction to obtain a high-dimensional feature map; The high-dimensional feature map is input into the channel attention module of the second intent recognition model to pool and compress the feature space. Based on the pooled compressed data, a fully connected shared fusion is performed to generate channel attention weights. The channel attention weights are multiplied channel by channel with the high-dimensional feature map to obtain the attention-enhanced feature map; The attention-enhanced feature map is input into the fully connected classification layer of the second intent recognition model to generate the braking intent confirmation result.

7. The electronic vacuum pump pre-start control method based on pedal behavior recognition as described in claim 1, characterized in that, When the braking intent confirmation result is "braking intent confirmed," a second pre-start control command is generated to switch the electronic vacuum pump to the second pre-start mode. When the braking intent confirmation result is "no braking intent," an exit control command is generated to stop the pre-start operation of the electronic vacuum pump. The method includes: Based on the braking intent confirmation result, a confidence analysis is performed to extract the braking intent confirmation confidence score; When the confidence score of the braking intention confirmation is greater than or equal to the first confirmation threshold, a second pre-start control command is generated to switch the electronic vacuum pump from the first pre-start mode to the full-power pre-start mode. When the confidence score of the braking intention confirmation is greater than or equal to the second confirmation threshold and less than the first confirmation threshold, the electronic vacuum pump is maintained in the first pre-start mode, and supplementary pedal behavior data is collected simultaneously for secondary confirmation of braking intention. When the confidence score of the braking intention confirmation is less than the second confirmation threshold, the exit control command is generated to stop the pre-start operation of the electronic vacuum pump; Wherein, the first confirmation threshold is greater than the second confirmation threshold.

8. An electronic vacuum pump pre-start control system based on pedal behavior recognition, characterized in that, The system is used to execute the electronic vacuum pump pre-start control method based on pedal behavior recognition as described in any one of claims 1 to 7, the system comprising: The behavior data acquisition module is used to collect multi-dimensional pedal behavior data of the brake pedal in real time through the pedal behavior sensor within a preset time window. The prediction result acquisition module is used to preprocess the multi-dimensional pedal behavior data to obtain a first pedal behavior feature set, synchronize the first pedal behavior feature set to the first intention recognition model, and obtain a braking intention prediction result. The control command generation module is used to obtain the real-time vacuum level of the electronic vacuum pump when the braking intention prediction result meets the preset pre-start conditions, generate a first pre-start control command based on the real-time vacuum level and the braking intention prediction result, and start the electronic vacuum pump to enter the first pre-start mode. The confirmation result acquisition module is used to perform real-time acquisition of the brake pedal in the first pre-start mode, obtain the second pedal behavior feature set, synchronize it to the second intention recognition model for brake intention confirmation, and obtain the brake intention confirmation result; the brake intention judgment module is used to generate a second pre-start control command when the brake intention confirmation result is brake intention confirmation, switch the electronic vacuum pump to the second pre-start mode, and generate an exit control command when the brake intention confirmation result is no brake intention, stop the pre-start operation of the electronic vacuum pump.

Images