Adaptive braking kinetic energy recovery control method and system based on multi-sensor fusion

Abstract

The present application relates to the technical field of automobile brake control, in particular, the present application relates to the adaptive braking kinetic energy recovery control method and system based on multi-sensor fusion, the present application collects tire wear and road feature data through multi-source perception fusion unit, intelligent decision calculation unit generates tire wear coefficient, road friction coefficient, the nonlinear correlation of both is integrated through rule and data fusion algorithm, output comprehensive friction coefficient, adaptive control execution unit dynamically adjusts braking kinetic energy recovery strength and response time accordingly, the system contains multi-source perception fusion unit, intelligent decision calculation unit, adaptive control execution unit, closed loop feedback calibration unit, the closed loop unit cross-verification correction data deviation through laser radar and motor torque inversion model, the present application solves the problem of poor working condition adaptation, unreliable data, improves kinetic energy recovery efficiency and brake safety, driving smoothness.

Description

Technical Field

[0001] This invention relates to the field of automotive braking control technology, and more specifically, to an adaptive braking kinetic energy recovery control method and system based on multi-sensor fusion. Background Technology

[0002] Automotive braking control technology is an important technology. In the context of the current pursuit of improved driving range and braking safety in new energy vehicles, this technology is a key support for balancing kinetic energy recovery efficiency and braking safety. It can not only extend the vehicle's range by accurately recovering kinetic energy during braking and converting it into electrical energy, but also dynamically adjust the recovery force and response time according to tire wear and road friction characteristics, avoiding problems such as wheel lock-up and braking jerking caused by fixed recovery parameters. At the same time, it is adaptable to different road surfaces such as dry, snowy, and wet, as well as different tire conditions such as new and worn tires, promoting the upgrade of new energy vehicle braking systems from fixed parameter control to scenario-adaptive control, and is widely applicable to various new energy passenger vehicles and commercial vehicles. Existing regenerative braking control technologies for new energy vehicles face core problems in practical applications, including poor adaptability to operating conditions, low data reliability, and insufficient control precision. Traditional systems often rely on single sensor data or use fixed recovery parameters, failing to fully integrate the dynamic correlation between tire wear and road friction. Tire wear continuously decreases with changes in mileage, load, temperature, and pressure. Tires with different wear levels exhibit significant differences in friction capacity on the same road surface. Traditional systems lack a coupled analysis mechanism between the wear coefficient and the road friction coefficient, setting recovery parameters only based on road type, ignoring the aggravated friction decay effect of high-wear tires on low-friction roads. Furthermore, sensor data is susceptible to environmental interference. Issues such as camera deviations in identifying road texture in rainy weather and signal drift from tire pressure and temperature sensors at low temperatures can lead to errors in road type judgment and tire condition perception. Moreover, traditional systems lack data uncertainty verification mechanisms and fail to correct deviations through multi-sensor cross-validation. These problems create a chain reaction, firstly affecting wear and... The lack of coupling relationship with the road surface leads to inaccurate calculation of the comprehensive friction coefficient, resulting in improper setting of the recovery force. High-wear tires are set to a high recovery force based on the parameters of new tires on icy and snowy roads, which can easily cause wheel lock-up. Conversely, new tires are set to a low recovery force based on conservative parameters on dry roads, wasting kinetic energy. Secondly, the lack of verification of sensor data errors can lead to misjudgment of road type, such as misjudging wet asphalt roads as dry roads. The excessively short recovery response time can cause braking shock. Finally, insufficient control precision will increase the mechanical braking load, shortening the life of the braking system in the long run, while reducing driving smoothness and affecting the user experience. Ultimately, the existing technology cannot meet the comprehensive requirements of new energy vehicles for efficient recovery, safe braking, and smooth driving under complex operating conditions. There is an urgent need for an adaptive braking kinetic energy recovery control method and system that can integrate multi-sensor data, process the nonlinear correlation between wear and road surface, and verify the reliability of data. To solve this technical problem, we provide an adaptive braking kinetic energy recovery control method and system based on multi-sensor fusion. Summary of the Invention

[0003] The purpose of this invention is to provide an adaptive braking kinetic energy recovery control method and system based on multi-sensor fusion to solve the problems mentioned in the background art.

[0004] 1. Since the comprehensive friction coefficient is inaccurate due to the lack of coupling analysis of tire wear and road friction, this case uses a multi-sensor fusion processing unit to adopt rule and data fusion algorithms to integrate the two nonlinear correlations to calculate the comprehensive friction coefficient, which can accurately adapt to the working conditions and avoid improper recovery force.

[0005] 2. Since sensor data is susceptible to interference and lacks verification, leading to misjudgments of road conditions, this case uses a closed-loop feedback calibration unit to call the lidar and motor torque fluctuation inversion model for cross-validation to correct data deviations, thereby improving data reliability and avoiding braking shocks.

[0006] To achieve the above objectives, one objective of this invention is to provide an adaptive braking kinetic energy recovery control method based on multi-sensor fusion, comprising the following steps: The tire wear coefficient is obtained in real time through the vehicle mileage sensor. This coefficient is calculated based on the vehicle's cumulative mileage and a preset wear model. Meanwhile, the road surface type detection unit identifies the type of the current driving road surface and generates the corresponding road surface friction coefficient. The tire wear coefficient and the road surface friction coefficient are then input into the multi-sensor fusion processing unit. This unit uses a rule-based and data-based fusion algorithm to dynamically calculate the comprehensive friction coefficient between the tire and the ground. The rule-based and data-based fusion algorithm integrates the nonlinear correlation between the tire wear coefficient and the road surface type. Based on the comprehensive friction coefficient, optimized control parameters are generated by the adaptive control unit to dynamically adjust the braking energy recovery force and response time, so as to achieve adaptive braking energy recovery control for different wear conditions and road surface conditions.

[0007] The second objective of this invention is to provide a system for implementing an adaptive braking kinetic energy recovery control method based on multi-sensor fusion, including any one of the above-mentioned features, comprising: The multi-source sensing fusion unit integrates a vehicle mileage sensor, a tire pressure and temperature sensor, an on-board gravity sensor, a multi-source heterogeneous road surface detection sensor array, and a tire tread depth ultrasonic detection module to collect multi-dimensional data related to tire wear and multi-modal feature data of the road surface in real time. The intelligent decision computing unit incorporates a dynamic wear assessment module, a road surface-friction relationship knowledge graph engine, a rule and data fusion algorithm processor, and an uncertainty quantification correction module. It receives output data from the multi-source perception fusion unit, generates the tire wear coefficient through the dynamic wear assessment module, generates the road surface friction coefficient through the road surface classification and friction mapping module, and integrates the nonlinear correlation between the wear coefficient and the road surface friction coefficient based on the fusion algorithm processor to output the comprehensive friction coefficient between the tire and the ground.

[0008] The adaptive control execution unit includes a two-layer optimization controller and a safety protection logic module. It receives the comprehensive friction coefficient output by the intelligent decision calculation unit, generates a basic recovery parameter set through the Pareto front strategy generator, combines the dynamic compensation amount with the model predictive controller, dynamically adjusts the kinetic energy recovery force gradient and response time constant using the sliding mode variable structure algorithm, and suppresses the mechanical braking impact during frictional abrupt changes through the gradual migration strategy.

[0009] The closed-loop feedback calibration unit connects the motor torque fluctuation inversion module and the lidar imprint scanner to capture tire ground contact pattern and actual friction boundary data in real time. The data is then fed back to the uncertainty quantification module in the intelligent decision calculation unit and the safety protection logic module in the adaptive control execution unit to drive multi-sensor cross-validation and online model optimization.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By collecting tire wear-related data and road feature data through a multi-source sensing fusion unit, the intelligent decision-making calculation unit adopts a rule-and-data collaborative fusion algorithm to integrate the nonlinear correlation between the two and construct a multi-dimensional response surface of wear-road coupling effect. The comprehensive friction coefficient is dynamically calculated to achieve the technical effect of accurately quantifying the tire-road coupling friction characteristics. This solves the problem of inaccurate comprehensive friction coefficient caused by the lack of coupling analysis of the relationship between the two in traditional technology. It has the advantages of avoiding improper recovery force and balancing braking safety and kinetic energy recovery efficiency.

[0011] 2. By embedding an uncertainty quantification module into the intelligent decision-making calculation unit, the confidence entropy value of sensor data is calculated in real time. When the bandwidth exceeds the limit, the closed-loop feedback calibration unit calls the lidar to scan the tire ground contact mark. Combined with the motor torque fluctuation inversion model, the actual friction boundary is calculated and the fusion result is corrected. This achieves the technical effect of improving data reliability and solves the problem of unreliable data caused by sensor interference and misjudgment of road surface. It has the advantages of enhancing the robustness of the comprehensive friction coefficient and avoiding braking shock.

[0012] 3. A two-layer optimization architecture is constructed through an adaptive control execution unit. The upper layer generates basic recovery parameters by using Pareto curves, while the lower layer uses a model predictive controller to superimpose dynamic compensation. The sliding mode variable structure algorithm adjusts the recovery force gradient and response time. When friction changes abruptly, a gradual migration strategy is initiated to achieve the technical effect of dynamically adapting to the working conditions. This solves the problem of braking jerking caused by low precision in traditional control. It has the advantages of improving driving smoothness and reducing mechanical braking load to extend its life. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the overall workflow of the present invention; Figure 2 This is a schematic diagram of the overall structure of the present invention; The meanings of the labels in the diagram are as follows: 1. Multi-source sensing fusion unit; 2. Intelligent decision computing unit; 3. Adaptive control execution unit; Closed-loop feedback calibration unit. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] Please see Figure 1 As shown, one of the objectives of this embodiment is to provide an adaptive braking kinetic energy recovery control method based on multi-sensor fusion, including the following steps: The tire wear coefficient is obtained in real time through the vehicle mileage sensor. This coefficient is calculated based on the vehicle's cumulative mileage and a preset wear model. Meanwhile, the road surface type detection unit identifies the type of the current driving road surface and generates the corresponding road surface friction coefficient. The tire wear coefficient and the road surface friction coefficient are input into the multi-sensor fusion processing unit. This unit uses a rule-based and data-based fusion algorithm to dynamically calculate the comprehensive friction coefficient between the tire and the ground. The rule-based and data-based fusion algorithm integrates the nonlinear correlation between the tire wear coefficient and the road surface type. Based on the comprehensive friction coefficient, the adaptive control unit generates optimized control parameters and dynamically adjusts the braking energy recovery force and response time to achieve adaptive braking energy recovery control for different wear conditions and road surface conditions.

[0016] The tire wear coefficient is obtained in real time using the vehicle's mileage sensor. The specific steps are as follows: A multi-dimensional data fusion mechanism is established to integrate the vehicle's cumulative mileage, real-time load distribution data collected by the vehicle's gravity sensor, and tire pressure and temperature change sequences from the tire pressure and temperature sensor. These data are then input into the dynamic wear assessment module. This module uses a pre-trained wear feature extraction network to analyze the historical attenuation patterns of tire contact deformation under different load and temperature / pressure conditions, generates a dynamic correction factor, and superimposes the dynamic correction factor with the baseline wear curve to output the initial tire wear coefficient.

[0017] The specific steps for constructing a pre-defined wear model are as follows: The wear evolution modeling framework based on reinforcement learning imports a material fatigue property database, a road adhesion coefficient historical database, and a user driving behavior profile database provided by the tire manufacturer during the initialization phase. Through a convolutional recurrent neural network architecture, spatiotemporal feature correlation analysis is performed on the three databases to construct a three-dimensional mapping relationship of wear degradation with mileage as the time axis and load-temperature-pressure as the environmental axis as the preset wear model. When the preset wear model is running online, it is combined with real-time feedback of tread depth ultrasonic detection data to drive the adversarial generative network to dynamically optimize the model weight coefficients, which is used to optimize the preset wear model.

[0018] Further explanation is needed regarding adaptive regenerative braking control based on multi-sensor fusion. The frictional characteristics between the tire and the road surface are the core factors determining the recovery force and response time. The tire wear coefficient directly reflects the degree of frictional attenuation; the difference in friction coefficient between a new tire and a severely worn tire can exceed 30%. Relying solely on fixed parameters can easily lead to inaccurate recovery force. Therefore, it is necessary to accurately obtain the tire wear coefficient using a vehicle mileage sensor combined with multi-dimensional data. The specific implementation method is as follows: To obtain the tire wear coefficient in real time using the vehicle mileage sensor, a multi-dimensional data fusion mechanism must first be established. This mechanism integrates relevant data from different sources and types of tires, eliminating the bias of single data sources. Developed by the automotive manufacturer's electronic control system team based on the Kalman filter algorithm, the core of this system is to transform scattered sensor data into a unified analytical data source. This avoids miscalculations of wear coefficients caused by focusing solely on mileage while neglecting load, temperature, and pressure. This mechanism integrates three key types of data: The first is the vehicle's cumulative mileage, provided by mileage sensors installed at the transmission or wheels, directly reflecting the tire's basic usage time and wear baseline. The second is real-time load distribution data collected by onboard gravity sensors. These high-precision sensors, installed on the vehicle chassis, detect the real-time load distribution of the front and rear axles and left and right wheels. Because greater load increases tire ground pressure and intensifies tire-to-ground friction, significantly accelerating wear, this data is crucial for correcting wear. The third category of key environmental parameters for the coefficient is the tire pressure and temperature change sequence from the tire pressure and temperature sensor. The tire pressure and temperature sensor is a TPMS sensor embedded inside the tire, and its output sequence represents the continuous changes in tire pressure and temperature during a certain period of driving. Low tire pressure leads to an increased contact area between the tire and the road surface, resulting in accelerated localized wear; high temperature accelerates tire rubber aging. Both must be included in wear analysis to avoid biased wear judgments due to static parameters. The above three types of data are aligned by timestamps to ensure that mileage, load, and temperature / pressure data at the same moment correspond before being input into the dynamic wear assessment module. The dynamic wear assessment module is a software module integrated into the intelligent decision-making computing unit 2, developed by the algorithm team, and includes a pre-trained wear feature extraction network. This network is a deep learning network trained using wear data from over 100,000 tires of different brands and models, including tread depth, contact patch deformation, mileage, load, and temperature / pressure. It can accurately identify key features affecting wear and avoid interference from irrelevant data. The core function of this module is to analyze the historical decay pattern of tire contact patch deformation under different load and temperature / pressure conditions. Contact patch deformation is the amount of deformation of the tire in contact with the ground under load. It is indirectly calculated using TPMS sensor data and suspension vibration sensor data. The contact patch deformation of a new tire is approximately 8-10mm, increasing to 12-15mm after severe wear. The historical decay pattern refers to retrieving contact patch deformation data from the tire over the past three months and analyzing its trend with increasing mileage. For every 1000 kilometers driven...The ground deformation increases by 0.2 mm. Based on these patterns, the network, combined with the differences between the current load, temperature, and pressure and standard operating conditions (2.5 bar tire pressure, 25°C temperature, half load), generates a dynamic correction factor. This factor is a coefficient that quantifies the difference in wear rate between the current and standard operating conditions. For example, if the current tire pressure is 2.3 bar and the tire is fully loaded, the correction factor is 1.2, meaning the wear rate under the current condition is 1.2 times that under the standard condition. If the tire pressure is 2.6 bar and the tire is unloaded, the correction factor is 0.8, indicating a slower wear rate. Finally, the dynamic correction factor is superimposed on the reference wear curve to output the initial tire wear coefficient. The reference wear curve is a standard wear curve provided by the tire manufacturer, based on... The wear test results for new tires under standard operating conditions are generated. When overlaying these results, the baseline wear coefficient corresponding to the current cumulative mileage on the reference curve must first be found. This coefficient is then multiplied by a dynamic correction factor to obtain the initial tire wear coefficient. This coefficient reflects both the baseline wear caused by mileage and corrects for the effects of current load, temperature, and pressure, providing a more accurate reflection of the actual tire wear state and offering precise input for subsequent calculations of the comprehensive friction coefficient. However, the calculation of the initial tire wear coefficient still relies on a static reference curve, making it difficult to adapt to the dynamic changes in different user driving habits and road conditions. Therefore, a preset wear model needs to be constructed to provide a more adaptable framework for real-time calculation of the wear coefficient. The specific implementation method is as follows: The pre-built wear model is based on a reinforcement learning-based wear evolution modeling framework. This framework involves the model interacting with driving conditions to continuously learn wear patterns and optimize the model. Built by the algorithm team using the DQN (Deep Q-Network) reinforcement learning algorithm, its core objective is to enable the model to dynamically adapt to wear changes under different usage scenarios, avoiding error accumulation caused by model rigidity. During the framework initialization phase, it is necessary to import material fatigue characteristic libraries, historical road adhesion coefficient databases, and user driving behavior profiles provided by tire manufacturers. The material fatigue characteristic library contains fatigue life data of the tire's rubber and cord core materials, directly determining the tire's basic wear resistance. The historical road adhesion coefficient database contains adhesion coefficient data collected by automakers for different road types under varying weather conditions. A lower adhesion coefficient indicates a greater likelihood of tire slippage during braking, leading to increased localized friction and indirectly accelerating wear. The user driving behavior profile database is built upon vehicle CAN bus data (frequency of rapid acceleration, braking, and sharp turns) to construct user habit data (e.g., aggressive drivers brake an average of 15 times per day, while mild-driving users brake only 5 times). Aggressive driving causes tires to frequently experience high loads and friction, significantly accelerating wear. Therefore, this database is crucial for model adaptation to different users. After importing data from these three databases, a convolutional recurrent neural network (CNN) architecture is used to perform spatiotemporal feature correlation analysis on the data. The architecture is a hybrid network combining convolutional neural networks (CNNs, which excel at extracting spatial features, such as the spatial distribution differences of tire wear under different loads) and recurrent neural networks (RNNs, which excel at extracting temporal features, such as the time trend of wear as mileage increases). Built and trained by the algorithm team, the CNN layers first extract spatial structural features of rubber and cord from a material fatigue property database, frictional spatial features of different road surfaces from a road adhesion coefficient database, and spatial load features of driving actions from a driving behavior database. The RNN layers then correlate these spatial features with the temporal dimension (mileage, usage time) to analyze the coupling relationship between material, road, and driving factors over time (e.g., high-fatigue materials + low-adhesion road surface + aggressive driving). The wear rate is twice that of a normal combination (due to excessive driving). Through this analysis, a three-dimensional mapping relationship of wear degradation is constructed as the preset wear model, with mileage as the time axis and load-temperature-pressure as the environmental axis. The time axis is the cumulative mileage of the vehicle (range 0-200,000 km), and the environmental axis is the two-dimensional environmental parameters composed of load (0-full load) and temperature-pressure (tire pressure 2.0-3.0 bar, temperature -20℃ to 80℃). The three-dimensional mapping relationship of wear degradation is the corresponding value of the tire wear coefficient under different mileage, load and temperature-pressure conditions (e.g., at 50,000 km, half load, 2.5 bar tire pressure, 25℃, the wear coefficient is 0.8; at the same mileage, full load, 2.3 bar tire pressure, 40℃, the wear coefficient is 0).7) This mapping relationship can cover most usage scenarios, providing a dynamic framework for wear coefficient calculation. The preset wear model is not fixed. During online operation, it combines real-time feedback of tread depth ultrasonic detection data to drive the generative adversarial network to dynamically optimize the model weight coefficients, thereby optimizing the preset wear model. The tread depth ultrasonic detection data is the actual tread depth data of the tire collected by the tread depth ultrasonic detection module installed on the vehicle chassis (directly facing the tire tread). Tread depth is a direct reflection of wear (new tire tread depth 6-8mm, wear limit 1.6mm). This data can verify in real time whether the wear coefficient predicted by the model is accurate. The generative adversarial network consists of two subnetworks: a generator and a discriminator. The model, composed of a network, was built by the algorithm team based on the GAN algorithm. The generator generates the corresponding tread depth based on the wear coefficient predicted by the preset model. The discriminator compares the generated tread depth with the actual detected data. If the difference is small (e.g., predicted 6mm, actual 5.8mm), the model is considered accurate, and the weight coefficients remain unchanged. If the difference is large (e.g., predicted 6mm, actual 5mm), the discriminator will provide an error signal, driving the generator to adjust the weight coefficients of mileage, load, and temperature / pressure dimensions in the model. Through this dynamic optimization, the preset wear model can continuously adapt to the actual wear changes of the tire, ensuring the accuracy of subsequent tire wear coefficient calculations and providing a reliable tire condition basis for adaptive braking energy recovery control.

[0019] The specific steps for analyzing the cumulative mileage of a vehicle within a pre-defined wear model are as follows: The vehicle's cumulative mileage is processed by the operating condition converter and converted into effective standard wear mileage based on three major characteristic parameters: load fluctuation rate, temperature and pressure change gradient, and frequency of rapid acceleration and deceleration. Through a dual attention mechanism embedded with a preset wear model, high-wear-risk mileage segments and their corresponding road surface roughness spectrum characteristics are identified. Based on the resonance effect of spectrum characteristics and material fatigue characteristics, the initial tire wear coefficient is analyzed, and the tire wear coefficient is output.

[0020] It needs further explanation that, after the pre-built wear model is completed, although the cumulative mileage of the vehicle can intuitively reflect the basic service life of the tires, the actual wear degree corresponding to the same mileage varies greatly under different driving conditions. For example, driving 100 kilometers under harsh conditions of full load, high temperature, and frequent emergency braking may result in tire wear equivalent to driving 200 kilometers under standard conditions (tire pressure 2.5 bar, tire temperature 25°C, half load, and smooth driving). If the original cumulative mileage is directly substituted into the model, the wear coefficient calculation will be biased due to ignoring the differences in driving conditions, which will affect the accuracy of subsequent regenerative braking control. Therefore, it is necessary to first process the cumulative mileage for driving conditions, and then combine it with the model's built-in dual attention mechanism to focus on key wear factors. The specific implementation method is as follows: First, the vehicle's cumulative mileage is processed by a performance condition converter. This converter is a hardware processing module integrated into the intelligent decision-making computing unit 2. Developed by the vehicle manufacturer's electronic control R&D team based on vehicle dynamics principles and extensive wear test data, its core function is to eliminate differences in driving conditions and convert actual mileage into effective mileage under a unified standard, ensuring the consistency of subsequent wear analysis benchmarks. The converter first extracts three core characteristic parameters affecting tire wear from vehicle sensors and the CAN bus, providing data support for subsequent conversion. Based on these three characteristic parameters—load fluctuation rate, temperature and pressure change gradient, and frequency of rapid acceleration and deceleration—it converts the mileage into effective standard wear mileage. These parameters are key to quantifying the severity of driving conditions. Load fluctuation rate refers to the frequency and amplitude of load changes during vehicle operation. Real-time load data for the front and rear axles and left and right wheels is collected every 10 seconds by onboard gravity sensors. The difference between two consecutive data collections is calculated as a percentage of the average load during that period. If the percentage exceeds 10% (considered a significant fluctuation), a fluctuation is recorded. The number of fluctuations within one minute is the load fluctuation rate. Higher fluctuation rates (e.g., 5 fluctuations per minute due to frequent loading and unloading of cargo in freight vehicles) result in frequent changes in tire contact pressure. The tire tread repeatedly experiences pressure impacts in localized areas, leading to wear rates more than 30% faster than under stable loads. The temperature-pressure gradient is the tire pressure-temperature gradient. The tire pressure and temperature change rate collected by the TPMS (Tire Pressure Monitoring System) over time, such as a 0.1 bar decrease in tire pressure or a 5°C increase in tire temperature within one minute, indicates a more significant gradient. A larger gradient value indicates more severe thermal expansion and contraction of the tire rubber due to temperature and pressure changes, leading to decreased tread structure stability and a higher risk of localized cracking or accelerated wear. The frequency of rapid acceleration and deceleration is collected via driving behavior data from the CAN bus. Counts are made when the accelerator pedal opening exceeds 80% (determined as rapid acceleration) and the brake pedal opening exceeds 70% (determined as emergency braking). The total frequency is calculated hourly. During rapid acceleration and deceleration, the friction between the tire and the road surface increases sharply (the friction during emergency braking can reach three times that of normal driving). Increased sliding friction between the surface and the ground leads to a multiple increase in wear. During the calculation process, the operating condition converter first maps these three parameters to corresponding efficiency coefficients (e.g., load fluctuation rate of 3 times / minute corresponds to coefficient 1.2, temperature and pressure change gradient of 0.1 bar / minute corresponds to coefficient 1.1, and rapid acceleration / deceleration frequency of 8 times / hour corresponds to coefficient 1.3). Then, the three coefficients are multiplied to obtain the total efficiency coefficient (1.2 × 1.1 × 1.3 = 1.716). Finally, the cumulative mileage of the vehicle is multiplied by the total efficiency coefficient to obtain the standard wear mileage. For example, if the vehicle has actually accumulated 150 kilometers and the total efficiency coefficient is 1.716, the standard wear mileage is 150 × 1.716 = 257 kilometers.The 4-kilometer mileage more accurately reflects the actual wear load borne by the tire, avoiding misleading underestimation or overestimation of the original mileage. After converting the effective standard wear mileage, a dual-attention mechanism embedded in the pre-defined wear model identifies high-wear-risk mileage segments and their corresponding road surface roughness spectrum features. The dual-attention mechanism, an algorithm module in the pre-defined wear model used to accurately locate key wear factors, was developed by the algorithm team based on the Transformer architecture and consists of two layers: mileage segment attention and road surface feature attention. It can filter out the segments and road surface features with the greatest impact on wear from massive mileage data, avoiding interference from invalid data. The mileage segment attention layer divides the effective standard wear mileage into units of 10 kilometers. The system is divided into several segments. For each segment, data on load fluctuation rate, temperature and pressure gradient, and frequency of rapid acceleration and deceleration are retrieved. The wear risk value for each segment is calculated (risk value = load fluctuation rate coefficient + temperature and pressure gradient coefficient + frequency of rapid acceleration and deceleration coefficient). If the risk value exceeds a preset threshold (experimentally calibrated to 1.5), the segment is marked as a high-wear-risk mileage segment (e.g., a 10km segment with a risk value of 1.8 is considered high-risk). Subsequently, the road surface feature attention layer connects to the historical road adhesion coefficient database to extract the road surface roughness spectrum features of the driving section corresponding to the high-risk mileage segment. This feature is obtained by detecting the frequency distribution of road surface reflected waves using millimeter-wave radar. Different road surface types correspond to different spectra, with high-frequency components accounting for a significant portion. A higher ratio indicates a rougher road surface, stronger friction and impact between the tire tread and the road surface, and a higher risk of wear. For example, high-risk segments correspond to gravel roads, where high-frequency components account for 28% of the spectral characteristics and will be highlighted. Finally, the initial tire wear coefficient is analyzed based on the resonance effect of spectral characteristics and material fatigue properties, and the tire wear coefficient is output. The resonance effect refers to the phenomenon where high-frequency components in the road surface roughness spectrum couple with the natural frequencies of the tire material, leading to increased wear. The material fatigue characteristic library provided by tire manufacturers contains the natural frequencies of rubber and cord core materials (such as the natural frequency of rubber 50-100Hz). When the high-frequency components in the road surface roughness spectrum (such as 80Hz) couple with the natural frequencies of the material... When frequencies overlap, the tire will generate a slight resonance during driving, increasing the impact friction between the tire tread and the ground, and the wear rate will be about 20% faster than when there is no resonance. During the analysis, the preset wear model first retrieves the natural frequency of the current tire material from the material fatigue characteristic library and compares it with the road roughness spectrum characteristics of the high wear risk segment to calculate the resonance matching degree (the higher the matching degree, the stronger the resonance effect). Then, a resonance correction coefficient is generated based on the matching degree (80% matching degree corresponds to a correction coefficient of 0.9, 50% corresponds to 0.95, and the lower the matching degree, the closer the correction coefficient is to 1). Finally, the initial tire wear coefficient (e.g., 0.8) is multiplied by the resonance correction coefficient (e.g., 0.9) to obtain the final tire wear coefficient (0.8 × 0.9 = 0).(72) This coefficient reflects both the basic wear caused by mileage and operating conditions, and corrects for the additional wear caused by road resonance. It accurately reflects the current actual wear state of the tire, providing a core basis for subsequent multi-sensor fusion calculation of the comprehensive friction coefficient and optimization of braking energy recovery control parameters.

[0021] The method for the road surface type detection unit to identify the type of the current driving road surface and generate the corresponding road surface friction coefficient is as follows: By fusing road texture semantic segmentation data collected by vehicle-mounted cameras, medium reflectivity matrix detected by millimeter-wave radar, and frequency domain energy distribution spectrum of suspension vibration sensors using a multi-source heterogeneous sensor array, a hierarchical classifier based on graph convolutional network is constructed. Based on texture features, reflection features, and vibration entropy values, probability distribution vectors for different road surface types are generated. The road surface-friction relationship knowledge graph is indexed according to the probability distribution vectors. Combined with the working condition compensation parameters of real-time rain sensor and infrared road surface thermometer, the road surface friction coefficient is output.

[0022] It needs further explanation that after determining the tire wear coefficient, the road surface friction coefficient becomes another core parameter for calculating the overall friction coefficient between the tire and the ground. The friction coefficient of dry asphalt roads can reach 0.8, while that of icy and snowy roads is only 0.1-0.2. If the road surface type cannot be accurately identified and the corresponding friction coefficient matched, the adjustment of the regenerative braking force is prone to deviation. For example, if the regenerative braking force is set to a high level based on the parameters for asphalt roads on icy and snowy roads, it may cause the wheels to lock up. Therefore, it is necessary to use a road surface type detection unit in conjunction with multi-source sensor data to achieve accurate determination of the road surface type and friction coefficient. The specific implementation method is as follows: The road surface type detection unit identifies the type of the current road surface and generates the road friction coefficient. This is achieved by fusing road texture semantic segmentation data collected by an onboard camera, the medium reflectivity matrix detected by millimeter-wave radar, and the frequency domain energy distribution spectrum of a suspension vibration sensor using a multi-source heterogeneous sensor array. This multi-source heterogeneous sensor array is a collaborative acquisition system composed of sensors of different types and functions (deployed by automakers according to vehicle driving safety requirements). It can acquire road information from multiple dimensions, including vision, radar, and vibration, avoiding misjudgments caused by environmental interference from a single sensor (such as blurred camera images in rain or abnormal radar reflections from ice and snow). Specifically, the road texture semantic segmentation data collected by the onboard camera is obtained by capturing real-time road images using a high-definition camera installed inside the vehicle's windshield, and then processed by the built-in semantic segmentation... The algorithm, based on a pre-trained U-Net network, is trained using 100,000 different road surface images to separate road surface regions from the background (such as pedestrians and vehicles) and annotate texture features (such as the particle texture of asphalt roads, the crack texture of cement roads, and the reflective texture of icy and snowy roads). The final output is road surface texture semantic segmentation data, which contains pixel-level texture classification results for road surface regions, intuitively reflecting the visual characteristics of the road surface. The medium reflectivity matrix detected by millimeter-wave radar is generated by a millimeter-wave radar installed on the vehicle bumper emitting radar waves towards the road surface and receiving reflected signals from different directions. Each element in the matrix represents the reflection intensity at the corresponding angle (e.g., asphalt roads have weaker radar wave reflection intensity, with matrix element values ​​of 0.3-0.5, while icy and snowy roads have stronger reflection intensity, with element values ​​of 0.8-1).0), the type of road surface medium can be distinguished by the difference in reflectivity. The frequency domain energy distribution spectrum of the suspension vibration sensor is obtained by collecting vibration signals caused by road bumps from vibration sensors installed at the vehicle suspension springs or shock absorbers. The time-domain vibration signal is then converted into a frequency-domain signal by Fourier transform to obtain the frequency domain energy distribution spectrum. Different road surfaces have different degrees of bumps (gravel roads are bumpy, with high-frequency energy accounting for 30%; asphalt roads are smooth, with low-frequency energy accounting for 80%). This spectrum can help judge the roughness of the road surface by the difference in energy distribution. After completing the multi-source data collection, a hierarchical classifier based on graph convolutional networks needs to be constructed. Graph convolutional networks are a type of deep learning network that is good at processing data with correlation relationships (the algorithm team developed it based on the GCN algorithm framework). It can regard the three features of road surface texture, reflection, and vibration as interrelated nodes and explore the hidden correlations between features (such as strong reflection + high-frequency vibration may correspond to a gravel road covered by ice and snow). The hierarchical classifier is a classification structure designed according to the two-level classification logic of major category - minor category. The first level first classifies the road surface into hard road surface (asphalt, The road surface is classified into three main categories: cement, soft road surface (sand, soil), and special road surface (ice, snow, water). The second level further subdivides these main categories into specific types (e.g., hard road surface is subdivided into dry asphalt, wet asphalt, dry cement, and wet cement) to avoid classification confusion caused by direct subdivision. During construction, the texture features of the vehicle camera, the reflection features of the millimeter-wave radar, and the frequency domain energy features of the suspension vibration are first aligned by timestamps to form feature vectors for each set of data. These vectors are then input into a graph convolutional network. The network undergoes hierarchical training (first training the classification ability of the main categories, then optimizing the classification accuracy of the subcategories) to ultimately form a hierarchical classifier that can output the probability of road surface types. When the classifier is working, it generates probability distribution vectors for different road surface types based on texture features, reflection features, and vibration entropy values. Vibration entropy is a parameter that measures the irregularity of the suspension vibration signal. The more intense and irregular the vibration, the higher the entropy value (0.8-1.0); the more gentle and regular the vibration (e.g., asphalt road), the lower the entropy value (0.2-0.4). It is calculated from the frequency domain energy distribution spectrum and is used to supplement the quantitative indicators of vibration features. The classifier analyzes these three types of features simultaneously. For example, if the texture feature shows uniform particles and no obvious cracks (prone to asphalt roads), the reflection feature shows a reflection intensity of 0.4 (within the range of asphalt roads), and the vibration entropy value of 0.3 (corresponding to smooth road surfaces), then the classifier will generate: dry asphalt road: 0.85, wet asphalt road: 0.1, and dry cement road: 0.05. Other types: A probability distribution vector of 0. The type with the highest probability in this vector is the preliminary judgment result of the current road surface. The probability distribution of the vector can reflect the confidence of the classification and avoid absolute misjudgment based on a single result. After initially determining the road surface type, the pre-set road surface-friction relationship knowledge graph is indexed according to the probability distribution vector. The pre-set road surface-friction relationship knowledge graph is a structured database built by the car manufacturer in conjunction with the road engineering team. It uses road surface type-environmental conditions-base friction coefficient as the core relationship and stores the baseline friction coefficient values ​​for different scenarios (such as 0.8 for dry asphalt road, 0.5 for wet asphalt road, and 0.15 for icy and snowy road). The knowledge graph will automatically index the corresponding baseline friction coefficient (0.8) according to the road surface type with the highest probability in the probability distribution vector (such as dry asphalt road, probability 0.85). However, the baseline value does not consider real-time environmental changes (such as sudden rainfall, road icing). Therefore, it is necessary to combine the working condition compensation parameters of the real-time rain sensor and the infrared road surface thermometer. The real-time rain sensor is an optical rain sensor installed on the roof of the vehicle (detection range 0-50mm / h) that can output rainfall in real time. Rainfall speeds are categorized into levels (e.g., light rain: 0-5 mm / h, moderate rain: 5-15 mm / h, heavy rain: above 15 mm / h). The greater the rainfall, the more significant the decrease in road surface friction coefficient (during moderate rain, the asphalt road friction coefficient drops from 0.8 to 0.5, requiring a compensation factor of 0.625). Infrared road surface thermometers are infrared temperature sensors installed on the front bumper of vehicles (measuring range -40℃ to 80℃, accuracy ±0.5℃). They detect real-time road surface temperature. When the temperature is below 0℃, even without significant rainfall, a thin layer of ice may form on the road surface, requiring an additional factor of 0.3-0. A compensation coefficient of 0.5 (e.g., 0.4 for a temperature of -5℃); the working condition compensation parameter is a friction coefficient correction coefficient calculated based on rainfall level and road surface temperature. Multiplying the baseline friction coefficient by this coefficient yields the actual road surface friction coefficient under the current working conditions. For example, the baseline value for dry asphalt road is 0.8, and a sudden moderate rain (compensation coefficient 0.625) results in an actual road surface friction coefficient of 0.8 × 0.625 = 0.5. This value accurately reflects the real-time friction capacity of the road surface, providing reliable road surface parameters for subsequent multi-sensor fusion calculation of the comprehensive friction coefficient.

[0023] A method for dynamically calculating the combined friction coefficient between a tire and the ground based on a rule-and-data fusion algorithm is described below: A collaborative decision-making architecture of rule engine and deep learning model is created. The rule engine loads the ISO road-tire coupling friction rule library and tire wear safety threshold constraints. The deep learning model uses a spatiotemporal graph neural network to process the spatiotemporal coupling characteristics of tire wear coefficient and road friction coefficient. The decision weights of rules and data are dynamically allocated through a gating fusion mechanism. When the wear coefficient exceeds the safety threshold, the rule-dominated mode is activated to generate a conservative friction estimate. For other working conditions, the data-driven mode is activated to generate an optimized friction estimate and output the comprehensive friction coefficient.

[0024] It needs further explanation that after obtaining the tire wear coefficient and the road friction coefficient separately, relying on a single coefficient cannot fully reflect the actual friction characteristics between the tire and the ground. For example, the overall friction capability of a high-wear tire (wear coefficient 0.3) on a dry asphalt road (road friction coefficient 0.8) differs significantly from that of a new tire (wear coefficient 0.9) on an icy or snowy road (road friction coefficient 0.2). If a single coefficient is used, it can easily lead to deviations in the setting of the regenerative braking force. Therefore, a rule-based and data-driven fusion algorithm is needed to couple and analyze the two to generate an accurate overall friction coefficient. The specific implementation method is as follows: The algorithm for dynamically calculating the comprehensive friction coefficient based on the fusion of rules and data first requires the creation of a collaborative decision-making architecture between the rule engine and the deep learning model. This collaborative decision-making architecture is a hybrid decision-making framework designed by the vehicle manufacturer's electronic control team in conjunction with the algorithm team, taking into account both braking safety requirements and data accuracy requirements. The core idea is to ensure a safety baseline through rules and improve adaptation accuracy through data, avoiding the overly conservative nature of single rule decisions (such as always taking a low friction coefficient, resulting in low recovery efficiency) or the potential risks of single data decisions (such as data anomalies leading to an overestimation of the friction coefficient and causing braking slippage). The architecture comprises two core modules: a rule engine module responsible for safety constraints and a deep learning model module responsible for precise calculations. These two modules communicate bidirectionally via a high-speed data interface. The rule engine outputs safety thresholds and constraints to the model, while the model feeds back real-time feature analysis results to the rule engine, ensuring that decisions comply with safety standards and are adapted to real-time driving scenarios. Next, the rule engine loads the ISO road-tire coupling friction rule library and tire wear safety threshold constraints. The ISO road-tire coupling friction rule library is a standardized friction reference library (such as ISO 8349 and ISO 18100 standards) developed by the International Organization for Standardization (ISO) based on massive road testing. The library contains reference ranges for coupling friction coefficients for different road surface types (asphalt, cement, gravel, snow) and tires with different wear levels (new tires, moderate wear, near-limit wear). For example, the reference value for the coupling friction coefficient of a new tire (wear coefficient 0.9) on a dry asphalt road is 0.7-0.9, and the reference value for a moderately worn tire (wear coefficient 0.5) on the same road surface is 0.5-0.7, providing an internationally recognized standard basis for rule-based decision-making. The tire wear safety threshold constraint is a critical value set based on the tire manufacturer's safety test data. Typically, a wear coefficient of 0.3 is set as the safety threshold. This threshold has been verified through 100,000 braking distance tests. When the tire wear coefficient is below 0.3, the tread depth approaches the wear limit of 1.6mm, friction capacity decreases sharply, and braking distance increases by more than 50% compared to a new tire. Therefore, the rule engine uses this threshold as the core constraint for safety decisions. Once a wear coefficient below 0.3 is detected, a high-level safety decision logic is immediately initiated. Subsequently, the deep learning model uses a spatiotemporal graph neural network to process the spatiotemporal coupling characteristics of tire wear coefficient and road friction coefficient. The spatiotemporal graph neural network is a deep learning model improved by the algorithm team based on the graph attention network (GATv2), specifically designed to handle the coupling relationship between wear changes in the time dimension and road friction differences in the spatial dimension. In the time dimension, the model retrieves the tire wear coefficient sequence from the past 5 minutes (obtained from the dynamic wear assessment module at 10-second intervals) and analyzes the decay trend of the wear coefficient with increasing mileage (e.g., the wear coefficient decreases from 0.51 to 0 for every 1 kilometer driven).50); Spatially, the model collects road surface friction coefficient data for the current driving segment and 1 kilometer before and after it (obtained from the road surface type detection unit at 50-meter intervals), analyzes the spatial distribution differences of the road surface friction coefficient (e.g., the friction coefficient of the current segment is 0.6, and the friction coefficient drops to 0.4 when entering a wet segment 500 meters ahead); Through the attention mechanism of the graph neural network, the model focuses on dangerous spatiotemporal segments with high wear and low friction (e.g., wear coefficient 0.32 and road surface friction coefficient 0.35 300 meters ahead), and assigns higher feature weights to such segments to ensure that the coupled features are not masked by the data of normal working conditions, thereby more accurately capturing the dynamic relationship between the two. After processing the coupled features, the decision weights of the rules and data are dynamically allocated through a gating fusion mechanism. This is the core unit in the architecture responsible for balancing rule constraints and data accuracy. It is implemented by an adaptive gating function, which calculates the weight allocation ratio based on the safety level of the current operating condition: the safety level is determined by the product of tire wear coefficient and road friction coefficient. A product > 0.3 indicates a safe operating condition (e.g., wear coefficient 0.5 × road friction coefficient 0.7 = 0.35), 0.2-0.3 indicates a critical operating condition, and < 0.2 indicates a dangerous operating condition. Under safe operating conditions, data-driven weights account for 70% and rule weights account for 30%, prioritizing the model's adaptability to real-time operating conditions to improve accuracy. Under critical operating conditions, both account for 50%, balancing safety and accuracy requirements. Under dangerous operating conditions, rule weights account for 80% and data weights account for 20%, prioritizing adherence to rule safety constraints to avoid risks caused by data fluctuations. For example, under safe operating conditions, the gating function assigns higher weights to the accurate calculation results output by the model, while retaining the basic reference of the rules to prevent the model from overfitting abnormal data. When the wear coefficient exceeds the safety threshold (i.e., <0.3), the rule-driven mode is activated to generate a conservative friction estimate. The core of the rule-driven mode is to be conservative rather than risky: the rule engine first retrieves the lowest reference value of the friction coefficient corresponding to the current wear coefficient from the ISO road-tire coupling friction rule library (e.g., when the wear coefficient is 0.28, the lowest reference value for all road types is 0.3), and then combines it with the measured value of the current road friction coefficient (e.g., 0.35), taking the smaller value of the two as the conservative friction estimate (i.e., 0.3); at the same time, the rule engine ignores any values ​​higher than the specified value in the model output. The calculation result of this value avoids overestimation of the friction coefficient due to insufficient adaptation of the model to high wear conditions. This conservative estimate ensures that even when the tire is close to the wear limit, the braking energy recovery force will be set based on a safe friction coefficient to prevent wheel lock-up or slippage. In other conditions where the wear coefficient does not exceed the safe threshold (i.e., ≥0.3), the data-driven mode is enabled to generate an optimized friction estimate. The data-driven mode fully combines the accurate calculation of the model with the safety reference of the rules: First, the spatiotemporal graph neural network inputs the processed coupled features into the fully connected layer and outputs a preliminary optimized friction coefficient (e.g., 0.3).62); Next, the gating fusion mechanism, based on the weights allocated to the current safety level (e.g., 70% data weight and 30% rule weight under safe operating conditions), weights and fuses the preliminary optimized value with the reference value output by the rule engine (e.g., the reference value of 0.58 for the corresponding operating condition in the ISO library). The calculation process is: optimized friction estimate = preliminary optimized value × data weight + rule reference value × rule weight (i.e., 0.62 × 0.7 + 0.58 × 0.3 = 0.602). This fusion method not only preserves the model's adaptability to real-time operating conditions (e.g., temporary road friction fluctuations and short-term load changes), but also corrects potential biases in the model through the rule reference value, ensuring that the friction estimate is both accurate and safe. Finally, the comprehensive friction coefficient is output. Before outputting, the optimized friction estimate or conservative friction estimate needs to be range-checked. The reasonable range of the comprehensive friction coefficient is set to 0.1-0.9 (0.1 corresponds to the lowest friction on extreme icy and snowy roads, and 0.9 corresponds to the highest friction on dry asphalt roads with new tires). If the calculation result exceeds this range, it is automatically corrected to the range boundary value (e.g., if the calculation result is 0.08, it is corrected to 0.1, and if it is 1.0, it is corrected to 0.9). After the check is passed, the comprehensive friction coefficient (e.g., 0.602) will be transmitted to the adaptive control execution unit (3) in real time via the CAN bus. This provides the core decision-making basis for the subsequent dynamic adjustment of the braking kinetic energy recovery intensity (e.g., setting a high recovery intensity when the friction coefficient is high, and reducing the recovery intensity when the friction coefficient is low) and response time (e.g., shortening the response time when the friction coefficient fluctuates greatly), ensuring that the kinetic energy recovery is both efficient and safe.

[0025] The rule-based and data-driven fusion algorithm integrates the nonlinear correlation between tire wear coefficient and road surface type, specifically as follows: A multidimensional response surface for the wear-road coupling effect is constructed. The horizontal axis of the surface represents the wear coefficient grading interval, the vertical axis represents the road surface type code, and the vertical axis represents the friction attenuation sensitive factor. By transferring learning, the wear-road-friction mapping relationship of vehicle big data from different climate zones is loaded. Prior knowledge is injected into key nodes of the surface. When the system detects that a high-wear tire has entered the preset road surface, the exponential compensation mechanism of the friction attenuation sensitive factor is activated to dynamically improve the safety margin of the comprehensive friction coefficient and realize the digital expression of the nonlinear effect of wear and road surface.

[0026] Further explanation is needed. After the rule-based and data-driven fusion algorithm performs basic friction calculations through a collaborative decision-making architecture, it is still necessary to further address the nonlinear relationship between tire wear coefficient and road surface type. This relationship is not a simple linear superposition, but rather the friction attenuation of a high-wear tire on a low-friction road surface is significantly greater than the combined effect of the two factors alone (e.g., the friction attenuation of a tire with a wear coefficient of 0.2 on icy or snowy roads is more than twice that of a tire with a wear coefficient of 0.5 on the same road surface). Ignoring this nonlinearity can easily lead to an overestimation of the overall friction coefficient, resulting in inaccurate regenerative braking force (e.g., misjudging an icy or snowy road as a dry road and setting a high regenerative braking force, causing wheel lock-up). Therefore, it is necessary to digitize and quantify this by constructing a multidimensional response surface. The specific implementation method is as follows: The rule-based and data-based fusion algorithm integrates this nonlinear relationship, which is first manifested in the construction of a multidimensional response surface for the wear-road coupling effect. This multidimensional response surface for the wear-road coupling effect is a three-dimensional mathematical model developed by the car company's algorithm team by combining vehicle dynamics and tribology principles to intuitively present the relationship between wear, road surface and friction decay. Its core is to transform the abstract nonlinear relationship into a calculable and queryable numerical surface through the combination of parameters of the three coordinate axes. The horizontal axis of the surface represents the wear coefficient grading range. This range is divided into three levels based on the tire wear safety threshold (0.3) mentioned earlier, combined with real vehicle test data: low wear range (0.6-1.0, corresponding to new tires or lightly worn tires, tread depth > 4mm), medium wear range (0.3-0.6, corresponding to mid-wear tires, tread depth 2-4mm), and high wear range (0-0.3, corresponding to tires nearing the wear limit, tread depth < 2mm). The grading setting allows tires with different wear levels to correspond to clear analysis ranges, avoiding excessive fluctuations in the surface due to continuous values. The vertical axis represents the road surface type code. The code is a standardized digital identifier for road surface types based on road friction characteristics and common scenarios (developed by the automaker in conjunction with the road engineering team). (Formulated), for example, dry asphalt road code 01, wet asphalt road code 02, dry cement road code 03, gravel road code 04, icy and snowy road code 05, and waterlogged road code 06. The code is convenient for surface storage index and can quickly associate with the reference value of road friction coefficient. The vertical axis is the friction attenuation sensitivity factor. The friction attenuation sensitivity factor is a parameter that quantifies the degree of friction coefficient attenuation caused by tire wear and road type (the value is 1.0-2.0, and the larger the value, the more serious the friction attenuation). For example, the sensitivity factor of a new tire (wear coefficient 0.9) on a dry asphalt road (code 01) is 1.0 (almost no attenuation), and the sensitivity factor of a high-wear tire (0.2) on an icy and snowy road (05) is 1.9 (90% friction attenuation). This factor is the core output of the surface and directly determines the correction direction and magnitude of the comprehensive friction coefficient. After the surface is constructed, the wear-road-friction mapping relationship of vehicle big data from different climate zones needs to be loaded through transfer learning. Transfer learning is a machine learning technique that uses model parameters trained on existing data to quickly adapt to new scenarios (this is based on the ResNet transfer learning framework). The reason for using this technique is that the road surface characteristics of different climate zones vary greatly (e.g., there are many icy and snowy roads in the north during winter, and many waterlogged roads in the south during the rainy season). If a surface model is trained separately for each climate zone, it will consume a lot of data and time. However, transfer learning can transfer the parameters of the temperate climate zone model that has been trained to the cold and subtropical scenarios, and can be adapted with only a small amount of local data for fine-tuning.The vehicle big data for different climate zones is collected by car manufacturers through in-vehicle T-BOX (remote information processing terminal) and covers the driving data of more than 100,000 vehicles in Northeast (cold zone), North China (temperate zone), and South China (subtropical zone). The data includes the tire wear coefficient, road surface type code, and actual measured friction coefficient of vehicles in different climate zones. Through these data, a mapping relationship between wear-road surface-friction can be established (for example, in cold winter, the actual friction coefficient of a tire with a wear coefficient of 0.3 on icy and snowy roads (05) is 0.15, corresponding to a sensitivity factor of 1.8). After loading these mapping relationships onto the multidimensional response surface, the surface will automatically adjust the sensitivity factor values ​​of each interval. For example, after loading the cold zone data, the sensitivity factor of all wear intervals on icy and snowy roads (05) is increased by 0.1-0.2, ensuring that the surface adapts to the road surface characteristics of different regions and avoiding errors caused by a one-size-fits-all approach. To further improve the accuracy of the surface, prior knowledge needs to be injected into the key nodes of the surface. Key nodes refer to the parameter combinations on the surface that have the greatest impact on braking safety. These are usually extreme or common working conditions such as high wear range + low friction road surface (e.g., wear coefficient 0.2 + icy / snowy road 05) or low wear range + high friction road surface (e.g., 0.9 + dry asphalt 01). If these nodes rely solely on data fitting, the sensitivity factor is easily biased due to data sparsity. Prior knowledge comes from international standards (e.g., ISO 8349 road surface - tire friction standard) and empirical values ​​from real-world vehicle tests by automakers. For example, the ISO standard stipulates that the wear limit is approaching. The coefficient of friction of the tires (wear coefficient 0.3) on icy and snowy roads must not exceed 0.2, corresponding to a sensitivity factor of ≥1.7. In real vehicle tests, the measured values ​​of the sensitivity factor for this combination are mostly 1.7-1.8. Therefore, prior knowledge of a sensitivity factor ≥1.7 is injected at the critical node (0.3,05) to forcibly correct any values ​​below 1.7 that may appear in the surface fitting, ensuring that the safety bottom line is not breached. At the same time, knowledge of a sensitivity factor ≤1.1 is injected at common working condition nodes (such as 0.8,01) to avoid over-correction that leads to a decrease in recovery efficiency. By combining prior knowledge with data, the surface can be made both accurate and safe.When the system detects a high-wear tire entering a preset road surface during vehicle operation, it activates the exponential compensation mechanism of the friction attenuation sensitivity factor. The detection of high-wear tires is performed by the intelligent decision calculation unit (2) in real time by comparing the tire wear coefficient with the safety threshold (wear coefficient < 0.3 is judged as high wear). The preset road surface refers to a pre-set low-friction risk road surface, namely the road surface type coded as 05 (ice and snow road), 06 (waterlogged road), and 04 (gravel road). The system identifies the tires by the code output by the road surface type detection unit. The exponential compensation mechanism is an enhanced safety correction mechanism relative to linear compensation. Linear compensation only adjusts the sensitivity factor by a fixed proportion (such as uniformly increasing it by 0.2), while exponential compensation increases the sensitivity factor according to the product of wear degree × road risk, such as high wear. When a worn tire (0.2) enters an icy road (05), the product is 0.2 × 0.8 (road risk coefficient, icy road risk coefficient 0.8) = 0.16. The index compensation will increase the sensitivity factor from the base of 1.8 to 1.8^(1+0.16) = 1.8 × 1.17 ≈ 2.1 (here the index base is the basic value of the sensitivity factor, and the index is 1 + risk product, ensuring that the compensation range increases significantly with the increase of risk); if a high-wear tire enters a low-risk dry asphalt road (01, risk coefficient 0.1), the product is 0.2 × 0.1 = 0.02, and the sensitivity factor only increases from 1.2 to 1.2 × 1.02 ≈ 1.22. This differentiated compensation can accurately match the needs of high-risk strong compensation and low-risk weak compensation, avoiding the waste of recycling efficiency caused by over-compensation. Finally, the safety margin of the comprehensive friction coefficient is dynamically improved to achieve a digital expression of wear and nonlinear road surface effects. The safety margin refers to the additional safety buffer space reserved when substituting the compensated sensitive factor into the calculation of the comprehensive friction coefficient. For example, if the comprehensive friction coefficient calculated by the sensitive factor is 0.25, and the safety margin is set to 10%, then the final output comprehensive friction coefficient is 0.25×(1-10%)=0.225. This value is lower than the actual calculated value, which can reserve more safety redundancy for the braking system and prevent the friction coefficient from being reduced due to sudden changes in the road surface (such as local icing on icy or snowy roads). Further reduction; while digital expression, through the three coordinate axes of the multidimensional response surface, transforms the originally abstract nonlinear relationship of more severe frictional attenuation of high-wear tires on icy and snowy roads into specific sensitive factor values ​​(such as 1.9), compensation range (such as increasing to 2.1), and comprehensive friction coefficient after safety margin (such as 0.225), so that the fusion algorithm can quantitatively calculate rather than qualitatively judge, ensuring that the output of each comprehensive friction coefficient can accurately reflect the nonlinear coupling effect of wear and road surface, and provide core parameter support that is more in line with actual working conditions for subsequent adaptive braking kinetic energy recovery control.

[0027] The rule-based and data-driven fusion algorithm further includes: An uncertainty quantification module is embedded to calculate the confidence entropy values ​​of the data from the tire wear sensor group and the road surface detection sensor group in real time. Based on the confidence entropy values, a Bayesian inference framework is constructed to derive the probability distribution band of the wear-road surface fusion result. When the bandwidth of the probability distribution band exceeds the preset tolerance, a multi-sensor cross-validation mechanism is triggered. That is, the on-board LiDAR is called to scan the geometry of the tire contact mark, and the actual friction boundary is calculated by combining the motor torque fluctuation inversion model. The inversion results are used to correct the output of the rule-based and data-based fusion algorithm online to ensure the robustness of the comprehensive friction coefficient.

[0028] Further explanation is needed. Based on the multi-dimensional response surface of the wear-road coupling effect and the rule-data collaborative decision-making architecture constructed above, the rule- and data-based fusion algorithm also needs to address the risks brought about by the uncertainty of sensor data. For example, fogging of the vehicle camera lens in rainy weather can lead to deviations in road texture recognition, and the tire pressure and temperature sensor signals can drift in low-temperature environments. These can cause errors in the calculation of tire wear coefficient or road friction coefficient. If directly used for the derivation of the comprehensive friction coefficient, it may cause inaccurate braking energy recovery control. Therefore, an uncertainty quantification module needs to be embedded to ensure the reliability of the fusion result through multi-stage verification. The specific implementation method is as follows: The rule-based and data-driven fusion algorithm further includes: First, an uncertainty quantification module is embedded. This module is an algorithm sub-module integrated into the intelligent decision-making computing unit 2. Developed by the automotive company's algorithm team based on information entropy theory and sensor reliability engineering, its core function is to evaluate the uncertainty level of the output data from the tire wear sensor group and the road surface detection sensor group, preventing erroneous data caused by sensor failure or environmental interference from being directly used in the fusion calculation. This module receives raw data and status signals (such as whether the sensors are in normal working mode and whether there is signal packet loss) from both types of sensor groups in real time, providing basic data support for subsequent confidence entropy value calculation. Next, the confidence entropy values ​​of the data from the tire wear sensor group and the road surface detection sensor group are calculated in real time. The tire wear sensor group is the set of sensors mentioned above used to collect tire wear-related data, including a vehicle mileage sensor (providing cumulative mileage), a tire pressure and temperature sensor (providing tire pressure and temperature sequences), and a tread depth ultrasonic detection module (providing real-time tread depth). The road surface detection sensor group is the set of sensors that collect road surface feature data, including an onboard camera (road surface texture), millimeter-wave radar (medium reflectivity), a suspension vibration sensor (frequency domain energy), a real-time rain sensor (rainfall level), and an infrared road surface thermometer (road surface temperature). Data confidence entropy is a quantitative indicator that measures the uncertainty of sensor data (ranging from 0 to 1; the closer the entropy value is to 1, the higher the data uncertainty; the closer it is to 0, the more reliable the data). During calculation, the uncertainty quantification module first sets a baseline reliability weight for each sensor (e.g., the ultrasonic tread depth detection module has high accuracy, so the weight is 0.9; the weight of the rain camera is reduced to 0.5). Then, it adjusts the weights based on the real-time status of the sensors (e.g., whether there are signal fluctuations, whether the data is within a reasonable range). Finally, it calculates the confidence entropy value for each of the two groups using the information entropy formula (simplified to a weighted average of the weight deviations of each sensor). For example, in the tire wear sensor group, the tread depth sensor data is stable (weight 0.9), and the TPMS signal has slight drift (weight 0.7), resulting in a final confidence entropy value of 0.2. In the road surface detection sensor group, due to the rain camera's weight of 0.5 and the millimeter-wave radar's weight of 0.8, the confidence entropy value is 0.4. This difference in entropy values ​​reflects a higher level of uncertainty in the road surface data. Then, a Bayesian inference framework is constructed based on the confidence entropy value. The Bayesian inference framework is an inference model that uses prior knowledge and real-time data to dynamically update the probability distribution (developed by the algorithm team based on Bayes' theorem and adapted to the real-time computing needs of vehicles). The prior knowledge here is the comprehensive friction coefficient reference range corresponding to different wear coefficients and road surface types in the ISO road-tire coupling friction rule base mentioned above (such as a wear coefficient of 0.5 and a reference range of 0.5-0.7 for dry asphalt roads). The real-time data is the confidence entropy value of the two sensor groups and the currently calculated tire wear coefficient and road friction coefficient.During construction, the framework first transforms prior knowledge into a prior probability distribution (e.g., a reference range of 0.5-0.7 corresponds to a probability peak of 0.6). Then, it adjusts the dispersion of the probability distribution based on the confidence entropy value. The higher the entropy value (the less reliable the data), the more dispersed the distribution; the lower the entropy value (the more reliable the data), the more concentrated the distribution. For example, if the entropy value of road surface data is 0.4 (high uncertainty), the prior distribution is broadened from 0.5-0.7 to 0.4-0.8, forming the basic framework of the posterior probability distribution. Subsequently, the probability distribution band of the wear-road surface fusion result is derived. The wear-road surface fusion result is the comprehensive friction coefficient initially calculated by the current rule-data fusion algorithm (e.g., 0.6). The probability distribution band is an interval based on the output of the Bayesian inference framework, containing the possible fluctuation range of the fusion result. The upper and lower limits of the interval are calculated by the fusion result ± (confidence entropy value × correction coefficient) (the correction coefficient is calibrated to 0.2 through real vehicle testing to balance accuracy and safety). For example, if the fusion result is 0.6 and the average confidence entropy value of the two sensor groups is 0.3, then the probability distribution band is 0.6 ± (0.3 × 0.2) = 0.54 - 0.66. This range directly reflects the uncertainty range of the current comprehensive friction coefficient. The narrower the range, the more reliable the result; the wider the range, the higher the uncertainty, providing a quantitative basis for subsequent risk assessment. When the bandwidth of the probability distribution band (i.e., the difference between the upper and lower limits, such as 0.66-0.54=0.12) exceeds the preset tolerance, the multi-sensor cross-validation mechanism is triggered. The preset tolerance is the maximum allowable uncertainty bandwidth set by the automaker based on braking safety requirements (set to 0.15 after verification by 100,000 braking distance tests). If the bandwidth exceeds 0.15 (e.g., distribution band 0.5-0.7, bandwidth 0.2), it indicates that the uncertainty of the current sensor data has exceeded the safety tolerance range and needs to be corrected by additional sensor verification. The multi-sensor cross-validation mechanism is a process of calling other high-precision sensors on the vehicle (not the original wear and road surface sensor group) to independently verify the fusion results. The core is to introduce third-party data to break the uncertainty dependence of the original sensor group and ensure the objectivity of the correction results. The cross-validation mechanism involves using an onboard LiDAR to scan the geometry of the tire contact patch. The onboard LiDAR is a high-precision laser scanning device (10Hz scanning frequency, ±1mm ranging accuracy) installed in the center of the vehicle chassis, directly facing the tire's contact patch area. It can penetrate light water mist or dust, avoiding interference from adverse weather conditions, unlike cameras. The geometry of the tire contact patch refers to the shape, size, and pressure distribution characteristics of the actual contact area between the tire and the ground (e.g., the contact patch of a new tire on a dry asphalt road is rectangular, with an area of ​​approximately 200cm²; the contact patch of a highly worn tire on icy or snowy roads is irregularly shaped, with an area increasing to 250cm²). These characteristics are directly related to the tire's actual friction capability. A larger patch area and clearer edges indicate more sufficient contact between the tire and the ground, and a friction coefficient closer to the calculated value. Conversely, a blurry patch or an abnormally large or small area indicates a potential deviation in the friction coefficient.The lidar transmits the scanned imprint length, width, and pressure distribution (indirectly calculated through laser reflection intensity) data to the intelligent decision-making calculation unit 2 in real time, serving as the geometric verification basis for friction capability. Simultaneously, it calculates the actual friction boundary using a motor torque fluctuation inversion model. This model, developed by the vehicle manufacturer's electronic control team based on motor dynamics principles, is based on the fixed correlation between tire friction and motor torque fluctuation. When the coefficient of friction between the tire and the ground changes, the reaction force of the tire on the drive motor changes, causing slight fluctuations in the motor's output torque (e.g., when the coefficient of friction decreases, the reaction force decreases, and the torque fluctuation amplitude increases). The model reads the instantaneous torque data of the motor from the vehicle's CAN bus in real time (sampling frequency 1kHz), calculates the standard deviation of torque fluctuation (e.g., 0.5 N·m under normal friction, increasing to 1.2 N·m under low friction), and then uses a preset torque fluctuation-friction force mapping table (generated by bench testing, e.g., a standard deviation of 1.2 N·m corresponds to a friction force of 3000 N) to deduce the actual friction force currently experienced by the tire. Finally, combined with the vehicle's real-time load (from the vehicle's gravity sensor), the actual friction boundary (i.e., the true friction coefficient range between the tire and the ground, e.g., 0.55-0.60) is calculated using the relationship of friction coefficient = friction force ÷ (load × gravitational acceleration). This boundary does not depend on the original wear and road surface sensor data and is an independent physical-level verification result. Finally, the output of the rule-based and data-driven fusion algorithm is corrected online using the inversion results to ensure the robustness of the comprehensive friction coefficient. Online correction refers to adjusting the output of the fusion algorithm in real time without interrupting the regenerative braking control: First, the overlapping area (0.55-0.60) between the inversion result (actual friction boundary 0.55-0.60) and the original probability distribution band (e.g., 0.5-0.7) is calculated, and the median of this overlapping area (0.575) is taken as the correction benchmark value; then, based on the grounding imprint morphology of the lidar scan (e.g., normal imprint area and clear edges indicate reliable inversion results), the benchmark value is assigned a weight of 0.8, and the original fusion result is assigned a weight of 0.2. The final output value (e.g., 0.575×0.8+0.6×0.2=0.58) is obtained by calculating the corrected comprehensive friction coefficient = benchmark value × 0.8 + original fusion result × 0.2. This correction retains the real-time performance of the original fusion algorithm while eliminating the uncertainty of the original sensor group through independent data from lidar and motor torque inversion. This reduces the fluctuation range of the comprehensive friction coefficient from 0.5-0.7 to 0.57-0.59, and the bandwidth is reduced to 0.02 (below the preset tolerance of 0.15). Ultimately, this ensures the robustness of the comprehensive friction coefficient, meaning that regardless of whether the original sensor data is disturbed, the reliability of the comprehensive friction coefficient can be maintained through cross-validation correction, providing stable core parameter support for adaptive braking kinetic energy recovery control.

[0029] The rule-based and data-driven fusion algorithm integrates the nonlinear correlation between tire wear coefficient and road surface type, specifically as follows: A multidimensional response surface for the wear-road coupling effect is constructed. The horizontal axis of the surface represents the wear coefficient grading interval, the vertical axis represents the road surface type code, and the vertical axis represents the friction attenuation sensitive factor. By transferring learning, the wear-road-friction mapping relationship of vehicle big data from different climate zones is loaded. Prior knowledge is injected into key nodes of the surface. When the system detects that a high-wear tire has entered the preset road surface, the exponential compensation mechanism of the friction attenuation sensitive factor is activated to dynamically improve the safety margin of the comprehensive friction coefficient and realize the digital expression of the nonlinear effect of wear and road surface.

[0030] Further explanation is needed. After the rule-based and data-driven fusion algorithm completes the basic comprehensive friction coefficient calculation through a collaborative decision-making architecture, it is also crucial to address the nonlinear relationship between tire wear coefficient and road surface type. This relationship is not a simple linear superposition, but rather exhibits the characteristic that the friction attenuation of a high-wear tire on a low-friction road surface is far greater than the sum of the individual effects of the two. For example, the friction attenuation of a tire with a wear coefficient of 0.2 on icy or snowy roads is more than 2.5 times that of a tire with a wear coefficient of 0.5 on the same road surface. Ignoring this nonlinearity can easily lead to an overestimation of the comprehensive friction coefficient, resulting in an overly strong setting of the regenerative braking force. For instance, if an icy or snowy road is mistakenly identified as a dry road, a high regenerative braking force may cause the wheels to lock up. Therefore, it is necessary to digitize and quantify this by constructing a multidimensional response surface. The specific implementation method is as follows: The rule-based and data-driven fusion algorithm integrates this nonlinear correlation, firstly manifested in the construction of a multidimensional response surface for the wear-road coupling effect. This multidimensional response surface is a three-dimensional mathematical model developed by the automotive company's algorithm team, combining vehicle dynamics and tribology principles. Its core is to transform the abstract nonlinear correlation into a calculable and queryable numerical surface through the combination of parameters on three coordinate axes, providing an intuitive quantitative carrier for subsequent friction decay analysis. The horizontal axis of the surface represents the wear coefficient grading interval. This interval is based on the previously mentioned tire wear safety threshold (0.3), combined with 100,000 sets of real-vehicle wear test data, and is divided into three levels: low wear interval (0.6-1.0, corresponding to new tires or lightly worn tires, tread depth > 4mm, extremely weak friction decay), medium wear interval (0.3-0.6, corresponding to mid-term worn tires, tread depth 2-4mm, moderate friction decay), and high wear interval (0-0.3, corresponding to tires nearing the wear limit, tread depth < 2mm, severe friction decay). This grading allows for... Tires with the same wear level correspond to a clearly defined analysis range to avoid excessive surface fluctuations caused by continuous values. The vertical axis represents the road surface type code, which is a standardized digital identifier developed by the car manufacturer in conjunction with the road engineering team based on the road surface friction characteristics. For example, dry asphalt road code 01 (friction coefficient baseline value 0.8), wet asphalt road code 02 (0.5), dry cement road code 03 (0.75), gravel road code 04 (0.4), icy and snowy road code 05 (0.15), and waterlogged road code 06 (0.3). The code facilitates surface storage and quick indexing, and can also directly associate with the road surface friction coefficient baseline value. The vertical axis represents friction decay. Sensitivity factor, friction attenuation sensitivity factor is the core parameter for quantifying the degree of friction coefficient attenuation caused by tire wear and road surface type (value ranges from 1.0 to 2.0, with larger values ​​indicating more severe friction attenuation). For example, the sensitivity factor for a new tire (wear coefficient 0.9) on a dry asphalt road (code 01) is 1.0 (almost no attenuation), while the sensitivity factor for a high-wear tire (0.2) on an icy or snowy road (05) is 1.9 (90% friction attenuation). This factor directly determines the direction and magnitude of the correction to the overall friction coefficient and is the core output item of the surface. After the surface is constructed, different parameters need to be loaded through transfer learning. The wear-road-friction mapping relationship of vehicle big data in different climate zones is explored. Transfer learning is a machine learning technique that uses model parameters trained on existing scenarios to quickly adapt to new scenarios (this is based on the ResNet transfer learning framework). The reason for using this technique is that the road surface characteristics of different climate zones vary greatly (e.g., there are many icy and snowy roads in the north during winter and many waterlogged roads in the south during the rainy season). If a surface model is trained separately for each climate zone, it will consume a lot of data and time. However, transfer learning can transfer the parameters of the temperate climate zone model that has been trained to the cold and subtropical scenarios, and can be adapted with only a small amount of local data for fine-tuning.The vehicle big data for different climate zones is the driving data of more than 150,000 vehicles in Northeast (cold zone), North China (temperate zone), and South China (subtropical zone) collected by car manufacturers through the vehicle-mounted T-BOX (remote information processing terminal). The data includes the real-time tire wear coefficient, road surface type code, and actual measured friction coefficient of vehicles in different climate zones. Through these data, a precise wear-road-friction mapping relationship can be established (for example, in cold winter, the actual friction coefficient of a tire with a wear coefficient of 0.3 on icy and snowy roads (05) is 0.15, corresponding to a sensitivity factor of 1.8). After loading these mapping relationships onto the multidimensional response surface, the surface will automatically adjust the sensitivity factor values ​​of each interval. For example, after loading the cold zone data, the sensitivity factor of all wear intervals on icy and snowy roads (05) is increased by 0.1-0.2, ensuring that the surface can adapt to the road surface characteristics of different regions and avoid errors caused by a one-size-fits-all approach. To further improve the accuracy of the surface, prior knowledge needs to be injected into the key nodes of the surface. Key nodes refer to the parameter combinations on the surface that have the greatest impact on braking safety. These are usually extreme or high-frequency operating conditions such as high wear range + low friction road surface (e.g., wear coefficient 0.2 + icy / snowy road 0.5) or low wear range + high friction road surface (e.g., 0.9 + dry asphalt 0.1). If these nodes rely solely on data fitting, the sensitivity factor is easily biased due to data sparsity. Prior knowledge comes from international standards (e.g., ISO 8349 Road Surface - Tire Friction Standard) and empirical values ​​from real-vehicle tests by automakers. For example, the ISO standard stipulates that tires approaching their wear limit (wear coefficient 0.2 + tire friction coefficient 0.5) are considered to have a certain friction factor. The friction coefficient of a loss coefficient of 0.3 on icy and snowy roads must not exceed 0.2, corresponding to a sensitivity factor of ≥1.7. However, in actual vehicle tests, the measured values ​​of the sensitivity factor for this combination are mostly between 1.7 and 1.8. Therefore, prior knowledge of a sensitivity factor ≥1.7 is injected at the critical node (0.3, 05) to forcibly correct any values ​​below 1.7 that may appear in the surface fitting, ensuring that the safety baseline is not breached. At the same time, knowledge of a sensitivity factor ≤1.1 is injected at high-frequency operating condition nodes (such as 0.8, 01) to avoid over-correction that would reduce the efficiency of brake energy recovery. Through the complementarity of prior knowledge and data, the surface is made both accurate and in line with actual safety requirements. When the vehicle is in motion, if the system detects that a high-wear tire has entered the preset road surface, it activates the exponential compensation mechanism of the friction attenuation sensitive factor. The detection of high-wear tires is completed in real time by the intelligent decision calculation unit (2). The tire wear coefficient is compared with the safety threshold. If the wear coefficient is <0.3, it is judged as high wear. The preset road surface refers to the low friction risk road surface set in advance, that is, the road surface type coded as 04 (gravel road), 05 (ice and snow road), and 06 (waterlogged road). The system identifies the road surface type by the code output by the road surface type detection unit.The exponential compensation mechanism is a reinforced safety correction mechanism relative to linear compensation. Linear compensation adjusts the sensitivity factor by a fixed proportion (e.g., a uniform increase of 0.2), while exponential compensation increases the sensitivity factor exponentially based on the product of wear level and road risk. For example, if a high-wear tire (0.2) drives onto an icy road (0.5, road risk coefficient 0.8, where a higher risk coefficient means a more dangerous road), the product of the two is 0.16. Exponential compensation will increase the sensitivity factor from the base of 1.8 to 1.8 × (1 + 0.16) = 2.1 (here, 1 + the product is the exponential adjustment coefficient, ensuring that the compensation range increases significantly with the risk). If a high-wear tire drives onto a low-risk dry asphalt road (0.1, risk coefficient 0.1), the product is 0.02, and the sensitivity factor only increases from 1.2 to 1.2 × 1.02 ≈ 1.22. This differentiated compensation can accurately match the needs of high-risk, high-compensation and low-risk, low-compensation, avoiding waste of recycling efficiency caused by overcompensation. Finally, the safety margin of the comprehensive friction coefficient is dynamically improved to realize the digital expression of wear and nonlinear road surface effects. The safety margin refers to the additional safety buffer space reserved when the compensated sensitive factor is substituted into the calculation of the comprehensive friction coefficient. For example, if the comprehensive friction coefficient calculated by the compensated sensitive factor is 0.25, and the safety margin is set to 10% (calibrated by actual vehicle braking test), then the final output comprehensive friction coefficient is 0.25×(1-10%)=0.225. This value is lower than the actual calculated value, which can reserve more safety redundancy for the braking system and prevent risks caused by sudden changes in the road surface (such as local icing on icy and snowy roads leading to a further reduction in the friction coefficient). Digital representation, on the other hand, uses the three coordinate axes of a multidimensional response surface to transform the originally abstract nonlinear relationship—such as the more severe friction decay of high-wear tires on low-friction roads—into specific sensitive factor values ​​(e.g., 1.9), exponential compensation magnitude (e.g., increased to 2.1), and comprehensive friction coefficient after safety margin (e.g., 0.225). This allows the fusion algorithm to perform quantitative calculations rather than qualitative judgments, ensuring that the output of each comprehensive friction coefficient accurately reflects the nonlinear coupling effect between wear and road surface, providing core parameter support that is more in line with actual working conditions for subsequent adaptive braking energy recovery control.

[0031] The method for generating optimized control parameters by the adaptive control unit to dynamically adjust the regenerative braking force and response time involves the following steps: A two-layer optimized control architecture is constructed. The upper layer generates a basic recovery strategy parameter set by querying the Pareto front curve of the recovery force-response time based on the comprehensive friction coefficient. The lower layer deploys a model predictive controller to receive real-time signals of motor speed, battery SOC, and brake pedal opening gradient. Dynamic compensation is superimposed on the basic recovery strategy parameter set. The calculation of dynamic compensation incorporates a feedforward control channel for the rate of change of friction coefficient and a feedback channel for recovery torque fluctuation. The recovery force gradient and response time constant are dynamically adjusted through a sliding mode variable structure algorithm, and a safety protection mechanism is activated simultaneously. When a sudden change in friction coefficient is detected, a gradual migration strategy of recovery force is adopted to avoid mechanical braking intervention impact.

[0032] It needs further explanation that after obtaining the accurate comprehensive friction coefficient, the adaptive control unit needs to convert it into specific braking energy recovery operation parameters. If only a single fixed parameter is used (such as setting a 50% recovery force and a 200ms response time regardless of the friction coefficient), it will lead to excessive recovery force on low-friction surfaces (such as icy and snowy roads) causing wheel lock-up, or insufficient recovery force on high-friction surfaces (such as dry asphalt roads) wasting kinetic energy. Therefore, a two-layer optimized control architecture needs to be constructed to balance the stability of the basic strategy and the adaptability to dynamic operating conditions. The specific implementation method is as follows: The adaptive control unit generates optimized control... To control parameters, a dual-layer optimized control architecture must first be constructed. This dual-layer optimized control architecture is a control framework designed by the vehicle manufacturer's electronic control team in combination with the stability and dynamic requirements of regenerative braking. It consists of an upper basic strategy layer and a lower dynamic compensation layer. The upper layer is responsible for outputting static optimal parameters that adapt to the current friction conditions, avoiding decision delays caused by real-time calculation pressure. The lower layer is responsible for adjusting parameters according to instantaneous operating conditions, solving dynamic changes that the basic strategy cannot cover (such as sudden acceleration or full battery charge). The two communicate in real time through a high-speed data bus to ensure that the control is both stable and flexible, avoiding the problem of either slow response or easy fluctuation in a single layer. After the architecture is built, the upper layer queries the Pareto front curve of the pre-set recovery force-response time based on the comprehensive friction coefficient. The Pareto front curve of the recovery force-response time is an optimized curve generated by the car manufacturer through bench testing and 100,000+ kilometers of real vehicle verification. The core is to find the Pareto optimal balance point between recovery force (measures the intensity of kinetic energy recovery, expressed as a percentage of the maximum recovery torque of the motor, such as 30% or 50%) and response time (the time from detecting braking demand to the start of the recovery system, in milliseconds). That is, it is not possible to shorten the response time while increasing the recovery force (or vice versa), otherwise it will lead to system instability (such as high force + short response easily causing torque shock). The curve construction process is as follows: Under different comprehensive friction coefficients (0.1-0.9, with each 0.1 as a node), different combinations of recovery force and response time are tested, and the recovery efficiency (the ratio of recovered electrical energy to braking kinetic energy) and braking smoothness (vehicle acceleration fluctuation value) are recorded. The combination with the highest efficiency and smoothness under each friction coefficient is selected, and these combination points are connected in order of friction coefficient to form the Pareto front curve (e.g., when the comprehensive friction coefficient is 0.8, the optimal combination is 50% recovery force and 150ms response time; when the friction coefficient is 0.3, the optimal combination is 20% recovery force and 100ms response time).After querying the curve, a basic recovery strategy parameter set is generated. This set is structured data containing core control parameters under the current operating conditions. Besides recovery force (e.g., 50%) and response time (e.g., 150ms), it also includes the upper limit of recovery torque (the maximum recovery torque value to avoid motor overload, e.g., 200N·m) and a pedal opening-recovery force mapping table (the recovery force ratio corresponding to different brake pedal openings, e.g., 30% pedal opening corresponds to 30% recovery force). These parameters are extracted from the optimal combination corresponding to the Pareto front curve, ensuring they represent the static optimal solution under the current comprehensive friction coefficient, providing a benchmark for lower-level dynamic adjustments. Corresponding to the upper-level static strategy, a model predictive controller (MPC) is deployed at the lower level. The MPC is an algorithm module that predicts future operating conditions based on real-time data and dynamically adjusts control quantities (developed by the algorithm team based on predictive control theory, adapted to the real-time computing needs of vehicles). Its core advantage is its ability to predict changes in operating conditions 100-200ms in advance, avoiding delayed adjustments. The controller needs to receive real-time signals of motor speed, battery SOC, and brake pedal opening gradient: Motor speed is transmitted in real-time by the motor controller via CAN bus (sampling frequency 1kHz), reflecting the current kinetic energy state of the motor (the higher the speed, the more kinetic energy can be recovered); Battery SOC (State of Charge, remaining battery power) is collected by the Battery Management System (BMS) (sampling frequency 1Hz). When the SOC is too high (e.g., >95%), the battery cannot receive a large amount of electrical energy, and the recovery force needs to be reduced; Brake pedal opening gradient is calculated from the pedal travel data collected by the brake pedal sensor (unit % / ms, e.g., 5% / ms represents an increase of 5000% in pedal opening per second). The larger the gradient, the stronger the driver's braking intention, and the faster the recovery response speed needs to be. After receiving the signal, the controller superimposes a dynamic compensation amount onto the basic recovery strategy parameter set. This dynamic compensation amount is an adjustment value used to correct the basic parameters and adapt to real-time operating conditions (it can be positive or negative; for example, if the basic recovery force is 50% and the compensation amount is -10%, the actual force is 40%). The superposition logic is as follows: when the motor speed is too high (e.g., >3000rpm), the compensation amount increases by 5% to improve recovery efficiency; when the battery SOC is >90%, the compensation amount decreases by 15% to avoid battery overcharging; when the brake pedal opening gradient is >8% / ms, the compensation amount decreases by 5% in recovery force and by 30ms in response time, prioritizing braking smoothness. This superposition transforms the control parameters from a static baseline to a dynamically adapted one, perfectly matching the instantaneous operating conditions.The calculation of dynamic compensation requires the introduction of a feedforward control channel for the rate of change of friction coefficient and a feedback channel for recovery torque fluctuation. The rate of change of friction coefficient is obtained by dividing the difference between two consecutive comprehensive friction coefficients by the time interval (e.g., if it decreases from 0.8 to 0.6 within 100ms, the rate of change is -2.0 / second). The feedforward control channel uses this rate of change to adjust the compensation amount in advance. If the friction coefficient drops rapidly (rate of change < -0.5 / second), it is predicted that the friction coefficient will be even lower in the future, and the compensation amount is reduced by 5%-10% in advance to avoid slippage caused by excessive recovery force, thus achieving early defense. The recovery torque fluctuation feedback channel collects the recovery torque data output by the motor in real time and calculates the standard deviation of torque fluctuation (e.g., normal fluctuation < 5 N·m, exceeding it is judged as excessive fluctuation). If the fluctuation exceeds the standard, the compensation amount is adjusted in feedback (e.g., when the fluctuation is large, the compensation amount is finely adjusted by 0.5% / time) to correct the torque output and achieve real-time correction.The combination of feedforward and feedback ensures both accuracy and stability of the compensation amount. Once the compensation amount is determined, the sliding mode variable structure algorithm dynamically adjusts the recovery force gradient and response time constant. This algorithm is a robust control algorithm (simplified and adapted to vehicle computing power by the algorithm team) suitable for handling scenarios with frequent parameter changes in braking systems. The recovery force gradient represents the rate of change of the recovery force (e.g., 5% / 100ms, meaning a 5% change in recovery force every 100ms). Adjusting the gradient avoids jerks caused by sudden changes in force. The response time constant is the core parameter determining the response speed (the smaller the time constant, the faster the response). During algorithm operation, these two parameters are adjusted according to the magnitude of the dynamic compensation amount: if the absolute value of the compensation amount is large (e.g., ±15%), it indicates a significant adjustment to the recovery force. In this case, the force gradient is set to 10% / 100ms (rapid adjustment), and the time constant is reduced by 20ms (faster response). If the absolute value of the compensation amount is small (e.g., ±5%), the gradient is set to 3% / 100ms (gradual adjustment), and the time constant remains unchanged, ensuring smooth adjustment. Smooth and efficient, the system simultaneously activates a safety protection mechanism while dynamically adjusting. This mechanism is a safety fallback designed to handle extreme conditions (such as sudden changes in the coefficient of friction), preventing risks caused by untimely dynamic adjustments. The core of the mechanism is the detection of sudden changes in the coefficient of friction. By continuously monitoring the comprehensive coefficient of friction, if the change in the coefficient of friction exceeds 0.2 within 100ms (a preset threshold, calibrated through real vehicle testing; exceeding this value can easily lead to braking instability), it is determined to be a sudden change in the coefficient of friction (e.g., the coefficient of friction suddenly drops from 0.6 to 0.3 when the vehicle suddenly enters an icy road). When a sudden change is detected, a gradual shift strategy for recovery force is adopted to avoid mechanical braking intervention shock. The gradual shift strategy for recovery force means slowly reducing the recovery force from the current value to the target value (e.g., from 40% to 15%), with a shift time set at 500ms (to avoid torque interruption due to excessive speed), rather than an instantaneous switch. Mechanical braking intervention shock means that when the recovery force suddenly disappears or drops significantly, the system needs to quickly switch to mechanical braking (brake pad braking). If the switch is too abrupt, it will cause the vehicle to suddenly jerk, affecting the driving experience and safety. Through gradual transition, the decrease in regenerative braking force is seamlessly integrated with the intervention of mechanical braking (e.g., when the regenerative braking force drops to 15%, mechanical braking gradually intervenes at a force of 5%). The combined effect maintains stable braking intensity, completely avoiding any impact and ensuring a smooth and safe braking process. Finally, the regenerative braking force (e.g., 45%) and response time (e.g., 120ms), after dual-layer optimization and adjustment, are transmitted in real time to the motor controller and battery management system via the CAN bus. This drives the motor to regenerate kinetic energy at the set force, ensuring that the entire process efficiently utilizes kinetic energy while also considering braking safety and smoothness, achieving the core control objective of adaptive braking kinetic energy recovery.

[0033] The second objective of this invention is to provide a system for implementing an adaptive braking kinetic energy recovery control method based on multi-sensor fusion, including any one of the above-mentioned features, comprising: The multi-source sensing fusion unit 1 integrates a vehicle mileage sensor, a tire pressure and temperature sensor, an on-board gravity sensor, a multi-source heterogeneous road surface detection sensor array, and a tire tread depth ultrasonic detection module, which are used to collect multi-dimensional data related to tire wear and multi-modal feature data of the road surface in real time. The intelligent decision computing unit 2 has a built-in dynamic wear assessment module, a road surface-friction relationship knowledge graph engine, a rule and data fusion algorithm processor, and an uncertainty quantification correction module. It receives the output data of the multi-source perception fusion unit 1, generates the tire wear coefficient through the dynamic wear assessment module, generates the road surface friction coefficient through the road surface classification and friction mapping module, and integrates the nonlinear correlation between the wear coefficient and the road surface friction coefficient based on the fusion algorithm processor to output the comprehensive friction coefficient between the tire and the ground. The adaptive control execution unit 3 includes a two-layer optimization controller and a safety protection logic module. It receives the comprehensive friction coefficient output by the intelligent decision calculation unit, generates a basic recovery parameter set through the Pareto front strategy generator, combines the dynamic compensation amount with the model prediction controller, dynamically adjusts the kinetic energy recovery force gradient and response time constant using the sliding mode variable structure algorithm, and suppresses the mechanical braking impact during frictional abrupt changes through the gradual migration strategy. The closed-loop feedback calibration unit 4 connects the motor torque fluctuation inversion module and the lidar imprint scanner to capture tire ground contact pattern and actual friction boundary data in real time. The data is then fed back to the uncertainty quantification module in the intelligent decision calculation unit 2 and the safety protection logic module in the adaptive control execution unit 3 to drive multi-sensor cross-validation and online model optimization.

[0034] This invention collects tire wear and road surface feature data through a multi-source sensing fusion unit, and generates tire wear coefficient and road surface friction coefficient through an intelligent decision-making calculation unit. The nonlinear correlation between the two is integrated through a rule and data fusion algorithm to output a comprehensive friction coefficient. The adaptive control execution unit dynamically adjusts the braking energy recovery force and response time accordingly. The system includes a multi-source sensing fusion unit, an intelligent decision-making calculation unit, an adaptive control execution unit, and a closed-loop feedback calibration unit. The closed-loop unit corrects data deviations through cross-verification using lidar and a motor torque inversion model. This invention solves the problems of poor working condition adaptability and unreliable data, and improves energy recovery efficiency, braking safety, and driving smoothness.

[0035] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An adaptive braking kinetic energy recovery control method based on multi-sensor fusion, characterized in that: Includes the following steps: Tire wear coefficient is obtained in real time through vehicle mileage sensors. The tire wear coefficient is calculated based on the vehicle's cumulative mileage and a preset wear model. Meanwhile, the road surface type detection unit identifies the type of the current driving road surface and generates the corresponding road surface friction coefficient. The tire wear coefficient and the road surface friction coefficient are then input into the multi-sensor fusion processing unit. The multi-sensor fusion processing unit uses a rule-based and data-based fusion algorithm to dynamically calculate the comprehensive friction coefficient between the tire and the ground. The rule-based and data-based fusion algorithm integrates the nonlinear correlation between the tire wear coefficient and the road surface type. A method for dynamically calculating the combined friction coefficient between a tire and the ground based on a rule-and-data fusion algorithm is described below: A collaborative decision-making architecture for a rule engine and a deep learning model is created. The rule engine loads the ISO road-tire coupling friction rule library and tire wear safety threshold constraints. The deep learning model uses a spatiotemporal graph neural network to process the spatiotemporal coupling characteristics of tire wear coefficient and road friction coefficient. The decision weights of rules and data are dynamically allocated through a gating fusion mechanism. When the wear coefficient exceeds the safety threshold, the rule-dominated mode is activated to generate a conservative friction estimate. For other working conditions, the data-driven mode is activated to generate an optimized friction estimate and outputs the comprehensive friction coefficient. The rule- and data-based fusion algorithm further includes: An uncertainty quantification module is embedded to calculate the confidence entropy values ​​of the data from the tire wear sensor group and the road surface detection sensor group in real time. Based on the confidence entropy values, a Bayesian inference framework is constructed to derive the probability distribution band of the wear-road fusion result. When the bandwidth of the probability distribution band exceeds the preset tolerance, a multi-sensor cross-validation mechanism is triggered. That is, the vehicle-mounted LiDAR is called to scan the geometry of the tire contact mark, and the actual friction boundary is calculated by combining the motor torque fluctuation inversion model. The inversion results are used to correct the output of the rule-based and data-based fusion algorithm online to ensure the robustness of the comprehensive friction coefficient. The rule-based and data-driven fusion algorithm integrates the nonlinear correlation between tire wear coefficient and road surface type, specifically as follows: A multidimensional response surface for wear-road coupling effect is constructed. The horizontal axis of the surface represents the wear coefficient grading interval, the vertical axis represents the road surface type code, and the vertical axis represents the friction attenuation sensitive factor. By transferring learning, the wear-road-friction mapping relationship of vehicle big data from different climate zones is loaded. Prior knowledge is injected into key nodes of the surface. When the system detects that a high-wear tire has entered the preset road surface, the exponential compensation mechanism of the friction attenuation sensitive factor is activated to dynamically improve the safety margin of the comprehensive friction coefficient and realize the digital expression of the nonlinear effect of wear and road surface. Based on the comprehensive friction coefficient, optimized control parameters are generated by the adaptive control unit to dynamically adjust the braking energy recovery force and response time, so as to achieve adaptive braking energy recovery control for different wear conditions and road surface conditions.

2. The adaptive braking kinetic energy recovery control method based on multi-sensor fusion according to claim 1, characterized in that: The specific steps for obtaining the tire wear coefficient in real time using a vehicle mileage sensor are as follows: A multi-dimensional data fusion mechanism is established to integrate the vehicle's cumulative mileage, real-time load distribution data collected by the vehicle's gravity sensor, and tire pressure and temperature change sequences from the tire pressure and temperature sensor. These data are then input into the dynamic wear assessment module. The dynamic wear assessment module uses a pre-trained wear feature extraction network to analyze the historical attenuation patterns of tire contact deformation under different load and temperature / pressure conditions, generating a dynamic correction factor. This dynamic correction factor is then superimposed on the baseline wear curve to output the initial tire wear coefficient.

3. The adaptive braking kinetic energy recovery control method based on multi-sensor fusion according to claim 2, characterized in that: The specific steps for constructing the preset wear model are as follows: The wear evolution modeling framework based on reinforcement learning imports a material fatigue property database, a road adhesion coefficient historical database, and a user driving behavior profile database provided by the tire manufacturer during the initialization phase. Through a convolutional recurrent neural network architecture, spatiotemporal feature correlation analysis is performed on the three databases to construct a three-dimensional mapping relationship of wear degradation with mileage as the time axis and load-temperature-pressure as the environmental axis as the preset wear model. When the preset wear model is running online, it is combined with real-time feedback of tread depth ultrasonic detection data to drive the adversarial generative network to dynamically optimize the model weight coefficients, which is used to optimize the preset wear model.

4. The adaptive braking kinetic energy recovery control method based on multi-sensor fusion according to claim 3, characterized in that: The specific steps for analyzing the cumulative mileage of a vehicle within a pre-defined wear model are as follows: The vehicle's cumulative mileage is processed by the operating condition converter and converted into effective standard wear mileage based on three major characteristic parameters: load fluctuation rate, temperature and pressure change gradient, and frequency of rapid acceleration and deceleration. Through a dual attention mechanism embedded with a preset wear model, high-wear-risk mileage segments and their corresponding road surface roughness spectrum characteristics are identified. Based on the resonance effect of spectrum characteristics and material fatigue characteristics, the initial tire wear coefficient is analyzed, and the tire wear coefficient is output.

5. The adaptive braking kinetic energy recovery control method based on multi-sensor fusion according to claim 1, characterized in that: The method for the road surface type detection unit to identify the type of the current driving road surface and generate the corresponding road surface friction coefficient includes the following specific steps: By fusing road texture semantic segmentation data collected by vehicle-mounted cameras, medium reflectivity matrix detected by millimeter-wave radar, and frequency domain energy distribution spectrum of suspension vibration sensors using a multi-source heterogeneous sensor array, a hierarchical classifier based on graph convolutional network is constructed. Based on texture features, reflection features, and vibration entropy values, probability distribution vectors for different road surface types are generated. The road surface-friction relationship knowledge graph is indexed according to the probability distribution vectors. Combined with the working condition compensation parameters of real-time rain sensor and infrared road surface thermometer, the road surface friction coefficient is output.

6. The adaptive braking kinetic energy recovery control method based on multi-sensor fusion according to claim 1, characterized in that: The method for generating optimized control parameters and dynamically adjusting the regenerative braking force and response time by the adaptive control unit includes the following steps: A two-layer optimized control architecture is constructed. The upper layer generates a basic recovery strategy parameter set by querying the Pareto front curve of the recovery force-response time based on the comprehensive friction coefficient. The lower layer deploys a model predictive controller to receive real-time signals of motor speed, battery SOC, and brake pedal opening gradient. Dynamic compensation is superimposed on the basic recovery strategy parameter set. The calculation of dynamic compensation incorporates a feedforward control channel for the rate of change of friction coefficient and a feedback channel for recovery torque fluctuation. The recovery force gradient and response time constant are dynamically adjusted through a sliding mode variable structure algorithm, and a safety protection mechanism is activated simultaneously. When a sudden change in friction coefficient is detected, a gradual migration strategy of recovery force is adopted to avoid mechanical braking intervention impact.

7. A system for implementing the adaptive braking kinetic energy recovery control method based on multi-sensor fusion as described in any one of claims 1-6, characterized in that, include: The multi-source sensing fusion unit (1) integrates a vehicle mileage sensor, a tire pressure and temperature sensor, an on-board gravity sensor, a multi-source heterogeneous road surface detection sensor array, and a tire tread depth ultrasonic detection module to collect multi-dimensional data related to tire wear and multi-modal feature data of the road surface in real time. The intelligent decision computing unit (2) has a built-in dynamic wear assessment module, a road surface-friction relationship knowledge graph engine, a rule and data fusion algorithm processor and an uncertainty quantification correction module. It receives the output data of the multi-source perception fusion unit (1), generates the tire wear coefficient through the dynamic wear assessment module, generates the road surface friction coefficient through the road surface classification and friction mapping module, and integrates the nonlinear relationship between the wear coefficient and the road surface friction coefficient based on the fusion algorithm processor to output the comprehensive friction coefficient between the tire and the ground. The adaptive control execution unit (3) includes a dual-layer optimization controller and a safety protection logic module. It receives the comprehensive friction coefficient output by the intelligent decision calculation unit, generates a basic recovery parameter set through the Pareto front strategy generator, combines the dynamic compensation amount with the model prediction controller, dynamically adjusts the kinetic energy recovery force gradient and response time constant using the sliding mode variable structure algorithm, and suppresses the mechanical braking impact during frictional abrupt changes through the gradual migration strategy. The closed-loop feedback calibration unit (4) connects the motor torque fluctuation inversion module and the lidar imprint scanner to capture the tire ground contact pattern and actual friction boundary data in real time. The data is then fed back to the uncertainty quantification module in the intelligent decision calculation unit (2) and the safety protection logic module in the adaptive control execution unit (3) to drive multi-sensor cross-validation and online model optimization.

Images