Child electric motorcycle intelligent safety control system based on Internet of Things

Abstract

The invention discloses a child electric motorcycle intelligent safety control system based on the Internet of Things, and relates to the technical field of electric motorcycle safety control. The problem of insufficient driving risk perception and control response of children is solved. The method comprises the following steps: firstly, fusing obstacle trajectory prediction, a road adhesion coefficient and an environment contour, constructing a dynamic risk vector field, and generating a collision probability and out-of-control coefficient fused risk assessment matrix; then, the matching degree of handle operation and vehicle body response is calculated through cross attention, and a third-level behavior label is output; according to the behavior level and the risk direction, anti-sideslip control or progressive power limitation is executed, and intervention characteristics are recorded; and finally, constructing a virtual training task by using parent operation data, extracting a personalized control threshold, dynamically optimizing system parameters, and realizing a child driving behavior-oriented closed-loop safety control system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electric motorcycle safety control, in particular to a child electric motorcycle intelligent safety control system based on the Internet of Things. BACKGROUND

[0002] With the upgrading of family consumption and the popularization of smart travel concept, children's electric motorcycles gradually evolve from a single toy attribute to "intelligentization + safety" direction, becoming an important traffic entertainment device in the parent-child scene. Parents' awareness of safety risks during children's driving is constantly improving, prompting product design to shift from traditional mechanical control to a comprehensive safety control system of "environment perception + behavior understanding + active intervention". The embedding of Internet of Things technology enables children's transportation tools to have the potential for real-time environmental perception, remote interactive management and data-driven optimization. How to achieve fine-grained risk control while ensuring children's riding pleasure has become an important research direction in this field.

[0003] However, existing safety control of children's electric motorcycles relies on fixed speed limit, electronic fence or single threshold emergency stop logic, and its control feedback often lags behind the risk occurrence, which is prone to the situation of "danger first and then control" in emergency situations, and the emergency stop or sudden deceleration brings discomfort to children. The coarse-grained discrimination based on GPS or acceleration sensor often leads to false alarm or missed alarm for complex conditions such as wet road and obstacle approach; single rule is difficult to consider multiple dynamic scenarios, and frequent hard threshold triggering not only affects continuous riding, but also cannot be personalized according to different children's driving styles and environmental conditions. There is no closed-loop data-driven mechanism, and the system control strategy is difficult to continuously evolve with user feedback, which restricts the improvement of safety protection effect and user acceptance. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a child electric motorcycle intelligent safety control system based on the Internet of Things, which solves the problems in the above background art.

[0005] In order to achieve the above object, the present application is realized by the following technical scheme: the child electric motorcycle intelligent safety control system based on the Internet of Things comprises the following modules: a risk modeling module, a behavior analysis module, a safety decision module and an optimization module; the risk modeling module is used for generating a dynamic risk vector field centered on the vehicle by fusing mobile obstacle trajectory prediction data, road adhesion coefficient analysis data and static obstacle three-dimensional contour data, and outputting a fusion risk assessment matrix containing a collision probability gradient and a road out-of-control coefficient; the behavior analysis module is used for receiving handle operation time sequence features and vehicle body dynamic response features from vehicle-mounted sensors, and analyzing the matching degree of operation intention and vehicle response through cross attention, wherein the road out-of-control coefficient is used as a weight factor to participate in the matching degree calculation, and a three-level behavior label of active risk behavior, passive loss behavior or normal operation behavior is output; the safety decision module is used for executing differentiated control according to the direction priority of the fusion risk assessment matrix and the behavior label level, starting anti-skid vector control based on the direction of the collision probability gradient when the behavior label is passive loss behavior, triggering progressive power limitation when the behavior label is active risk behavior and the operation direction coincides with the high-risk area, and simultaneously recording the spatial features and operation features of the intervention scene in real time; and the optimization module is used for generating a mirror virtual training task according to the intervention scene features, converting parent operation data into individualized safety threshold through cluster analysis, and dynamically updating the judgment criteria of the behavior analysis module and the control parameters of the safety decision module.

[0006] Further, by fusing mobile obstacle trajectory prediction data, road adhesion coefficient analysis data and static obstacle three-dimensional contour data, a dynamic risk vector field centered on the vehicle is generated, comprising the following steps: firstly, time sequence modeling is performed on the mobile target point cloud collected by the radar, the probability distribution of the multi-path motion trajectory is deduced, the three-dimensional topological map of the environment is constructed through depth vision, the physical properties of the static obstacles are identified through semantic understanding, and the passing feasibility evaluation of each direction is output, then the tire pressure distribution and the vehicle body posture change data are fused, and the maximum lateral grip force critical value of the current road is dynamically analyzed; finally, the mobile target trajectory probability distribution, the static obstacle passing evaluation and the road grip force critical value are spatially vector fused to form a directional risk field, wherein each vector points to the direction of the risk source, and the vector length reflects the risk intensity level.

[0007] Further, the output includes a fusion risk assessment matrix containing the collision probability gradient and the road loss of control coefficient, including the following steps: dividing the fan-shaped area radially in the dynamic risk vector field along the polar coordinate system, extracting the maximum risk vector intensity in each area, and calculating the collision probability gradient combined with the relative speed of the moving target; At the same time, according to the real-time deviation degree of the maximum lateral adhesion coefficient and the road grip force, the vehicle loss of control risk coefficient is dynamically quantified; Then map the collision probability gradient to the Cartesian coordinate system grid, and nonlinearly couple with the road loss of control coefficient at the same grid position to form a spatial evaluation matrix containing collision risk and loss of control risk, wherein the high-risk area presents a gradient transition feature.

[0008] Further, the handle operation timing features and vehicle body dynamic response features from the vehicle-mounted sensor are received, and the matching degree of the operation intention and the vehicle response is analyzed through cross attention, wherein the road loss of control coefficient is used as a weight factor in the matching degree calculation, including the following steps: time window segmentation is performed on the handle operation signal to generate operation feature sequence, and the vehicle body dynamic response signal is processed synchronously to generate response feature sequence; The attention weight distribution of the response feature sequence is calculated based on the operation feature sequence, and the road loss of control coefficient is used as a dynamic adjustment factor to dynamically correct the attention weight strength; Finally, according to the weighted and fused feature sequence, the dynamic consistency degree of the operation intention and the actual response of the vehicle is calculated through vector similarity, and the matching degree score of the operation intention and the vehicle response is output, wherein the weight proportion of the response feature is automatically enhanced in the low adhesion force road scene.

[0009] Further, the output includes a fusion risk assessment matrix containing the collision probability gradient and the road loss of control coefficient, including the following steps: dividing the fan-shaped area radially in the dynamic risk vector field along the polar coordinate system, extracting the maximum risk vector intensity in each area, and calculating the collision probability gradient combined with the relative speed of the moving target; At the same time, according to the real-time deviation degree of the maximum lateral adhesion coefficient and the road grip force, the vehicle loss of control risk coefficient is dynamically quantified; Then map the collision probability gradient to the Cartesian coordinate system grid, and nonlinearly couple with the road loss of control coefficient at the same grid position to form a spatial evaluation matrix containing collision risk and loss of control risk, wherein the high-risk area presents a gradient transition feature.

[0010] Further, difference control is performed according to the direction priority of the fusion risk assessment matrix and the behavior label level, when the behavior label is passive loss behavior, the anti-skid vector control is started based on the gradient direction of the collision probability, including the following steps: extracting the peak area azimuth of the collision probability gradient in the fusion risk assessment matrix, combining the vehicle body roll angular velocity change trend, determining the dominant direction of vehicle slip; dynamically allocating the output torque difference of the double wheel hub motor according to the dominant direction, reducing the torque output of the high-risk side wheel, keeping or increasing the torque output of the low-risk side wheel, forming a correction torque to suppress skidding; real-time monitoring of the vehicle body posture recovery state, when the lateral acceleration returns to the safety threshold, gradually releasing the torque difference allocation and restoring symmetric driving.

[0011] Further, when the behavior label is active risk behavior and the operation direction coincides with the high-risk area, the progressive power limitation is triggered, and the spatial features and operation features of the intervention scene are recorded in real time, including the following steps: comparing the handle steering angle with the high-risk area azimuth of the fusion risk assessment matrix, calculating the angle deviation between the operation direction and the risk core area; when the angle deviation is less than the critical angle, starting three-stage progressive power attenuation, maintaining the current power output in the initial stage, linearly reducing the motor power in the middle stage, and stabilizing at a safe power level in the final stage; when the angle deviation is greater than the critical angle, only the vibration warning is triggered; intercepting the environmental spatial data and operation time sequence data before and after the intervention, extracting the feature combination including static obstacle distribution form, operation peak intensity and power attenuation curve.

[0012] Further, according to the characteristics of the intervention scene, mirror virtual training tasks are generated, and parent operation data is converted into individualized safety thresholds through cluster analysis, including the following steps: based on the recorded static obstacle distribution form and road surface adhesion parameters, reconstructing a three-dimensional risk scene in a virtual environment, presetting challenge nodes of the same type as the original intervention event in the mirror scene, including sharp turn obstacle avoidance and wet road speed control; collecting operation trajectory data of parents completing the training task, extracting steering smoothness and acceleration stability as feature vectors, identifying safety operation mode clusters through density clustering algorithm, extracting cluster center features as ideal operation templates, mapping key parameters of ideal operation templates to vehicle control thresholds, including maximum allowed roll angle and minimum adhesion coefficient requirement on curves.

[0013] Further, the determination criteria of the dynamic update behavior analysis module and the control parameters of the safety decision module include the following steps: converting the parent turning smoothness range into a matching degree score threshold, adjusting the weight proportion of the road uncontrolled coefficient in the attention calculation according to the curve adhesion requirement; writing the maximum allowed roll angle into the safety decision module as the angular velocity threshold of the passive loss behavior triggering the anti-sideslip control, obtaining the parent acceleration stability data, reconstructing the three-stage attenuation curve slope of the progressive power limitation, collecting the actual intervention frequency and scene type distribution after the new threshold takes effect, and feeding back to the optimization module to start the threshold iteration adjustment.

[0014] The present application has the following advantages:

[0015] (1) The child electric motorcycle intelligent safety control system based on the Internet of Things significantly improves the safety decision accuracy in complex scenarios through dynamic risk vector field modeling and environment-behavior collaborative analysis. The risk modeling module integrates mobile obstacle trajectory prediction, static obstacle three-dimensional profile and road adhesion coefficient, breaks through the limitations of traditional single distance threshold judgment, realizes multi-dimensional risk space quantization, and effectively avoids misjudging normal turning on low adhesion road as dangerous operation. The behavior analysis module introduces the road uncontrolled coefficient as a dynamic weight factor of the cross-attention mechanism, accurately distinguishes between children's subjective risk intention and vehicle passive loss state, and solves the risk omission problem caused by behavior recognition mispositioning of existing systems.

[0016] (2) The child electric motorcycle intelligent safety control system based on the Internet of Things realizes the collaborative optimization of safety protection and child growth through hierarchical dynamic control and capability evolution closed loop. The safety decision module performs differentiated control according to the spatial coupling relationship between risk direction and behavior label: starts the direction adaptive anti-sideslip vector control when passive loss occurs, and triggers the progressive power limitation when active risk and direction coincide, maximizes the riding freedom degree under the premise of ensuring safety; the optimization module mirrors the real intervention scene to a virtual training task, generates individualized safety threshold through clustering analysis of parent operation data, dynamically updates the system determination criteria and control parameters, breaks through the limitations of fixed threshold on child capability development, and forms an adaptive closed loop of "risk intervention-capability evaluation-strategy evolution".

[0017] Of course, implementing any product of the present application does not necessarily require all the advantages described above. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The flowchart of the child electric motorcycle intelligent safety control system based on the Internet of Things of the present application. DETAILED DESCRIPTION

[0019] The embodiments of the present application solve the problems of the traditional safety control scheme, such as response lag to complex dynamic environment, non-precise risk quantification, high behavior recognition misjudgment rate and lack of personalized closed-loop optimization, through the child electric motorcycle intelligent safety control system based on the Internet of Things.

[0020] The scheme in the embodiments of the present application has the following general idea:

[0021] The dynamic risk vector field is constructed by using mobile obstacle trajectory prediction, road adhesion coefficient analysis and static obstacle three-dimensional profile, so as to realize the simultaneous quantification of collision probability in each direction and road out-of-control risk.

[0022] The handle operation intention and vehicle body dynamic response are deeply aligned through the cross-attention network on the vehicle side, and the road out-of-control coefficient is introduced as the weight, so as to realize the high-confidence distinction of active risk-taking and passive out-of-control behavior.

[0023] According to the direction priority and behavior label of the risk assessment matrix, the vectorized anti-skid control or the gradual power restriction is triggered respectively, so as to ensure the intervention timeliness and take into account the riding comfort.

[0024] The real-time recorded intervention scene is mapped to the digital twin platform, the personalized safety threshold is generated by combining the parent operation data clustering, the end-to-cloud closed loop is formed, and the behavior judgment benchmark and control parameters are dynamically updated.

[0025] Please refer to Figure 1The embodiment of the present application provides a technical scheme: a child electric motorcycle intelligent safety control system based on the Internet of Things, comprising the following modules: a risk modeling module, a behavior analysis module, a safety decision module, and an optimization module; the risk modeling module is used to generate a dynamic risk vector field centered on a vehicle by fusing mobile obstacle trajectory prediction data, road adhesion coefficient analysis data, and static obstacle three-dimensional contour data, and output a fusion risk assessment matrix containing a collision probability gradient and a road out-of-control coefficient; the behavior analysis module is used to receive handle operation time sequence features and vehicle body dynamic response features from vehicle-mounted sensors, and analyze the matching degree of operation intention and vehicle response through cross attention, wherein the road out-of-control coefficient is used as a weight factor to participate in the matching degree calculation, and a three-level behavior label of active risk behavior, passive loss behavior, or normal operation behavior is output; the safety decision module is used to perform differentiated control according to the direction priority of the fusion risk assessment matrix and the behavior label level, start anti-skid vector control based on the collision probability gradient direction when the behavior label is passive loss behavior, trigger progressive power limitation when the behavior label is active risk behavior and the operation direction coincides with a high-risk area, and simultaneously record the spatial features and operation features of the intervention scene in real time; and the optimization module is used to generate a mirror virtual training task according to the intervention scene features, convert parent operation data into individualized safety thresholds through cluster analysis, and dynamically update the judgment criteria of the behavior analysis module and the control parameters of the safety decision module.

[0026] In this embodiment, the risk modeling module: this module first obtains three types of input data: mobile obstacle trajectory prediction data, based on radar or visual sensor, the historical trajectory of dynamic targets such as surrounding pedestrians and vehicles is time series fitted, and the future movement trend is predicted; Road adhesion coefficient analysis data, through the tire mechanics model combined with environmental sensors (such as temperature, humidity, illumination) to output the friction performance index of different road sections (dry, wet, etc.); Static obstacle three-dimensional contour data, using the point cloud obtained by the depth camera, after point cloud segmentation and clustering, the three-dimensional boundary of the surrounding static objects (utility poles, flowerpots, etc.) is reconstructed. Map the above three types of information to the vehicle coordinate system, and fuse to form a "dynamic risk vector field" through vector kernel density estimation method, which gives the probability gradient of collision in each direction: representing the probability rate of change of collision between the vehicle moving in this direction and the dynamic obstacle; Road loss coefficient: quantifying the possibility of vehicle slip or rollover in this direction due to insufficient adhesion coefficient. Finally output "fusion risk assessment matrix", that is, in multiple sampling directions and distance grids, the above two risk indicators are presented at the same time for subsequent module calling. Dynamic risk vector field: directional risk distribution with vehicle centroid as origin, taking into account the comprehensive influence of dynamic and static obstacles and road conditions; Collision probability gradient: in the risk field, the gradient size of the collision probability along a certain vector direction changes with distance or time; Road loss coefficient: based on the friction characteristics of tires and road, a numerical index is predicted to predict the possibility of vehicle side slip or loss of control in a particular direction. Behavior analysis module: receive two types of time series signals collected by vehicle-mounted sensors: handle operation time series features, throttle, brake and steering handle opening / angle change sequence over time; Vehicle body dynamic response characteristics, three-axis acceleration and angular velocity time series collected by IMU (inertial measurement unit). First, extract the time-frequency domain features (mean, variance, spectral entropy, kurtosis, etc.) of the two, forming two groups of feature vectors; Build cross attention network: take operation features as reference Query, and response features as Key / Value; Introduce the road loss coefficient from the "risk modeling module" as an attention weight factor in attention calculation, to enhance the focus on high-risk direction signals; Output behavior matching degree distribution, and according to the matching result and the preset strategy, mark as: active risk behavior (strong matching between driving intention and high-risk direction); Passive loss behavior (intention deviates from response or response points to high loss coefficient direction); Normal operation behavior (intention and response are consistent and risk is low). Cross attention network: a commonly used mechanism in Transformer architecture, used to measure the correlation between the two groups of features; Spectral entropy: an index describing the uniformity of signal frequency spectrum distribution, reflecting the complexity of action change; Matching degree distribution: the quantitative output of operation-response alignment degree in different directions or scenarios.The safety decision module receives the "fusion risk assessment matrix" from the risk modeling module and the "three-level behavior label" from the behavior analysis module; according to the "direction priority", that is, the intervention is performed on the direction with the highest risk first, and the "behavior level" is differentiated for control: passive loss behavior: extract the direction with the largest collision probability gradient, generate a vectorized anti-skid instruction, and separately close-loop distribute the front / rear wheel braking and driving force to quickly correct the vehicle body posture; active risk behavior: when the operation direction coincides with the high-risk grid (high collision probability, low adhesion coefficient), start the gradual power limitation, gradually reduce the motor output through a smooth curve to avoid the secondary risk caused by sudden stop; after each intervention, the current geographical position, environmental parameters and operation sequence are recorded in real time for the optimization module. Direction priority: in a multi-direction risk field, the collision probability gradient or the loss of control coefficient is sorted, and the most dangerous direction is selected for control; gradual power limitation: a non-sudden power reduction strategy, which realizes smooth transition through a multi-step decreasing curve to improve the riding comfort of children. The optimization module generates a mirror virtual training task on the digital twin platform based on the "intervention scene features" and "operation features" reported by the safety decision module, automatically replays various environmental disturbances (such as adhesion coefficient change and obstacle position offset); the actual operation data of parents are subjected to density clustering and time sequence pattern mining, different driving style clusters are induced, and corresponding "individualized safety threshold" is extracted; the above threshold and model updating strategy are packaged and distributed to each vehicle edge unit, and the behavior analysis module judges the dynamic update of the benchmark; the safety decision module controls the real-time iteration of the parameters. Digital twin platform: a cloud system that replicates real vehicles and scenes in a simulated environment to verify and optimize control strategies; density clustering: an unsupervised learning method for extracting high-frequency behavior patterns from multi-dimensional operation data; individualized safety threshold: a set of risk trigger parameters customized according to the driving characteristics of different children / parents.

[0027] Specifically, by fusing mobile obstacle trajectory prediction data, road adhesion coefficient analysis data and static obstacle three-dimensional contour data, a dynamic risk vector field centered on the vehicle is generated, including the following steps: first, time series modeling is performed on the mobile target point cloud collected by the radar, the probability distribution of the multi-path motion trajectory is deduced, and the three-dimensional topological map of the environment is constructed through deep vision, the physical properties of static obstacles are identified through semantic understanding, and the passing feasibility evaluation of each direction is output, then the tire pressure distribution and vehicle body posture change data are fused to dynamically analyze the maximum lateral grip force critical value of the current road; finally, the mobile target trajectory probability distribution, static obstacle passing evaluation and road grip force critical value are spatially vector fused to form a directional risk field, wherein each vector points to the direction of the risk source, and the vector length reflects the risk intensity level.

[0028] In this embodiment, the radar collected moving target point cloud is time series modeled, and the probability distribution of multi-path motion trajectory is deduced. First, the target point cloud sequence is obtained by the millimeter wave radar or laser radar, the spatial clustering and ID tracking of each moving object are performed, and the continuous time series trajectory data are formed. The future position distribution of the target is predicted based on the Bayesian filtering or RNN structure, and the multi-path behavior assumption is introduced: formula: Parameter description: θ: direction angle in the vehicle body coordinate system; r: radial distance of the target relative to the vehicle; π k : probability weight of the kth motion mode (such as straight running, turning, deceleration); The prediction probability of the target appearing at (θ, r) in the kth mode; K: total number of motion modes. The formula outputs the probability density of the future appearance of the moving obstacle in each direction and distance, which is used to represent the potential collision possibility. The three-dimensional topological graph of the environment is constructed by the deep vision, and the physical characteristics of the static obstacle are identified by the semantic understanding, and the passing feasibility in each direction is output. The vehicle-mounted deep vision system (such as stereo vision or structured light) collects the environment point cloud, adopts the SLAM method to construct the dense three-dimensional topology, and identifies the static object categories and attributes through the semantic segmentation model (such as DeepLabv3+). In the vehicle around, the passing feasibility score is constructed as a variable of direction angle: formula: Parameter description: E static (θ): static passing feasibility score of direction θ; N θ : number of static objects identified in the direction; s i : semantic category of the ith obstacle (“wall”, “fence”, “greening”); η(s i ): passing influence factor of the semantic category (“greening” is low, and “wall” is high); d i : relative distance of the obstacle to the vehicle; δ(d i ): distance penalty function, the weight is higher as the distance is closer, which can be defined as The tire pressure distribution and the vehicle body posture change data are fused, and the maximum lateral grip force critical value of the current road surface is dynamically analyzed. The tire pressure sensor provides the current pressure value of each tire; the IMU sensor provides the pitch angle and roll angle of the vehicle. According to the current dynamic load distribution of the vehicle, the effective ground contact area of each tire is estimated, and the maximum available lateral adhesion is judged combined with the road surface type: formula: Parameter description: μ(θ): road adhesion coefficient (i.e. maximum lateral grip force critical value) in direction θ; F y,max (θ): maximum available lateral force; F z (θ): normal load borne by the tire in direction θ; A contact(θ): actual contact area, varies with tire pressure and load; σ(θ): lateral friction capacity per unit area, influenced by road type; m: vehicle mass; g: gravitational acceleration; φ: roll angle, affects tire normal pressure distribution. This lateral grip threshold decreases rapidly in low adhesion coefficient or dramatic attitude change, and is an important input for identifying the risk of losing control. The three types of data described above are fused into a directional risk field. Based on the foregoing data, a risk fusion function is established to output a risk intensity vector for each direction θ. The three factors of moving targets, static obstacles, and road grip are weighted and calculated: Formula: Parameter description: R(θ): risk intensity of direction θ; P collision (θ): integrated collision probability of moving obstacles in the direction, defined as where w r (r) is the distance weight function; E static (θ): passing feasibility score of direction θ; μ(θ): maximum lateral grip threshold of direction θ; μ min , μ max : empirical or simulated grip value boundary; w1, w2, w3: weight coefficients of the three risk factors, satisfying w1+w2+w3=1. The final output vector is where represents the unit vector of the direction. A full directional vector field can be constructed around the vehicle for control decision.

[0029] Specifically, the output includes a fusion risk assessment matrix containing collision probability gradient and road loss coefficient, including the following steps: dividing the fan-shaped area in the dynamic risk vector field along the polar coordinate system, extracting the maximum risk vector intensity in each area, and calculating the collision probability gradient combined with the relative speed of the moving target; at the same time, according to the real-time deviation degree of the maximum lateral adhesion coefficient and the road grip, the vehicle loss risk coefficient is dynamically quantified; then the collision probability gradient is mapped to the Cartesian coordinate system grid, and the road loss coefficient at the same grid position is nonlinearly coupled to form a spatial evaluation matrix containing collision risk and loss risk, where the high-risk area presents a gradient transition feature.

[0030] In this embodiment, the fan-shaped area is divided in the dynamic risk vector field along the polar coordinate system, and the maximum risk vector intensity in each area is extracted. In the polar coordinate system with the vehicle centroid as the origin, the surrounding space is divided into M fan-shaped areas (angle interval Δφ, radius interval Δρ). In each sector, the risk vector field is traversed, and the maximum vector modulus is extracted, which represents the local risk aggregation degree of the area: Formula: Parameter description: Λm,n: Maximum risk vector intensity in the m,nth sector; V(ρ,φ): Risk vector at the polar coordinate position (ρ,φ); ρn: Starting value of the nth radial interval; φm: Starting value of the mth angular interval. Calculate the collision probability gradient based on the relative velocity of the moving target. For each moving target in a sector, the collision risk gradient is estimated based on its radial velocity component relative to the vehicle. The greater the speed and the direction toward the vehicle, the higher the risk. After integrating multiple targets, define the sector collision probability gradient: Formula: Parameter description: Γm,n: collision probability gradient in the mth and nth sectors; Qm,n: number of moving targets in the mth and nth sectors; ξq: relative radial velocity of the qth target (positive value towards the vehicle); τq: estimated time for the target to reach the reference line in front of the vehicle; τ0: time scale coefficient, which controls the attenuation of long-term collisions; Zm,n: normalization factor, which is used to adjust the dimensional uniformity. This formula reflects that a target approaching the vehicle at high speed in a short period of time will significantly increase the collision gradient of the sector. According to the degree of deviation between the maximum lateral adhesion coefficient and the real-time grip, the vehicle loss of control risk coefficient is dynamically quantified. The current lateral acceleration monitored by the sensor is compared with the estimated maximum lateral adhesion limit to obtain the critical degree of loss of control. The loss of control coefficient on each sector is defined as follows: Formula: Parameter Description: m,n : Road loss of control coefficient of m, n sectors; γ m,n : The current actual lateral acceleration of the sector; The maximum lateral acceleration allowed under the road conditions in this sector. The square of this ratio is used to amplify the high-risk area signal, showing a nonlinear growth characteristic. The collision probability gradient is mapped to the Cartesian coordinate grid and nonlinearly coupled with the road out-of-control coefficient at the same grid position. The sector risk in polar coordinates is mapped to an equally divided Cartesian grid (x u ,y v ), match the corresponding out-of-control coefficient, perform nonlinear fusion, and generate the fusion risk assessment matrix F(u,v). The fusion process considers cross terms and nonlinear amplification coefficients: Formula: Parameter description: F(u,v): Cartesian grid points (x u ,y v ) on the fusion risk value; Collision probability gradient obtained from polar sector interpolation mapping; The risk coefficient for loss of control is derived from the polar coordinate sector interpolation mapping. α and β are fusion weight indices that control the relative influence of the two types of risk. κ is the nonlinear interaction coefficient that regulates the synergistic amplification effect of risk. This formula ensures the generation of a "risk transition" characteristic in the intersection of high collision and high loss of control, manifested as a high-gradient transition zone in the matrix, which serves as an important input for subsequent control strategy determination.

[0031] Specifically, the handle operation timing features and the vehicle body dynamic response features from the vehicle sensors are received, and the matching degree of the operation intention and the vehicle response is analyzed through cross attention, wherein the road surface loss of control coefficient is used as a weight factor to participate in the matching degree calculation, including the following steps: time window segmentation is performed on the handle operation signal to generate an operation feature sequence, and the vehicle body dynamic response signal is processed through synchronous alignment to generate a response feature sequence; the attention weight distribution of the response feature sequence is calculated based on the operation feature sequence, and the road surface loss of control coefficient is used as a dynamic adjustment factor to dynamically correct the attention weight strength; finally, the dynamic consistency degree of the operation intention and the actual vehicle response is calculated through vector similarity based on the weighted and fused feature sequence, and the matching degree score of the operation intention and the vehicle response is output, wherein the weight proportion of the response feature is automatically enhanced in the low adhesion road surface scene.

[0032] In the embodiment, the time window segmentation generates the operation feature sequence and the response feature sequence: the handle operation signal (the original time domain sequence of the opening or angle of the accelerator, brake, steering, etc. with time) and the vehicle body dynamic response signal (the acceleration and angular velocity sequence collected by the IMU with time) are first segmented by the same length or overlapping sliding window. In each time window, the signal is preprocessed (such as denoising, normalization), and feature extraction (such as time domain and frequency domain features) is performed to obtain fixed-dimension operation feature vectors and response feature vectors. Synchronous alignment: ensure that the pth operation window and the pth response window of the two sequences are consistent in time range, so as to match subsequently. Formula: X p = FeatOp(SegOp[t p ,t p +ΔT]), Y p = FeatRsp(SegRsp[t p ,t p +ΔT]); Parameter description: X p : the operation feature vector extracted in the pth time window; Y p : the response feature vector extracted in the pth time window; SegOp[t p ,t p +ΔT]: the sampling sequence of the operation signal in the time interval [t p ,t p +ΔT]; SegRsp[t p ,t p+ ΔT]: the sampling sequence of the response signal in the same interval; ΔT: the length of the time window; FeatOp(·): the operation feature extraction function, which can include mean, variance, spectral entropy extraction module, etc.; FeatRsp(·): the response feature extraction function, which can include acceleration, angular velocity time-frequency domain feature extraction module; index p = 1, 2, …, P, representing the total number of windows divided in the current analysis period. Note: if the edge unit capacity allows, X p , Y p Further dimension reduction or projection processing is performed to unify the input dimension of subsequent attention calculation. The attention weight distribution of the response feature sequence is calculated based on the operation feature sequence: a cross-attention (CrossAttention) mechanism is constructed, taking the operation feature Xp as the "query" (Query) and the response feature Yq as the "key-value pair" (Key / Value). First, Xp and Yq are projected into the same attention space through linear transformation to obtain the projection vectors. Then the original attention score is calculated. The attention score is normalized between different response windows q to form the attention weight distribution, which represents the importance of each response window under the pth operation intention. Formula: U p = X p A, V q = Y q B, Parameter description: Xp: the pth operation feature vector (dimension can be denoted as Dop); Yq: the qth response feature vector (dimension can be denoted as Drsp); A: linear projection matrix with size Dop x D; B: linear projection matrix with size Drsp x D; Up: operation feature projection vector, dimension D; Vq: response feature projection vector, dimension D; D: attention space dimension; S p,q : original attention score of operation window p and response window q; scaling factor for balancing the numerical magnitude of point multiplication result; Ap, q: attention weight of operation window p to response window q (after normalization), satisfying ∑q=1QAp,q=1; index p = 1, …, P, q = 1, …, Q; usually P = Q, because the time windows are synchronized and aligned, but they can still be processed separately. The road uncontrolled coefficient is taken as a dynamic adjustment factor to correct the attention weight strength: the road uncontrolled coefficient sequence ζp (scalar) corresponding to the current time window is obtained from the upper level risk modeling module or the real-time grip evaluation module, representing the quantitative degree of the road adhesion ability and uncontrolled tendency of the vehicle in the pth time period. In the window with low adhesion (high uncontrolled risk), the influence of the response feature on the matching degree judgment needs to be enhanced, that is, the importance of the response vector is given higher weight when calculating the attention. The specific method is: the original attention score is dynamically amplified or scaled to obtain the corrected score S' p,q , and the final attention weight is generated after normalization Formula: S' p,q = S p,q ×(1+ηζ p ), Parameter description: ζp: the road uncontrolled coefficient corresponding to the pth time window, scalar; the larger the value, the higher the uncontrolled tendency (lower adhesion); η: adjustment coefficient (constant), used to control the strength of ζp amplifying attention score; S p,q : uncorrected attention score (see previous step); S' p,q : corrected attention score; The corrected and normalized attention weight. When ζp is large (low adhesion scene), (1+ηζp) amplifies the original score, making the attention distribution more inclined to respond to the feature, increasing the weight of the response in the matching degree calculation; if ζp is very small, it is close to the original attention. Weighted fusion of response features to get the fused feature under each operation window: for the pth operation window, according to the corrected attention weight Weighted sum of each response feature Y q to get the fused response representation C p . At the same time, X p and C p can be further combined for subsequent consistency measurement. Formula: Parameter description: C p : the fused response feature vector under the pth operation window, the dimension is consistent with Y q ; The corrected attention weight corresponding to the pth operation window; Y q : the qth response feature vector; index q = 1, …, Q. C p Emphasizes the more critical response feature part in the current operation intent and in the low adhesion scene, providing a focused semantic representation for subsequent similarity calculation. The consistency degree between the operation intent and the actual response is calculated by vector similarity, and the matching degree score is output. The similarity between the operation feature X p and the corresponding fused response feature C p is calculated in each time window to measure the consistency between the intent and the actual response. Cosine similarity or other distance measures can be used, and the design of increasing the weight of the response part in the low adhesion scene has been reflected in the previous step. Finally, the matching degrees of each time window can be aggregated (such as taking the average or weighted average), and the overall matching degree score is output, which is used to determine whether it is normal operation, active risk or passive loss of control. Formula: Parameter description: sim p : the cosine similarity score of the pth time window; X p : the pth operation feature vector; C p: the pth fused response feature vector; ||·||: vector norm (Euclidean norm); M: overall matching score, which can be used for subsequent threshold judgment; P: total number of time windows. When sim p is high, it indicates that the operation intention and the response are highly consistent; when sim p is low and accompanied by a high-risk scenario (high ζ p ), it can be determined as passive loss of control; when sim p is low but the risk scenario is low, it can be active risk. The automatic enhancement of the response feature weight in the low adhesion scenario is reflected by the attention score correction: when the road loss of control coefficient ζ p is large, the response feature occupies a higher proportion in the weighted fusion, so that more attention is paid to the inconsistency between the response and the intention in the similarity calculation. If it is necessary to further strengthen the influence of low adhesion in the final matching degree M level, a weighted average can be used when summarizing: Parameter description: p: adjustment coefficient when globally summarizing; ζ p : the pth window loss of control coefficient; sim p : the pth window similarity score; M': the overall matching score after weighting, which highlights the consistency change in the low adhesion scenario. This summary strategy is an optional enhancement design, which can be flexibly selected according to system requirements and resources.

[0033] Specifically, the three-level behavior labels of active risk behavior, passive loss of control behavior or normal operation behavior are output, including the following steps: when the matching score is greater than or equal to the first threshold and the vehicle body dynamic response is smooth, it is marked as normal operation behavior; when the matching score is less than the first threshold but the operation flow vector shows active steering or acceleration, it is marked as active risk behavior; when the matching score is less than the second threshold and the response flow vector shows abnormal yaw, it is marked as passive loss of control behavior; for the data marked as active risk behavior, it is verified whether the operation direction coincides with the high-risk area in the fused risk assessment matrix, if it coincides, the label is maintained, otherwise it is downgraded to normal operation behavior, combined with the classification benchmark and the environment verification result, the final three-level behavior label is output.

[0034] In this embodiment, the matching score is calculated and the normal operation behavior condition is judged: the matching score of each time window has been obtained (corresponding to the similarity or weighted similarity described above), and the vehicle body dynamic response stability metric wherein, the vehicle response fluctuation degree can be evaluated based on the variance or yaw rate of the response feature sequence. Let the first threshold constant be C norm , and the response stability threshold be S stab When and At this time, it is considered that the current window operation intention is highly consistent with the response and the response is smooth, and it can be marked as "normal operation behavior". Formula: if then label q = normal operation; Parameter description: Matching score of the qth window; C norm : Normal operation determination threshold (constant); Response stability measure of the qth window; S stab : Response stability threshold (constant); ∧: logical and operator; Label q: preliminary behavior label of the corresponding window. Note: It can be defined as the variance of the response acceleration or yaw rate sequence, or the average yaw rate in the sliding window, and the purpose is to quantify the degree of "smoothness" or "fluctuation". The preliminary condition for judging active risk behavior is when (matching degree decreases), it is necessary to judge whether it is active risk: check whether the "operation flow vector" shows an active steering or acceleration tendency. Define the operation flow vector O q of the qth window, which can be represented by the steering rate or throttle acceleration extracted from the operation characteristics, and if the vector changes significantly in size or direction within the current time window, it indicates that the driver has an active adjustment intention. Set the active intention threshold A thr When and ||O q || ≥ A thr or some components (such as the acceleration component) exceed the preset range, it is preliminarily marked as "active risk behavior". Formula: if then the preliminary label q = active risk. Parameter description: O q : Operation flow vector of the qth window; ||O q ||: Vector norm, representing operation intensity; A thr : Active risk determination threshold (constant); Matching score; C norm : Normal operation threshold; Preliminary label q: label candidate obtained at this step. If the operation flow vector dimension includes steering and acceleration components, further threshold judgment can be performed on specific components, such as acceleration component exceeding a certain value or steering rate being too large. The preliminary condition for judging passive loss of control behavior is when (lower matching degree, second threshold C loose ), and the vehicle response flow vector shows an abnormal yaw signal, it is preliminarily marked as "passive loss of control". Define the response flow vector R q , which can be represented by the yaw rate or lateral acceleration extracted from the response characteristics. If ||R q || or some yaw components exceed the loss of control threshold L thr , it is judged as loss of control. The second threshold C loose and the loss of control threshold L thr are used together: if and ||R q ||≥L thr , then the preliminary label is "passive loss of control". Formula: if then the preliminary label q = passive loss of control. Parameter description: R q : response flow vector of the qth window; ||R q ||: vector norm, representing the response abnormal intensity; L thr : loss of control threshold (constant); matching degree score; C loose : passive loss of control matching degree threshold (constant, lower than C norm ); preliminary label q: label candidate obtained at this step. Check the coincidence of the operation direction and the high-risk area: for each window q with the preliminary label of "active risk-taking", further check whether its operation direction is truly in the high-risk area to avoid misjudgment. Extract the high-risk direction vector set of the qth window from the risk assessment matrix or dynamic risk vector field , and define the operation direction vector O q of the window. Calculate the cosine of the angle or the projection ratio of O q and each high-risk vector . If there exists a such that , it is considered that the operation direction coincides with the high-risk area, and the label is maintained as "active risk-taking"; otherwise, the label is downgraded to "normal operation". If the preliminary label q = active risk-taking, and there exists a such that , then the final label q = active risk-taking; otherwise, the final label q = normal operation. Parameters: O q : operation direction vector of the qth window; high-risk direction vector set of the qth window; rth high-risk direction vector in the set; ||·||: vector norm; Δ thr : direction coincidence threshold (constant, between 0 and 1); ··: vector inner product operation; "exists" logic means that as long as the coincidence degree with a high-risk vector is sufficient, the risk-taking label is maintained; otherwise, it is downgraded; final label q: label result after verification. The high-risk direction set can be selected from the direction corresponding vector in the risk assessment matrix whose risk intensity exceeds a certain threshold; the threshold is Δ thrIt can be dynamically adjusted in the optimization module. Combining the classification benchmark with the environmental verification results, the final three-level behavior labels are output: for the downgraded "active risk" window, the "active risk" label is retained; for the windows that are judged as "normal operation", properly trigger "passive loss of control" and have not been further verified and downgraded, the corresponding labels are output respectively; at the same time, additional environmental verification can be combined: for example, the current risk level, abnormal impact or slip event verification detected by the inertial sensor. If the environmental verification result is inconsistent with the preliminary label, it can be re-corrected (for example, in extreme environments, even if the match is high, it may be judged as out of control due to external emergencies). Finally, the labels of all time windows are output in time series, or summarized into paragraph labels within the sliding analysis cycle (such as several consecutive windows with the same label are regarded as a behavior segment) for use by the safety decision module. The system reads the "final label" of each window and compares it with the environmental verification signals within the same analysis cycle (such as sudden braking impact, abnormal vibration sensor alarm, etc.); if the environmental verification indicates that there is a sudden loss of control, it can be upgraded to "passive loss of control" even if the previous label may be normal; conversely, if external verification clearly indicates safety, the original label is retained; the output format can be a time series label list or a behavior paragraph summary, so that the decision module can distinguish the behavior type and implement the corresponding intervention strategy.

[0035] Specifically, differentiated control is performed based on the directional priority and behavior label level of the fused risk assessment matrix. When the behavior label is a passive loss of control behavior, anti-skid vector control is initiated based on the direction of the collision probability gradient, including the following steps: extracting the peak area azimuth of the collision probability gradient in the fused risk assessment matrix, and determining the dominant direction of vehicle slip in combination with the trend of vehicle body yaw angular velocity change; dynamically allocating the output torque difference of the dual-hub motors according to the dominant direction, with the wheels on the high-risk side reducing torque output and the wheels on the low-risk side maintaining or increasing torque output to form a corrective torque to suppress skidding; real-time monitoring of the vehicle posture recovery status, and when the lateral acceleration returns to the safety threshold, gradually releasing the torque difference distribution and restoring symmetrical drive.

[0036] In this embodiment, the peak area azimuth of the collision probability gradient in the fusion risk assessment matrix is ​​extracted: from the fusion risk assessment matrix (Cartesian grid form), for the collision probability gradient part, find the grid area with the largest gradient value. Let the collision probability gradient be in the Cartesian coordinate grid (x u ,y v ) is recorded as the matrix H u,v Traverse all grids, select the grid point or several peak areas corresponding to the maximum value, and calculate its azimuth. The azimuth is calculated in the vehicle coordinate system based on the vector from the vehicle's center of mass to the grid point, representing the direction with the greatest potential collision risk. Formula: Parameter Description: H u,v : Cartesian grid points (xu ,y v ) on the collision probability gradient value; (u * ,v * ): grid index that makes H u,v reach the peak value; lateral and longitudinal coordinates of the corresponding peak grid point in the vehicle coordinate system; χ: extracted peak region azimuth angle (with the front of the vehicle as the reference zero degree, the right side as the positive direction); atan2(y, x): azimuth angle function of point (x, y) relative to the vehicle center of mass. Note: If there are multiple local peaks, multiple sets of (u * ,v * ) and corresponding χ can be calculated, and the most significant one is selected in combination with the yaw trend. In combination with the change trend of the body yaw rate, the dominant direction of vehicle slip is determined: the body yaw rate is provided by the vehicle-mounted IMU, denoted as the sequence ω(t) changing with time. In the passive loss of control scene, the change trend needs to be analyzed to determine the slip direction: if the yaw rate increases continuously in a short time

[0037] , it indicates that the vehicle is sliding to one side. The time derivative of the yaw rate is calculated (which can be obtained by filtering and difference or sliding window fitting), and the slip is to the left side or the right side in combination with the current yaw rate sign. Finally, the dominant slip direction vector is obtained, which is combined with the collision risk direction χ extracted in the previous step to correct the torque distribution. Formula: Parameter description: ω(t): current body yaw rate; Δt: time interval for difference calculation; approximate time derivative of the yaw rate, indicating the yaw change trend; s: dominant slip direction identifier; +1 indicates a right-side slip trend; 1 indicates a left-side slip trend; 0 indicates that the slip trend is not obvious or in a critical recovery stage; "other cases" can include scenarios where ω and the signs are not consistent or the values are close to zero. Note: In actual implementation, ω(t) can be low-pass filtered to remove noise, and then differentiated or fitted with a slope; if there are multiple peak azimuth angles χ, the most matched one can be further selected in combination with the slip direction s. According to the dominant direction, the output torque difference of the double-wheel hub motors is dynamically distributed: the vehicle is equipped with left and right double-wheel hub motors, which can generate a correction torque by asymmetrically distributing the torque of the left and right motors to suppress the side slip. According to the risk direction χ and the slip direction s, the high-risk side wheel and the low-risk side wheel are determined: if the risk peak direction χ points to the right and the slip direction s = +1, the right wheel is considered as the high-risk side and the torque needs to be reduced; the left wheel is the low-risk side and the torque can be maintained or slightly increased; vice versa. Define the baseline torque τ0. According to the coupling of the slip intensity and the risk intensity , the torque difference Δτ is calculated and distributed to the left and right motors. Formula: Parameter description: Δτ: torque difference absolute value, used for left and right distribution; κ1: slip intensity coefficient, used to map the yaw change to torque correction; κ2: risk sensitivity coefficient, used to amplify or adjust the influence of risk gradient value in torque distribution; extracted collision probability gradient peak value; τ0: baseline driving torque (both sides of the motor are this value when there is no side slip intervention or recovery stage); τ left , τ right : real-time torque allocated to left / right motor; s: slip dominant direction identifier, +1 indicates right side slip risk → right side reduces torque (so left side increases); 1 indicates left side slip risk → left side reduces torque; exp(·): exponential function, used for nonlinear amplification of risk intensity; "left and right" definition is based on vehicle coordinate system: right side is the right side of the passenger. Form a correction moment to suppress side slip and issue control instructions: τ left and τ right are issued to left and right hub motor drive units, and torque difference distribution is performed through control motor output to generate a lateral correction moment to resist the slip tendency. At the same time, ω and risk gradient H can be updated in real time within the control cycle and iteratively adjusted to ensure continuous suppression. Real-time monitoring of vehicle body posture recovery state and gradual release of torque difference distribution: continuously monitor vehicle body lateral acceleration a lat (t) and yaw rate ω(t) change after intervention, and determine whether the side slip is suppressed. Set a safe lateral acceleration threshold A safe . When |a lat (t)|≤A safe and |ω(t)| and When it returns to the small fluctuation range, it is considered that the side slip has been basically recovered. In the recovery stage, the torque difference Δτ is gradually reduced until τ left = τ right = τ0, and symmetric driving is restored. Formula: when Δτ new = μ·Δτ prev , τ left = τ0 + s·Δτ new , τ right = τ0 - s·Δτ new ; repeat the above update until Δτ new < ∈, then τ left = τ right = τ0. Parameter description: a lat (t): current lateral acceleration; A safe : safe lateral acceleration threshold; ω(t): yaw rate; Ω thr : yaw rate regression threshold; derivative of yaw rate; Yaw rate derivative regression threshold; Δτ prev : Torque difference of the last intervention cycle; Δτ new : Torque difference after reduction in the current cycle; μ: Recovery attenuation coefficient (0<μ<1), used to gradually reduce the torque difference; ∈: Torque difference termination threshold, when Δτ<∈, recovery is considered complete; τ0: Baseline driving torque; s: Slip dominant direction identifier.

[0038] Specifically, when the behavior label is active risky behavior and the operation direction coincides with the high-risk area, progressive power restriction is triggered, and the spatial characteristics and operation characteristics of the intervention scenario are recorded in real time, including the following steps: comparing the steering angle of the handle with the high-risk area orientation of the fusion risk assessment matrix, and calculating the angle deviation between the operation direction and the risk core area; when the angle deviation is less than the critical angle, starting the three-level progressive power attenuation, maintaining the current power output in the initial stage, linearly reducing the motor power in the middle stage, and stabilizing at a safe power level in the final stage; when the angle deviation is greater than the critical angle, only triggering a vibration warning; intercepting the environmental spatial data and operation time series data before and after the intervention, and extracting a feature combination including the distribution pattern of static obstacles, operation peak intensity, and power attenuation curve.

[0039] In this embodiment, it is determined whether the operation direction coincides with the high-risk area, the current operation direction, i.e., the steering angle of the handle, is extracted, and is recorded as the absolute azimuth angle in the vehicle reference coordinate system; at the same time, the main direction of the high-risk area identified at the current moment is extracted from the fused risk assessment matrix; the included angle deviation between the above two directions is calculated, if the deviation is less than the preset direction critical angle threshold, it is determined that the current operation has a clear trend of moving towards the high-risk area, and the power limiting control process is entered; if the included angle deviation is greater than the critical angle threshold, only vibration warning is triggered to remind, and power output is not intervened. Three-stage progressive power limitation is performed, after meeting the above trigger conditions, the system will gradually adjust the power output of the vehicle in three stages, as follows: Stage one: buffer maintenance stage, in the early stage of identifying risks, the current power output is temporarily kept unchanged, the purpose is to give the system and the user a buffer reaction time, to avoid sudden power drop causing riding discomfort or other risks. The duration of this stage is usually a few hundred milliseconds, which can be adaptively adjusted according to the vehicle speed and the age group of the driver. Stage two: linear decay stage, after entering this stage, the vehicle control system will gradually reduce the motor output power in a linear manner, so that it decreases from the current level to the preset safe power value; the power reduction rate can be dynamically adjusted according to the risk level or historical intervention data to ensure sensitive response and smooth process. Stage three: steady-state maintenance stage, when the motor output power decreases to the safe power, the system maintains the power unchanged to ensure that the vehicle is in a controllable state; at the same time, the environmental risk level and the change of user behavior are continuously monitored, if the risk is removed or the behavior label changes, the power limitation can be gradually removed. If the user's operation direction deviates from the core direction of the high-risk area, the system considers that it is not directly pointing to the dangerous area, and does not trigger power limitation to avoid interference with normal operation; but for the purpose of safety prompt, light vibration feedback such as handle short vibration can still be triggered by the control system to remind the driver to pay attention to the operation direction. The system will automatically record the following multi-dimensional features before and after the intervention process to support subsequent optimization control and model learning: environmental space features are extracted by camera, radar or visual SLAM sensing modules, including the spatial distribution form of static obstacles in front of and around the vehicle; the features include: obstacle type (such as railings, trees), distribution density, relative azimuth with the vehicle, etc. Operation timing features, collect the time sequence curves of handle steering angle, acceleration / brake signals before and after triggering; extract key feature indicators such as steering rate peak value, acceleration change abrupt point, etc. Record the whole process power change trajectory from the start of the progressive power limitation to the end of the steady-state stage; extract the stage division time point, linear decline rate, safe power level, etc. feature parameters for evaluating intervention effect and model optimization.

[0040] Specifically, according to the intervention scene characteristics, a mirror virtual training task is generated, and the parent operation data is converted into an individualized safety threshold through cluster analysis, including the following steps: based on the recorded static obstacle distribution form and the road surface adhesion parameters, a three-dimensional risk scene is reconstructed in a virtual environment, challenge nodes of the same type as the original intervention event are preset in the mirror scene, including sharp turn obstacle avoidance and wet road speed control; the operation trajectory data of the parents completing the training task is collected, the turning smoothness and acceleration stability are extracted as feature vectors, the safety operation mode cluster is identified through the density clustering algorithm, the cluster center features are extracted as the ideal operation template, and the key parameters of the ideal operation template are mapped into vehicle control thresholds, including the maximum allowed roll angle and the minimum adhesion coefficient requirement on the curve.

[0041] In the embodiment, the mirror three-dimensional risk scene is reconstructed, and the system automatically identifies the environmental elements at the time of the event based on the spatial data recorded in the foregoing intervention event, including: the distribution form of static obstacles (such as the relative position and density of obstacles in space); the physical properties of the road surface (such as local adhesion coefficient variation, slope, etc.). Based on this, a mirror three-dimensional scene is constructed in a virtual simulation environment as a training task environment. At the same time, a number of challenge nodes are preset in the scene to simulate actual high-risk operation events, for example: sharp turn obstacle avoidance task: simulate the vehicle entering the curve at high speed and a sudden static obstacle appearing in the path; wet road speed control task: set a regional low adhesion road section and require stable passing. When the parents perform the training task in the mirror scene, the continuous operation data of the parents is collected, and the time sequence is recorded, including: the steering angle input curve; the acceleration / braking force distribution; the vehicle body posture change. And the following key operation feature vectors are extracted therefrom: turning smoothness feature (denoted as η s ): representing the mean square error of the steering angle change per unit time, defined as: wherein: T: total number of operation samples; θ k : steering wheel steering angle of the kth sample; average value of the steering angle. Acceleration stability feature (denoted as ζ a ): representing the maximum deviation amplitude of acceleration change, used to evaluate the continuity and smoothness of power control: ζ a = max|a k -a k-1 |; wherein: a k : longitudinal acceleration of the kth sample; a k-1 : acceleration at the previous sampling time. The above two features form a two-dimensional feature vector group for the clustering model. Based on the density clustering, the safety operation mode is extracted from the collected parent training operation feature set {ζ1, ζ2,..., ζ N}application density clustering algorithm (such as DBSCAN) is used to divide and identify high-density clusters of operation modes in the feature space. Let the main cluster corresponding to the safe behavior in the clustering result be C safe , the center feature vector of the cluster is extracted as the ideal safe operation template Step four: mapping the operation template into personalized control threshold According to the ideal template feature extracted, a mapping relationship is established to convert it into a personalized safety threshold that can be called by the vehicle control system, including but not limited to: the maximum allowed roll angle threshold φ max , which is mapped from the steering smoothness feature , and a larger value corresponds to a stronger body posture control ability; the lower limit of the road adhesion coefficient μ min , which is mapped from the acceleration stability feature , is used to set the dynamic output limit trigger standard in the control system under wet road conditions. The mapping relationship can be completed using a linear or empirical model, for example: Where: κ1, κ2: system calibration constants; The upper limit of the acceleration instability is empirically set.

[0042] Specifically, the judgment criteria of the dynamic update behavior analysis module and the control parameters of the safety decision module include the following steps: converting the parent steering smoothness range into a matching score threshold, adjusting the weight proportion of the road loss coefficient in the attention calculation according to the road adhesion requirement; write the maximum allowed roll angle into the safety decision module as the angular velocity threshold for triggering the anti-skid control of passive loss behavior, obtain the parent acceleration stability data, and reconstruct the three-level decay curve slope of the progressive power limitation, after the new threshold takes effect, collect the actual intervention frequency and scene type distribution, and feedback to the optimization module to start the threshold iteration adjustment.

[0043] In this embodiment, the steering smoothness range is converted into a matching score threshold in the behavior analysis module, and the matching score is the core index for distinguishing the consistency of operation intention and vehicle response. The system maps the steering smoothness range extracted from the parent operation feature into the judgment interval of operation response matching degree to construct a new scoring threshold model: if ρp tends to be stable (i.e. small variance), a higher matching tolerance corresponds. The first updated scoring threshold γ1 and the second scoring threshold γ2 are estimated by the following empirical function: Where: a1, a2: system calibration coefficients, used to match the subjective handling stability and the objective score correspondence. Step two: adjust the weight proportion of the road loss of control coefficient in attention calculation The system takes the parent's control performance on the low adhesion road in the curve as input, analyzes the tolerance ability of the lateral stability, and then dynamically adjusts the participation proportion of the road loss of control coefficient Λr in the weight distribution of the attention mechanism in the behavior analysis module. The adjustment method is as follows: let the original attention weight distribution be A(u)_t, where u represents the time index of the response feature sequence; introduce a proportional adjustment factor δΛ, and update the weight formula as: Where: δΛ: fitted by the adhesion control ability of the parent, the enhancement amplitude of the control loss risk participation weight. Write the maximum allowed roll angle to the anti-skid control logic In the safety decision module, the maximum allowed roll angle is directly written into the control rule engine as the trigger condition of the anti-skid control: when the vehicle yaw rate and roll angle combination exceeds this threshold, it is determined as a potential loss of control state; the differential vector regulation logic of the wheel hub motor is activated to realize dynamic stability control. Reconstruct the three-stage attenuation curve slope of the progressive power limitation system takes the stability characteristics of the parent in the acceleration control as the benchmark, adjusts the output power curve slope of the three stages in the progressive power limitation: initial segment: maintain the original output unchanged; intermediate segment: the linear descending slope is set as: Where: β0: the system default mid-term power descending slope; reference stability level; β m : updated dynamic limitation slope. The constant output level of the final segment can be estimated by the lowest safe passing power in the training scene. The intervention effect is fed back to the optimization module to realize iterative optimization. After the new threshold takes effect, the system real-time statistics the following indexes: total number of intervention triggers; type distribution of intervention scenarios; vehicle posture recovery time after intervention; user operation compensation amplitude, etc. Take this as feedback to trigger the threshold re-evaluation task in the optimization module. If it is found that the intervention frequency of a certain type of scene abnormally rises, the personalized threshold is dynamically recalled, a new training mapping chain is constructed, and the behavior judgment and control logic are iteratively adjusted.

[0044] In summary, the present application has at least the following effects:

[0045] The child electric motorcycle intelligent safety control system based on the Internet of Things realizes spatial global perception of potential collision risk and road out-of-control risk by constructing a dynamic risk vector field integrating mobile obstacle prediction, static environment modeling and adhesion characteristic analysis, and improves the forward-looking identification ability of the system to dangerous sources in complex environments, thereby providing more targeted input basis for safety control. The cross-attention mechanism is adopted to couple and analyze the operation time sequence data and dynamic response data collected by the vehicle-mounted sensor, and based on the matching degree score and the weighted participation of the out-of-control factor, the three behavior modes of active risk-taking, passive out-of-control and normal operation are accurately distinguished, and the accuracy and response efficiency of the driving behavior identification are improved. The safety decision module combines the risk assessment direction priority and the behavior label level, adopts the anti-skid vector control and the progressive power limitation dual-path regulation logic, realizes the hierarchical intervention on the out-of-control behavior and the high-risk risk-taking behavior, effectively controls the accident risk diffusion, and guarantees the driving stability of children. Through the optimization module, a virtual training task based on the mirror reconstruction of the intervention scene is constructed, and the actual operation data of parents are introduced for clustering analysis, the individualized operation template is extracted, the behavior judgment and safety control parameters are updated in reverse, the intelligent control migration from "rule-driven" to "behavior-driven" is realized, and the adaptation ability of the system to individual differences is improved. The system automatically records the scene and behavior characteristics before and after each safety intervention, feeds back to the optimization module for threshold reevaluation, realizes the dynamic iterative adjustment of the control parameters, so that the system can continuously approach the optimal individualized safety configuration in the long-term operation process, and improves the robustness and evolution ability of the safety control strategy.

[0046] Those skilled in the art will appreciate that embodiments of the application can be provided as methods, systems, or computer program products. Accordingly, the application can be embodied in the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the application can be embodied in the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage media, etc.) having computer usable program code embodied therein.

[0047] The present application is described in reference to flowcharts and / or block diagrams of systems, apparatuses (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in a flow or multiple flows and / or blocks Figure 1 The functions specified in a flow or multiple flows and / or blocks

[0048] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions Figure 1 function specified in the flow Figure 1 block or blocks.

[0049] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions Figure 1 function specified in the flow Figure 1 block or blocks.

[0050] While the preferred embodiments of the application have been described, additional variations and modifications can be made to the embodiments by those skilled in the art once they learn of the basic inventive concepts. Therefore, it is intended that the appended claims shall cover all such modifications and variations as fall within the true spirit and scope of the application. Further, it is intended that each definition in the claims be read into each alternative of that claim.

[0051] Obviously, numerous modifications and variations of the present application are possible in light of the above teachings. It is therefore to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. The intelligent safety control system for children's electric motorcycle based on the Internet of Things is characterized by: Includes the following modules: risk modeling module, behavior analysis module, security decision module, and optimization module; The risk modeling module is used to generate a dynamic risk vector field centered on the vehicle by fusing mobile obstacle trajectory prediction data, road adhesion coefficient analysis data, and static obstacle three-dimensional contour data, and output a fused risk assessment matrix containing a collision probability gradient and a road loss of control coefficient; The behavior analysis module is used to receive the handle operation timing characteristics and vehicle body dynamic response characteristics from the vehicle sensor, analyze the matching degree between the operation intention and the vehicle response through cross-attention, wherein the road loss of control coefficient is used as a weight factor in the matching degree calculation, and output a three-level behavior label of active risky behavior, passive loss of control behavior or normal operation behavior; The safety decision module is used to perform differentiated control based on the directional priority and behavior label level of the fused risk assessment matrix. When the behavior label is a passive loss of control behavior, anti-skid vector control is initiated based on the collision probability gradient direction. When the behavior label is an active risky behavior and the operation direction coincides with a high-risk area, progressive power limitation is triggered. At the same time, the spatial characteristics and operation characteristics of the intervention scenario are recorded in real time. The optimization module is used to generate mirror virtual training tasks according to the characteristics of the intervention scenario, convert parent operation data into personalized safety thresholds through cluster analysis, and dynamically update the judgment criteria of the behavior analysis module and the control parameters of the safety decision module.

2. The intelligent safety control system for children's electric motorcycles based on the Internet of Things according to claim 1 is characterized by: By fusing the predicted trajectory data of moving obstacles, the road adhesion coefficient analysis data, and the three-dimensional contour data of static obstacles, a dynamic risk vector field centered on the vehicle is generated. The following steps are involved: First, the system performs time-series modeling on the moving target point cloud collected by the radar to deduce the probability distribution of multi-path motion trajectories. Simultaneously, it constructs a 3D topological map of the environment through deep vision. Combined with semantic understanding, it identifies the physical characteristics of static obstacles and outputs a feasibility assessment of all directions. It then integrates tire pressure distribution and vehicle posture change data to dynamically analyze the maximum lateral grip threshold of the current road surface. Finally, the probability distribution of moving target trajectories, static obstacle passage assessment, and road grip critical value are spatially fused to form a directional risk field, where each vector points to the direction of the risk source and the vector length reflects the risk intensity level.

3. The intelligent safety control system for children's electric motorcycles based on the Internet of Things according to claim 2 is characterized by: The output includes a fusion risk assessment matrix of the collision probability gradient and the road loss of control coefficient, which includes the following steps: The dynamic risk vector field is divided into sector areas along the polar coordinate system, the maximum risk vector intensity in each area is extracted, and the collision probability gradient is calculated based on the relative speed of the moving target. At the same time, the vehicle loss of control risk factor is dynamically quantified based on the real-time deviation between the maximum lateral adhesion coefficient and the road grip; The collision probability gradient is then mapped to a Cartesian coordinate grid and nonlinearly coupled with the road surface loss of control coefficient at the same grid position to form a spatial assessment matrix that includes both collision risk and loss of control risk, in which high-risk areas exhibit gradient jump characteristics.

4. The intelligent safety control system for children's electric motorcycles based on the Internet of Things according to claim 3 is characterized by: Receive the handle operation timing characteristics and vehicle body dynamic response characteristics from the vehicle sensor, and analyze the matching degree between the operation intention and the vehicle response through cross-attention. The road loss of control coefficient is used as a weight factor in the matching degree calculation, including the following steps: The handle operation signal is segmented into time windows to generate an operation feature sequence, and the vehicle body dynamic response signal is synchronously aligned to generate a response feature sequence; The attention weight distribution of the response feature sequence is calculated based on the operation feature sequence, and the road loss of control coefficient is used as a dynamic adjustment factor to dynamically modify the attention weight intensity; Finally, based on the weighted fused feature sequence, the dynamic consistency between the operation intention and the actual vehicle response is calculated through vector similarity, and a matching score between the operation intention and the vehicle response is output, among which the weight ratio of the automatic enhanced response feature in the low-adhesion road scene is given.

5. The intelligent safety control system for children's electric motorcycle based on the Internet of Things according to claim 4 is characterized in that: Outputting three-level behavior labels of active risky behavior, passive out-of-control behavior, or normal operating behavior includes the following steps: When the matching score is greater than or equal to the first threshold and the vehicle body dynamic response is stable, it is marked as normal operation behavior; When the matching score is less than the first threshold but the operation flow vector shows active steering or acceleration, it is marked as active risky behavior; When the matching score is less than the second threshold and the response flow vector shows abnormal yaw, it is marked as passive out-of-control behavior; For data marked as active risky behavior, check whether its operation direction coincides with the high-risk area in the fusion risk assessment matrix. If so, maintain the label; otherwise, downgrade it to normal operation behavior. Combined with the classification benchmark and environmental verification results, output the final three-level behavior label.

6. The intelligent safety control system for children's electric motorcycles based on the Internet of Things according to claim 5 is characterized by: Differentiated control is performed based on the directional priority and behavior label level of the fused risk assessment matrix. When the behavior label is a passive loss of control behavior, anti-skid vector control is initiated based on the collision probability gradient direction. The following steps are included: The peak azimuth angle of the collision probability gradient in the fusion risk assessment matrix is ​​extracted, and the dominant direction of vehicle slip is determined by combining the trend of vehicle body yaw rate change. Dynamically distribute the output torque difference of the dual-hub motors according to the dominant direction. The wheel on the high-risk side reduces torque output, while the wheel on the low-risk side maintains or increases torque output, forming a corrective torque to suppress sideslip. The vehicle body posture recovery status is monitored in real time. When the lateral acceleration returns to the safety threshold, the torque difference distribution is gradually released to restore symmetrical driving.

7. The intelligent safety control system for children's electric motorcycle based on the Internet of Things according to claim 6 is characterized by: When the behavior is labeled as active risk-taking and the operation direction coincides with the high-risk area, progressive power restriction is triggered, and the spatial and operational characteristics of the intervention scene are recorded in real time, including the following steps: Compare the steering angle of the handle with the high-risk area orientation of the integrated risk assessment matrix, and calculate the angle deviation between the operating direction and the risk core area; When the angle deviation is less than the critical angle, a three-stage progressive power reduction is initiated. In the initial stage, the current power output is maintained, the motor power is linearly reduced in the middle stage, and it stabilizes at a safe power level in the final stage. When the angle deviation is greater than the critical angle, only the vibration warning is triggered; The environmental spatial data and operation time series data before and after the intervention are intercepted to extract the characteristic combination including the distribution pattern of static obstacles, operation peak intensity, and power attenuation curve.

8. The intelligent safety control system for children's electric motorcycles based on the Internet of Things according to claim 7 is characterized in that: Generate a mirrored virtual training task based on the characteristics of the intervention scenario, and transform the parent's operation data into a personalized safety threshold through cluster analysis, including the following steps: Based on the recorded static obstacle distribution patterns and road adhesion parameters, a 3D risk scenario is reconstructed in a virtual environment. Challenge nodes of the same type as the original intervention event are preset in the mirrored scenario, including sharp bend obstacle avoidance and speed control on slippery roads. The operation trajectory data of parents completing training tasks is collected, and steering smoothness and acceleration stability are extracted as feature vectors. The safe operation mode clusters are identified through the density clustering algorithm, and the cluster center features are extracted as the ideal operation template. The key parameters of the ideal operation template are mapped to vehicle control thresholds, including the maximum allowable roll angle and the minimum adhesion coefficient requirement for cornering.

9. The intelligent safety control system for children's electric motorcycles based on the Internet of Things according to claim 8 is characterized by: Dynamically updating the judgment criteria of the behavior analysis module and the control parameters of the security decision module includes the following steps: Convert the parent's steering smoothness range into a matching score threshold, and adjust the weight ratio of the road loss coefficient in the attention calculation according to the cornering adhesion requirements; The maximum allowable roll angle is written into the safety decision-making module as the angular velocity threshold for triggering anti-skid control for passive loss of control behavior. Parent acceleration stability data is obtained, and the slope of the three-level attenuation curve of progressive power limitation is reconstructed. After the new threshold takes effect, the actual number of interventions and scenario type distribution are collected and fed back to the optimization module to initiate iterative adjustment of the threshold.

Images