An AI-based intelligent meeting early warning sentry system

Abstract

The application relates to the technical field of narrowband Internet of Things, and discloses an AI-based intelligent meeting early warning sentry system, which comprises nodes arranged at two ends of a blind area and a cloud server, an asynchronous double-loop heterogeneous communication architecture is constructed, and low-frequency wide-area road risk updating and high-frequency local vehicle state synchronization are separated. A processor allocates a braking weight coefficient according to a target radar scattering cross-section area extracted by a radar, combines a road risk parameter, performs nonlinear scaling on basic theoretical collision time, and outputs a scene collision time to participate in early warning. Meanwhile, the processor deduces a shortest theoretical braking distance of the vehicle in real time, when the spatial position of the vehicle is less than the distance and the actual deceleration does not reach a judgment threshold, failure arbitration is triggered, and a local preset limit adhesion coefficient is forced to be reloaded to execute bottom-layer deviation correction. The application solves the problems that the existing system has a single judgment model and lacks a deviation correction mechanism when external parameters are mismatched, thereby causing early warning lag.

Description

Technical Field

[0001] This invention relates to the field of narrowband Internet of Things (IoT) technology, specifically to an AI-based intelligent vehicle passing warning sentinel system. Background Technology

[0002] On mountain roads, sharp bends, or narrow sections with limited visibility, oncoming traffic safety is a key focus of road traffic management. Existing road warning systems typically deploy radar and display equipment at both ends of the blind spot, detecting oncoming vehicles and calculating relative distance and speed to provide audio and visual alerts. However, in practical applications, the warning logic of conventional systems has certain limitations. Current warning decisions mostly rely on fixed linear kinematic models, which do not take into account the differences in physical mass between different vehicle types and the decrease in road surface adhesion coefficient due to rain and snow. Because heavy vehicles and light vehicles have different braking requirements, and road friction conditions change with weather, using a uniform theoretical collision time for judgment results in inaccurate warning triggering timing under complex road conditions.

[0003] To improve the adaptability of early warning systems, some systems connect to external networks to obtain environmental risk parameters to assist in adjusting early warning strategies. However, in this architecture, wide-area network communication inevitably introduces latency. Traditional synchronous data processing methods can cause the reception and updating of external parameters to interfere with local high-frequency radar detection tasks, affecting the continuity of the underlying computation cycle. Furthermore, externally issued risk parameters are usually generated based on macroscopic meteorological data, which sometimes fails to fully reflect the actual local road conditions. When the actual local road surface adhesion conditions are worse than the assessment results of external parameters, the actual vehicle deceleration effect will not meet expectations. Existing systems lack a low-level correction mechanism based on the vehicle's micro-motion response, and cannot intervene promptly when external parameters are mismatched, resulting in delayed early warning actions and potential safety hazards. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an AI-based intelligent oncoming traffic warning sentinel system. This system solves the problems of existing systems easily interfering with local high-frequency calculations when acquiring external network parameters, their single linear judgment model not combining vehicle type differences and dynamic road surface adhesion status, and the lack of a low-level local correction mechanism when external environmental parameters and actual road conditions are mismatched, resulting in inaccurate and delayed warning timing in complex road conditions.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an AI-based intelligent vehicle-passing warning sentry system, comprising:

[0006] The first and second nodes are arranged at the ends of the road with limited visibility. The first and second nodes are physically symmetrically deployed at both ends of the target curve or narrow blind spot, and are used to monitor vehicles entering from the two opposite lanes respectively.

[0007] Both the first and second nodes include radar modules, processors, wide-area modules, local-area modules, display modules, and alarm modules; as well as cloud servers independently located on the external network;

[0008] The radar module is used to extract the multi-dimensional kinematic features of the target vehicle; the wide-area module is used to establish a low-frequency asynchronous data channel between the processor and the cloud server; and the local module is used to establish a high-frequency local synchronization loop between the first node and the second node.

[0009] The cloud server is used to identify the current road surface risk status and send road surface risk parameters to the processor via the wide area module;

[0010] The processor establishes electrical connections with the radar module, wide-area module, local module, display module, and alarm module respectively. The processor is used to receive the multi-dimensional kinematic features of the target vehicle extracted by the radar module, obtain the state dataset of the oncoming vehicle through the local module, combine the road risk parameters to solve the scenario-based collision time, and drive the display module and alarm module to execute alarms based on the solution results.

[0011] Furthermore, the high-frequency local synchronization loop includes a first loop and a second loop; the first loop is a low-frequency wide-area asynchronous confirmation loop composed of a cloud server, a wide-area module, and a processor, used by the cloud server to send road risk parameter data packets containing road risk coefficients to the processor, and by the processor to send parameter reception confirmation information back to the cloud server; the second loop is a high-frequency local bidirectional synchronization loop composed of a first node, a second node, and their respective local modules, used for bidirectional exchange of vehicle kinematic state data and confirmation of link health status within bidirectional blind spots.

[0012] Furthermore, an AI environmental risk identification model is deployed inside the cloud server. The AI ​​environmental risk identification model is implemented using a decision tree model. The decision tree model includes a data preprocessing module, a feature construction module, a decision tree inference module, and an output mapping module. The decision tree model is used to receive the current precipitation, ambient temperature, relative humidity, historical road surface adhesion coefficient, and historical vehicle deceleration response data of the target road area and convert them into feature vectors. It performs node judgments level by level in the tree structure and generates the road surface risk coefficient corresponding to the current road segment based on the leaf node results.

[0013] Furthermore, the processor is configured with a serial port receive interrupt triggering mechanism and a parameter parsing task branch. When a road risk parameter data packet is received, the processor responds to the interrupt and stores the data in the receive buffer. The parameter parsing task branch then performs timestamp verification, cyclic redundancy check, and numerical range check. After the verification passes, the processor calls the mutex lock mechanism or temporarily disables the global interrupt, performs an overwrite operation on the specified memory address storing the basic calculation parameters, and calculates the dynamic road adhesion coefficient and dynamic suggested speed limit based on the updated road risk coefficient. If the verification fails, the data expires, or the value goes out of bounds, the processor maintains the original parameters in memory or rolls back to the preset conservative road risk coefficient.

[0014] Furthermore, the radar module is used to extract the instantaneous spatial distance, instantaneous scalar velocity, and target radar cross section of the target vehicle; the processor is used to classify the target into light and heavy vehicles based on the target radar cross section and assign corresponding braking weight coefficients; when the target radar cross section is greater than or equal to the classification threshold, it is determined to be a heavy vehicle and a heavy vehicle braking weight coefficient is assigned; when the target radar cross section is less than the classification threshold, it is determined to be a light vehicle and a standard braking weight is assigned; the processor generates a micro kinematic state dataset from the instantaneous spatial distance, instantaneous scalar velocity, and braking weight coefficients.

[0015] Furthermore, the processor calculates the local link packet loss rate by counting the number of heartbeat packets that are not successfully received within the internal sliding time window in real time. When the local link packet loss rate is greater than or equal to the communication failure threshold, the processor forcibly switches to the asymmetric degradation defense mode. The processor's memory manager intercepts the listening thread and virtually constructs a virtual vehicle model under extreme physical boundary conditions in the data structure reserved for oncoming vehicles. The dataset of the virtual vehicle model is then forcibly substituted into the bidirectional collision calculation cache to compensate for the lack of input.

[0016] Furthermore, the extreme physical boundary conditions include: setting the spatial position of the virtual vehicle to the physical vertex at the other end of the blind zone, and setting the instantaneous spatial distance to zero; setting the instantaneous scalar speed of the virtual vehicle to reach the maximum theoretical safe passage speed benchmark value determined based on the longitudinal adhesion coefficient of the static reference road surface for the current road segment; and setting the braking weight coefficient of the virtual vehicle to the braking weight coefficient of heavy vehicles.

[0017] Furthermore, the processor internally incorporates a scenario-based collision time calculation algorithm. This algorithm determines the equivalent relative distance using the physical arc length of the blind spot, the instantaneous spatial distance between the local target vehicle and the oncoming vehicle, and the relative speed using the instantaneous scalar speeds of both vehicles. It also uses the ratio of the dynamic road surface adhesion coefficient to the static reference road surface longitudinal adhesion coefficient, as well as the average braking weight coefficient of both vehicles, as penalty factors. Through algebraic operations, the basic theoretical collision time calculated from the relative distance and relative speed is nonlinearly scaled, and the scenario-based collision time is output to participate in early warning decision-making.

[0018] Furthermore, the processor is internally configured with a warning control state machine and a hysteresis comparator logic. When an alarm command is triggered, the warning control state machine enters a one-way latch cycle to force the alarm intensity to be maintained or upgraded. The processor must meet the following conditions to release the monotonic latch state: the target vehicle speed is measured to drop to the value of the dynamic suggested speed limit minus the preset hysteresis bias within several consecutive operation cycles, the latest calculated scenario-based collision time is greater than the upper limit of the preset time threshold group, and the risk of insufficient braking space is not triggered.

[0019] Furthermore, the processor performs AI road risk coefficient and local microscopic vehicle response divergence event capture; the processor calculates in real time the shortest theoretical braking distance required for the target vehicle to reduce its speed to the dynamically suggested speed limit; when the instantaneous spatial distance is determined to be less than or equal to the shortest theoretical braking distance, the processor monitors the microscopic deceleration characteristics of the target vehicle; if the target vehicle speed exceeds the dynamically suggested speed limit within a consecutive preset time period, and the actual deceleration value is less than the system's preset deceleration judgment threshold, the processor triggers the failure arbitration logic to execute local physical limit parameter forced reload, clears the road risk coefficient, dynamic road adhesion coefficient and dynamic suggested speed limit in the calculation cache, and forcibly calls the preset local extreme harsh environment limit adhesion coefficient for memory overwrite.

[0020] This invention provides an AI-based intelligent vehicle-passing warning sentinel system. It has the following beneficial effects:

[0021] 1. This invention constructs a heterogeneous communication architecture comprising a low-frequency wide-area asynchronous loop and a high-frequency local synchronous loop. By configuring a serial port receive interrupt triggering mechanism and mutex lock technology, the system performs an overwrite operation on a specified memory address after receiving road risk parameters from the cloud. This structure physically separates the external parameter update process from the local high-frequency radar detection task, avoiding interference from wide-area network communication delays on the underlying collision calculation cycle and ensuring the continuity of local early warning control at the nodes.

[0022] 2. This invention introduces a nonlinear scaling mechanism for multidimensional physical parameters during the solution of scenario-based collision time. The processor allocates vehicle braking weight coefficients based on the target radar cross-section extracted by radar, and constructs a penalty factor by combining the attenuation ratio of the road adhesion coefficient, performing algebraic scaling on the basic theoretical collision time. This algorithm directly incorporates the mass differences of different vehicle models and the changes in current road friction conditions into the kinematic model, correcting the calculation deviation of conventional linear judgment logic under complex road conditions, and improving the accuracy of warning triggering timing.

[0023] 3. This invention establishes a low-level parameter correction logic based on the microscopic motion characteristics of the vehicle. The processor calculates the shortest theoretical braking distance required for the target vehicle to decelerate in real time. When the vehicle's spatial position is less than this braking distance and the actual deceleration does not reach the system's judgment threshold, the system determines that the externally issued risk parameters are mismatched with the actual road conditions. At this time, the processor triggers a failure arbitration, forcibly calling the locally preset extreme environmental adhesion coefficient for memory reload. This mechanism provides low-level, cross-level intervention when external parameter evaluation is insufficient, compensating for the safety hazards caused by the lag in environmental parameters. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the system architecture of the present invention;

[0025] Figure 2 This is a schematic diagram of the method flow of the present invention;

[0026] Figure 3 This is a flowchart of the static boundary initialization and normalization control logic of the present invention;

[0027] Figure 4 This is a time-series diagram of the road risk parameters and dynamic speed limit updates of the present invention.

[0028] Figure 5 This is the high-frequency detection and link degradation defense control logic diagram of the present invention;

[0029] Figure 6 This is the scene-based collision solving and state machine control logic diagram of the present invention;

[0030] Figure 7 This is the logic diagram for braking boundary deduction and AI road surface risk coefficient failure determination of the present invention;

[0031] Figure 8 This is a comparison diagram of the collision time evolution under dark ice conditions according to the present invention;

[0032] Figure 9 This is a comparison chart of the smoothness of the system alarm output under microscopic driving disturbances according to the present invention.

[0033] Among them, 100 is the first node; 200 is the second node; 110 is the radar module; 120 is the processor; 130 is the wide area module; 140 is the local area module; 150 is the display module; and 160 is the alarm module. Detailed Implementation

[0034] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] See attached document Figure 1 This invention provides an AI-based intelligent vehicle-passing warning sentry system, which may include:

[0036] A first node 100 and a second node 200 are arranged at the ends of the road's visibility-limited area. The first node 100 and the second node 200 are physically symmetrically deployed at both ends of the target curve or narrow blind spot, respectively, to monitor vehicles entering from the two opposite lanes.

[0037] The first node 100 and the second node 200 adopt the same resource configuration at the hardware level. Taking the first node 100 as an example, the internal hardware resources specifically include radar module 110, processor 120, wide area module 130, local area module 140, display module 150, and alarm module 160.

[0038] The radar module 110 is located on the physical plane of the first node 100 facing the direction of the oncoming vehicle, and is used to extract the target vehicle's distance, speed, and radar cross-section.

[0039] The processor 120 establishes an electrical connection with the radar module 110 through a general-purpose digital interface, and receives high-frequency sampling baseband data output by the radar module 110. The processor 120 undertakes multi-threaded computing tasks in this system, and its data output terminals are connected to the driver board of the display module 150 and the control pins of the alarm module 160, respectively.

[0040] The processor 120 generates a corresponding display control data frame based on the alarm threshold level output by the internal calculation model and sends it to the driver board of the display module 150 to control the character switching, brightness level or flashing mode of the display module 150; at the same time, the processor 120 outputs pulse width modulation signals with different duty cycles or different frequencies to drive the audible and visual alarm frequency adjustment of the alarm module 160.

[0041] In terms of communication network topology design, the system constructs an asynchronous dual-ring heterogeneous communication link with independent communication modules, heterogeneous communication protocols, and isolated data buffers. The first ring relies on the wide-area module 130 for construction, and the second ring relies on the local area module 140 for construction.

[0042] The first loop is a low-frequency wide-area asynchronous confirmation loop consisting of a cloud server 300, a wide-area module 130, and a processor 120. The cloud server 300 sends road risk parameter data packets to the processor 120. After memory overwriting or parameter discarding, the processor 120 can send parameter reception confirmation information back to the cloud server 300 via the wide-area module 130. The parameter reception confirmation information may include node identifier, road segment identifier, data packet sequence number, verification result, and the currently used road risk coefficient. Source identification, dynamic road surface adhesion coefficient Source identification and dynamic speed limit suggestions And the node running status field.

[0043] The second loop is a high-frequency local bidirectional synchronous loop consisting of the first node 100, the second node 200 and their respective local modules 140, used for interactive confirmation of bidirectional vehicle micro-kinematic state data and link health status.

[0044] The wide-area module 130 adopts narrowband IoT technology specifications to establish a low-frequency asynchronous data channel between the processor 120 and the cloud server 300. The cloud server 300 is independently located on the external network and generates road risk coefficients by calling external meteorological data. The road risk parameters are then injected into the processor 120 via the wide area module 130.

[0045] Local module 140 uses a Sub-GHz band communication protocol to establish a point-to-point data connection between local module 140 of the first node 100 and local module 140 of the second node 200. The first node 100 and the second node 200 utilize their respective local modules 140 to form a high-frequency local synchronization loop. This high-frequency local synchronization loop is independent of the low-frequency asynchronous data channel constructed by the wide-area module 130, and is used for bidirectional exchange of vehicle kinematic state data and confirmation of link health status within bidirectional blind spots.

[0046] The cloud server 300 is also equipped with an AI environmental risk identification model, which is used to identify the current road surface risk status based on the rainfall, ambient temperature, relative humidity, historical road surface adhesion coefficient, and historical vehicle deceleration response data of the target road area, and output the corresponding road surface risk coefficient, so that the processor 120 can calculate the dynamic road surface adhesion coefficient and dynamic suggested speed limit.

[0047] See attached document Figure 2This invention provides an AI-based intelligent vehicle meeting warning method. Based on the aforementioned system operation, the overall workflow includes the following steps:

[0048] S10, Initialize static spatial boundary parameters and calculate the theoretical safe passage speed of the current road segment;

[0049] S20, the cloud server 300 asynchronously sends road risk parameters to the processor 120 through the wide area module 130; the processor 120 performs timestamp verification, cyclic redundancy check, and numerical range check on the received road risk parameter data packet, and updates the local road risk coefficient after the verification passes. ,based on Calculate the dynamic road surface adhesion coefficient and dynamic speed limit suggestions If the verification fails or the data expires, the previous valid parameter will be maintained or a preset conservative parameter will be used.

[0050] S30, radar module 110 extracts multi-dimensional kinematic features of target vehicle, processor 120 evaluates local module 140 communication link status and performs two-way status synchronization or degraded defense.

[0051] S40, Processor 120 combined with dynamic road adhesion coefficient Dynamic speed limit recommendations The scenario-based collision time is solved by combining the multi-dimensional kinematic characteristics of the target vehicle, and the display module 150 and alarm module 160 are driven to perform monotonic latching alarm.

[0052] S50, processor 120 determines the braking space boundary based on real-time theoretical braking. Whether there is a mismatch with the local micro-vehicle response, and whether reverse correction and cross-level defense intervention are implemented.

[0053] To further clarify the implementation of each technical aspect of the present invention, the following will provide a detailed description of the implementation of each functional module involved above and its internal processing flow.

[0054] See attached document Figure 3 Step S10 is mainly used to solidify the environmental physical characteristics and establish a safety baseline in the early stage of system deployment, and may specifically include the following steps:

[0055] S101, obtain static spatial geometric boundary parameters.

[0056] During the on-site equipment installation and commissioning phase at the first node 100 and the second node 200, technicians can write the physical topology features of the target road into the system via an external configuration interface. This external configuration interface could be, for example, an RS485 serial communication interface or a debugging serial port reserved for the processor 120. The written physical feature parameters are used as the basis for subsequent kinematic calculations, specifically including the curvature radius of the target curve. Road superelevation cross slope The physical arc length of the blind spot between the first node 100 and the second node 200 caused by the obstruction of mountains or buildings. Combination, and the static reference longitudinal adhesion coefficient of the pavement calibrated by the pavement material under dry weather conditions. .

[0057] For the calibration of the above parameters, the radius of curvature of the curve... The value can be determined based on the actual road survey data; road superelevation cross slope It can be calculated by measuring the lateral inclination angle of the road, and its value is usually between 0.02 and 0.08; the static reference pavement longitudinal adhesion coefficient. The calibration is usually based on the road paving material. For example, asphalt pavement usually takes a value of 0.7 to 0.8, while cement concrete pavement usually takes a value of 0.6 to 0.7.

[0058] After receiving the above parameters, the processor 120 burns them into its non-volatile memory via the internal bus, so that the environmental topology data can be directly accessed after a system restart. Simultaneously, technicians can also input the target road number, curve radius of curvature, road superelevation cross slope, blind spot physical arc length, and static reference pavement longitudinal adhesion coefficient. The static spatial geometric boundary parameters are synchronously written to the cloud server 300; or, upon initial network connection, the processor 120 uploads the aforementioned static spatial geometric boundary parameters to the cloud server 300 via the wide-area module 130. Thus, the cloud server 300 subsequently generates the road surface risk coefficient. and dynamic speed limit recommendations It can call the same method as the edge node. Processor 120 is based on and local storage Calculate the dynamic road surface adhesion coefficient and dynamic speed limit recommendations .

[0059] S102, Establish a theoretical safe passage speed boundary model.

[0060] Processor 120 reads the stored static spatial geometric boundary parameters from non-volatile memory and performs dynamic modeling based on the critical physical conditions for lateral slippage of the vehicle while driving in a curve. The parameters obtained in this step... Based on road geometry and static reference pavement longitudinal adhesion coefficient The established static safe speed limit benchmark is used as the basic boundary for routine speed limit reminders and subsequent violation determinations; dynamic speed limit recommendations... This is used to assess different road surface risk conditions. Perform dynamic corrections.

[0061] Processor 120 calculates the maximum theoretical safe passage speed of the current road segment under static physical conditions by executing internal arithmetic instructions. The corresponding calculation formula is as follows:

[0062] ;

[0063] In the formula, The current static maximum theoretical safe passage speed reference value for the road segment, output by processor 120; The acceleration due to gravity is constant, and its value can be 9.8 m / s². 2 ; The radius of curvature of the bend read from non-volatile memory; The aforementioned static reference road surface longitudinal adhesion coefficient; The aforementioned road superelevation cross slope.

[0064] After processor 120 completes the above solution, it will... The values ​​are written into the internal high-speed static random access memory as a safety benchmark for subsequent real-time comparison and limit determination.

[0065] S103, implement routine early warning display control.

[0066] Radar module 110 scans the road physical space by transmitting frequency-modulated continuous waves at a set frame rate. If radar module 110 fails to detect the echo of a moving target within its detection line of sight, or if it detects a target vehicle but processor 120 calculates its actual speed to be less than or equal to the current dynamically recommended speed limit, the radar will proceed as follows: When the scenario-based collision time threshold, braking space insufficiency threshold, or environmental parameter failure risk conditions are not triggered, the control state machine of processor 120 remains in the normal monitoring state.

[0067] In this state, the processor 120 continuously sends data including the current dynamic suggested speed limit to the display module 150. The standard display driver data frame drives the display module 150 to normally emit light and output statically or in a loop. It is recommended that the target vehicle control its speed. The system displays a prompt character within a specified range. Simultaneously, the processor 120 cuts off or sets the duty cycle of the pulse width modulation signal output to the alarm module 160 to zero, keeping the warning light array and sound-emitting device of the alarm module 160 in low-power sleep mode to avoid unnecessary visual or auditory interference to vehicles driving normally and not exceeding the speed limit.

[0068] See attached document Figure 4 Step S20 is mainly used for the road surface risk coefficient during system operation. The generation and dynamic road surface adhesion coefficient Conversion and dynamic speed limit suggestions The underlying asynchronous update mechanism may include the following steps:

[0069] S201 generates road surface risk coefficients and determines dynamic speed limit parameters.

[0070] The cloud server 300 has a scheduled task program running internally, which calls the external meteorological data service interface at preset time intervals. This preset time interval can be set to a specific value between 15 and 60 minutes, such as 30 minutes, based on communication power consumption constraints and weather change patterns. Through the meteorological data service interface, the cloud server 300 obtains the current real-time meteorological parameters of the target road area, primarily including precipitation, ambient temperature, and relative humidity.

[0071] The cloud server 300 further inputs the acquired meteorological parameters into the AI ​​environmental risk identification model, and outputs the road surface risk coefficient. Processor 120 received Then, the dynamic road surface adhesion coefficient is calculated. :

[0072] ;

[0073] Subsequently, processor 120 according to Calculate dynamic suggested speed limit :

[0074] ;

[0075] When the road surface is dry Smaller near , Approximate static safe passage speed When the road surface is wet, flooded, icy, or oily, Increase reduce, This reduces the speed limit accordingly. Therefore, differentiated speed limit prompts can be applied to vehicles based on the degree to which frictional resistance decreases under different road conditions.

[0076] In one specific implementation, the AI ​​environmental risk identification model can be implemented using a decision tree model. The decision tree model includes a data preprocessing module, a feature construction module, a decision tree inference module, and an output mapping module connected sequentially. The data preprocessing module receives and cleans data on precipitation, ambient temperature, relative humidity, historical pavement adhesion coefficient, and historical vehicle deceleration response of the target road area; the feature construction module converts the above data into feature vectors that the model can read; the decision tree inference module performs node judgments level by level in the tree structure based on the feature vectors; and the output mapping module generates the pavement risk coefficient corresponding to the current road segment based on the leaf node results output by the decision tree inference module. .

[0077] The decision tree inference module includes a root node, multiple intermediate decision nodes, and multiple leaf nodes. The root node performs the initial segmentation of the main risk factors in the input features; these factors can be precipitation, ambient temperature, or relative humidity. Intermediate decision nodes further branch the input features based on different thresholds. Leaf nodes output the road risk coefficient corresponding to the current input feature. Or the road surface risk coefficient range. Nodes are connected in the direction from parent node to child node. The input feature vector enters from the root node, passes through one or more intermediate judgment nodes, and reaches the corresponding leaf node, thus obtaining the current road environment risk identification result.

[0078] The feature vector can be represented as:

[0079] ;

[0080] In the formula, This indicates the precipitation in the target road area during the current sampling period. Indicates ambient temperature. Indicates relative humidity. This represents the historical road surface adhesion coefficient recorded or calibrated under similar historical weather conditions for the target road. This represents historical vehicle deceleration response data.

[0081] Historical vehicle deceleration response data This can include the vehicle's speed change sequence after entering the detection area, the actual deceleration calculated between adjacent radar detection frames, the triggered alarm level, and the road risk coefficient used at that time. Dynamic road surface adhesion coefficient Dynamic speed limit recommendations And whether it is triggered Event identifier.

[0082] The output of the decision tree model is the road surface risk coefficient. Road surface risk coefficient The value range is from 0 to 1. The closer it is to 0, the drier the road surface is, and the greater the frictional resistance between the tire and the road surface. The closer the value is to 1, the closer the road surface is to abnormal conditions such as water accumulation, black ice, oil stains, or sand and gravel coverage, and the smaller the frictional resistance between the tire and the road surface.

[0083] In one specific implementation, when the AI ​​environmental risk identification model determines that the current road surface is a dry, low-risk condition, The value ranges from 0 to 0.1; when the current road surface is determined to be wet or slightly slippery, The value ranges from 0.1 to 0.3; when the current road surface is determined to be covered by rain, water, or mud, The value ranges from 0.3 to 0.6; when the current road surface is determined to be at high risk of icing, black ice, or oil contamination, The value ranges from 0.6 to 0.9.

[0084] The training samples for the decision tree model were constructed from a historical sample database of 300 road segments on a cloud server. Each sample includes a set of input features and a corresponding label. The input features include the precipitation P and ambient temperature at the time of sample collection. relative humidity Historical road surface adhesion coefficient and historical vehicle deceleration response data The corresponding label is the road surface risk coefficient label for that sample time. The training samples include historical meteorological data returned from external meteorological data service interfaces, road surface adhesion coefficient data calibrated during road installation, commissioning, or maintenance, vehicle speed change sequences and actual deceleration data uploaded by the first node 100 and the second node 200 during operation, and alarm levels recorded by the processor 120. Trigger event data.

[0085] Regarding labeling methods, the cloud server 300 can determine the actual road surface condition corresponding to the sample based on historically calibrated road surface adhesion coefficients or manual maintenance records. When the historically calibrated road surface adhesion coefficient is close to the dry road surface adhesion range, the corresponding road surface risk coefficient label will be applied to the sample. The range is 0 to 0.1; when the historically calibrated road surface adhesion coefficient is lower than the dry road surface adhesion range but has not yet reached the icing or extreme slippery threshold, it will be... The value is marked as 0.1 to 0.6; when the historically calibrated road surface adhesion coefficient is lower than the preset icing threshold, or when the same road section is triggered multiple times under the same weather conditions. During the event, The values ​​are labeled as 0.6 to 0.9.

[0086] For road surface risk factor labels The road surface adhesion coefficient can be determined based on the historical measured or calibrated values ​​corresponding to the sample. Longitudinal adhesion coefficient of static reference pavement The ratio is determined as follows:

[0087] ;

[0088] When according to and Calculated When it is less than 0, Corrected to 0; when When it is greater than 1, Adjust it to 1 so that the training labels are always within the valid range of 0 to 1.

[0089] During the training process, the cloud server 300 first reads multiple training samples from the road segment historical sample database and converts the input features of each training sample into a feature vector. Then, using road surface risk coefficient labels. As supervisory labels, a decision tree training algorithm is used to recursively partition the samples. During each partition, the cloud server (300) analyzes precipitation data. Ambient temperature relative humidity Historical road surface adhesion coefficient and historical vehicle deceleration response data The system selects features and corresponding thresholds that minimize prediction errors as the current decision nodes, continuing this process until the preset maximum tree depth, minimum leaf node sample size, or prediction error convergence condition is met. After training, the cloud server saves the node features, node thresholds, branch directions, and leaf node output values ​​of the decision tree, and calls this decision tree model to perform inference on real-time meteorological data and historical vehicle deceleration response data during system runtime.

[0090] Historical vehicle deceleration response data can be generated by the first node 100 and the second node 200 during daily operation. Specifically, it includes the speed change sequence after the vehicle enters the detection zone, the actual deceleration, and the triggered alarm level. Values, Values, Values ​​and whether to trigger The event identifier. The processor 120 can periodically upload the above-mentioned historical vehicle deceleration response data to the cloud server 300 through the wide area module 130, and the cloud server 300 stores it in the road segment historical sample library for use in training, correction or inference of the AI ​​environmental risk identification model.

[0091] S202, perform asynchronous low-frequency data downlink and road risk parameter overwrite.

[0092] The cloud server 300 distributes road risk parameter data packets to the wide area module 130 connecting the first node 100 and the second node 200 via a cellular communication base station network. The road risk parameter data packets may include a frame header, protocol version number, road segment identifier, node identifier, and road risk coefficient. The data includes a timestamp, an expiration date field, and a cyclic redundancy check field. After receiving the RF signal and completing baseband decoding, the wide area module 130 transmits the data packet to the processor 120 via a universal asynchronous transceiver interface.

[0093] To reduce the time delay impact of low-frequency downlink data reception on the main control loops of the processor 120, such as local radar high-frequency detection and alarm calculation, the processor 120 uses a combination of hardware receive interrupts and background low-priority tasks to process the aforementioned data. The processor 120 is configured with a serial port receive interrupt triggering mechanism.

[0094] When data sent by the wide-area module 130 arrives at the serial port register, the processor 120 responds to the interrupt, suspends the execution of regular instructions in the current main loop, and jumps the program counter to the interrupt service function. Within the interrupt service function, the processor 120 temporarily stores the received data bytes in an internally allocated receive buffer, sets the data ready flag, and then exits the interrupt to resume the main loop.

[0095] The processor 120's main loop includes a dedicated parameter parsing task branch. This parsing task branch periodically polls the status of the data ready flag. When a valid flag is detected, the parsing task extracts data packets from the receive buffer and performs cyclic redundancy check, timestamp check, and numerical range check.

[0096] If the cyclic redundancy check fails, it indicates that a bit error occurred in the data during wireless transmission. The processor 120 discards the data packet and maintains the original parameters in memory until the data is reacquired in the next update cycle. If the cyclic redundancy check passes, the processor 120 further performs timestamp verification and numerical range verification.

[0097] If a valid road risk parameter data packet has not been received after the system starts, the processor 120 will use a preset default road risk coefficient. It is used as the default value in the calculation, where It can be configured to 0.1 to 0.3, or a more conservative value depending on the system security level; if the timestamp of the received data packet exceeds the preset validity period, or if the data packet carries... If the value exceeds the preset range of 0 to 1, the processor 120 determines that the data packet is invalid and maintains the previous valid value. , or The system reverts to a preset conservative road surface risk coefficient. If the cyclic redundancy check, timestamp check, and numerical range check all pass, the processor 120 extracts the data packet... and calculate and .

[0098] After successful extraction, processor 120 performs an overwrite operation on the specified memory address storing the basic calculation parameters, replacing the old parameters with the new ones. , and During the instruction cycle of the overwrite operation, the processor 120 invokes the operating system's mutex mechanism or temporarily disables global interrupts to suppress high-frequency collision time calculation threads running in parallel from reading the memory address at this time, thereby reducing the risk of abnormal algorithm output due to reading residual scrambled data during the update process.

[0099] After completing the memory overwrite, processor 120 releases the mutex and clears the receive buffer. Through the cooperation of the aforementioned hardware interrupt and memory lock mechanism, the system maintains the continuity of the local high-frequency detection thread while implementing low-level updates of the road risk coefficient, dynamic road adhesion coefficient, and dynamic suggested speed limit at the edge nodes. After completing the memory overwrite or parameter discarding process, processor 120 can also send parameter reception confirmation information back to cloud server 300 through wide-area module 130 to form a low-frequency wide-area asynchronous confirmation loop.

[0100] See attached document Figure 5 Step S30 is mainly used to perform multi-dimensional feature extraction of targets in the local high-frequency loop, bidirectional vehicle status data synchronization, and asymmetric degradation defense strategy to address defects in local communication links. Specifically, it may include the following steps:

[0101] S301, Perform target multidimensional feature extraction and braking weight mapping.

[0102] When the target vehicle enters the monitoring line-of-sight area of ​​the first node 100, the radar module 110 generates baseband data containing the difference frequency signal by transmitting a frequency-modulated continuous wave and receiving the target's reflected echo. The processor 120 processes this baseband data using a fast Fourier transform to extract the target vehicle information. instantaneous spatial distance With instantaneous scalar velocity .

[0103] After acquiring the distance and speed parameters, the processor 120 further extracts the target radar cross-section features carried in the radar echo signal. Based on the target radar cross-section, the processor 120 classifies the target into light and heavy vehicles and assigns corresponding braking weight coefficients. To reduce the risk of misclassification caused by the influence of attitude angle, range fluctuations, or local occlusion on the radar cross section of a single frame, the processor 120 can perform moving average, median filtering, or confidence weighting on the radar cross section of multiple consecutive radar detection frames, and perform light and heavy vehicle classification based on the processed radar cross section features.

[0104] The specific classification rule is as follows: the processor 120 has a preset radar cross-section classification threshold. The threshold value is typically configured to range from 6 to 8 square meters. When the extracted target radar cross-section is greater than or equal to the classification threshold... At that time, processor 120 will target the vehicle The vehicle was classified as a heavy vehicle and assigned a heavy vehicle braking weighting coefficient. Braking weighting coefficient for heavy vehicles Typically, a value between 0.7 and 0.8 is used to characterize the attenuation of braking capability. When the extracted target radar cross-section is smaller than the classification threshold... At that time, processor 120 will target the vehicle It was classified as a light vehicle and assigned standard braking weights. The standard braking weight is usually set to 1.0.

[0105] Through the above mapping, the processor 120 generates the target vehicle in memory. Micro kinematic state dataset When radar module 110 detects multiple moving targets within the same detection cycle, processor 120 extracts the distance, velocity, and radar cross-section features of each moving target and generates corresponding candidate micro-kinematic state datasets. Processor 120 can determine the highest-risk target from multiple candidate moving targets based on risk priority (from near to far, from high to low, from low to high, or from short to long scenario-based collision time), and uses this highest-risk target as the target vehicle participating in the early warning calculation for this cycle. .

[0106] S302 performs local link status monitoring and normal bidirectional synchronization.

[0107] While conducting high-frequency radar detection, the first node 100 and the second node 200 establish a high-frequency local synchronization loop using their respective local modules 140. To assess the channel health status under the communication environment, the local modules 140 on both sides exchange heartbeat data packets at a fixed high-frequency reference frequency, such as 10Hz. The processor 120 is internally configured with a fixed-length sliding time window, for example, a time length of 1 second, to count the number of heartbeat packets that are not successfully received within the sliding time window in real time and to calculate the current local link packet loss rate. .

[0108] Processor 120 will calculate the local link packet loss rate in real time. Compare with the internally preset communication failure threshold. Communication failure threshold The value can be set to 20% to 30% depending on the hardware fault tolerance capability of the local module 140.

[0109] When determining the local area link packet loss rate Less than the communication failure threshold At this time, processor 120 determines that the current bilateral communication link is in a normal state. In this state, processor 120 of the first node 100 drives local module 140 to display the locally generated target vehicle. State dataset Send to the second node 200. The local status data packet may include a frame header, sending node identifier, receiving node identifier, target validity flag, target number, and target distance. Target speed Radar cross-section characteristics, braking weighting coefficient The data includes timestamps, link heartbeat sequence numbers, and verification fields. Similarly, the first node 100 receives information about oncoming vehicles sent by the second node 200 through the local module 140. State dataset .

[0110] If the radar on the second node 200 side does not detect an oncoming vehicle, the second node 200 will send a no-payload data packet containing a target identifier bit to the first node 100 to prevent deadlock in the receiving thread of the first node 100. After recognizing the target identifier bit, the processor 120 of the first node 100 marks the oncoming vehicle state as invalid and skips the two-vehicle collision time calculation in this cycle, relying solely on the local vehicle speed. With dynamic speed limit suggestions Implement single-vehicle speeding warnings or regular speed limit reminders.

[0111] After the processor 120 completes the data handshake by verifying the packet header and frame tail, if the target valid identifier bit in the data packet is valid, the corresponding kinematic feature data of the oncoming vehicle is written into the pre-allocated bidirectional collision calculation buffer for subsequent spatiotemporal solution; if the target valid identifier bit in the data packet is invalid, the oncoming no-target identifier bit is written into the bidirectional collision calculation buffer, and the solution of the two-vehicle meeting collision time is skipped in this cycle.

[0112] S303, executes link degradation judgment and virtual vehicle asymmetric defense.

[0113] Due to factors such as geological obstruction or electromagnetic interference, the high-frequency local synchronization loop may experience intermittent failures. When processor 120 determines the local link packet loss rate within the sliding time window... Greater than or equal to the communication failure threshold This indicates that a single-sided node can no longer obtain the true status of the oncoming vehicle in the blind spot.

[0114] To compensate for the lack of algorithm input under the condition of impaired communication, processor 120 forcibly switches to asymmetric degradation defense mode under this trigger condition. The memory manager of processor 120 directly intercepts the listening thread waiting to receive data from the other party, and virtually constructs a virtual vehicle under extreme physical boundary conditions in the data structure reserved for the other vehicle. Model.

[0115] Processor 120 forcibly writes preset extreme physical boundary parameters into the data structure via register direct addressing. The specific parameter combinations written include: setting the virtual vehicle... The spatial position has reached the physical vertex at the other end of the blind zone, i.e., the set instantaneous spatial distance. Setting up virtual vehicles The vehicle's speed has reached the safe limit allowed on the current road section, i.e., the set instantaneous scalar speed. Among them, increased road surface risks led to Below At that time, continue to use As the speed input for virtual vehicles, to improve the conservative early warning strength in link failure states, the virtual vehicle is set. For the heavy-duty freight vehicle with the worst braking performance, the braking weight of this virtual target is set. .

[0116] After completing the above memory assignment, processor 120 will assign the virtual vehicle Virtual state dataset The actual oncoming vehicle data received by the local module 140 is forcibly incorporated into the subsequent collision time calculation buffer. This allows the subsequent warning decision-making process to be driven solely by radar detection data from one side, compensating for input loss due to physical communication link failures and achieving unidirectional conservative control in the event of a network outage.

[0117] When processor 120 detects local area link packet loss rate In continuous Within each sliding time window, the value was below the preset recovery threshold. When processor 120 continuously receives peer status data packets containing valid frame headers, frame trailers, and check fields, it exits the asymmetric degradation defense mode and stops writing virtual vehicles to the peer vehicle data structure. The extreme physical boundary parameters are determined, and the actual opposing vehicle state data received using local module 140 is recovered. The recovery threshold is then determined. Less than the communication failure threshold This is to avoid frequent switching between normal and degraded modes caused by critical fluctuations in the link. The number of consecutive sliding time windows required for the downgrade recovery determination can be configured to be 2 to 5 based on the communication stability of the local module 140; The link recovery threshold can be configured to be 10% to 15% based on the link quality of the local module 140.

[0118] See attached document Figure 6 Step S40 is mainly used to perform nonlinear solution of bidirectional vehicle scenario-based collision time by combining multidimensional physical environment parameters, and to shield micro-driving disturbances through the underlying hysteresis state machine logic to realize alarm latching and degradation release. Specifically, it may include the following steps:

[0119] S401 performs a scenario-based collision time-multidimensional nonlinear solution.

[0120] Before performing scene-based collision time calculation, the processor 120 uses the locations of the first node 100 and the second node 200 as the spatial boundary points at both ends of the blind zone to establish equivalent one-dimensional distance coordinates along the road centerline or curve arc direction. Indicates local target vehicle The remaining distance along the road centerline to the boundary of the blind spot entrance where the first node 100 is located. Indicates oncoming vehicles The remaining distance along the road centerline to the boundary of the blind spot entrance where the second node 200 is located. This represents the physical arc length of the blind spot between the first node 100 and the second node 200, formed by the curvature of mountains, buildings, or roads.

[0121] Based on the above definition, the equivalent relative distance between two-way vehicles before entering the blind spot is determined by... , and The distance is determined jointly. When the target vehicle has passed the corresponding node and entered the blind spot, the processor 120 can set the remaining distance of the vehicle to zero, or correct the equivalent relative interval distance according to the distance already traveled after entering the blind spot along the road centerline.

[0122] Processor 120 further extracts the local target vehicle from the bidirectional collision calculation cache. State dataset If the bidirectional collision calculation cache contains a targetless flag, the processor 120 will not perform the dual-vehicle scenario-based collision time calculation, but will instead enter the single-vehicle risk determination branch; in this single-vehicle risk determination branch, the processor 120 will only determine the local target vehicle. Are there risks of speeding, insufficient braking distance, or mismatch in AI road risk coefficients?

[0123] Specifically, if the target vehicle is located in the local area instantaneous scalar velocity Speed ​​exceeding the dynamic recommended speed limit The processor 120 will trigger a single-vehicle overspeed alarm; if the local target vehicle Instantaneous spatial distance between the probe surface and the probe surface Less than or equal to the shortest theoretical braking distance If the actual deceleration response is insufficient, the processor 120 will trigger a braking space insufficient alarm; if the actual deceleration response insufficient risk condition is met at the same time, the processor 120 will trigger the AI ​​road risk coefficient failure arbitration logic.

[0124] Once the processor 120 enters the single-vehicle risk assessment branch, the following dual-vehicle scenario-based collision time calculation process will no longer be executed in this cycle; when the oncoming vehicle status is valid or the system is in asymmetric degradation defense mode, the processor 120 will continue to execute the following dual-vehicle scenario-based collision time calculation process.

[0125] For data extraction from oncoming vehicles, processor 120 depends on the current local link status: when communication is normal, it extracts the actual oncoming vehicles. State dataset When the asymmetric degradation defense mode is triggered, because the cache has been overloaded at the underlying level, processor 120 directly extracts the virtual vehicle. Virtual state dataset Participate in subsequent calculations.

[0126] In conventional linear kinematic models, collision time is typically calculated simply by dividing relative distance by relative velocity. However, such models fail to account for the nonlinear amplification effect of reduced road surface friction and decreased braking efficiency of heavy vehicles on actual collision risk. To address this, processor 120 internally implements a scenario-based collision time calculation algorithm. This algorithm uses the dynamic road surface adhesion coefficient calculated by AI... Microscopic vehicle characteristics are introduced as a penalty factor into the basic time model, and the basic theoretical collision time is scaled through algebraic operations.

[0127] The specific mathematical formula is as follows:

[0128] ;

[0129] In the formula, The scene-based collision time scalar obtained by processor 120 is in seconds; The physical arc length of the blind zone is expressed in meters. and These are the instantaneous spatial distances from the local and the oncoming vehicles to their respective node detection surfaces, in meters; and These are the instantaneous scalar speeds of vehicles in both directions, in meters per second. A minimum speed constant is preset for the system, for example, a value of 0.1 m / s, to prevent division by zero calculation errors when both vehicles are stationary; and These are the dynamic road surface adhesion coefficient and the static reference road surface longitudinal adhesion coefficient, respectively, calculated from the AI ​​output. The ratio of these two values ​​is used to characterize the degree of braking capacity degradation caused by the decrease in frictional resistance under the current road conditions. and These are the braking weighting coefficients for both vehicles, and the average of the two is used to characterize the degree of degradation in the overall mechanical braking capability of both vehicles.

[0130] When the system is in asymmetric degradation defense mode, the parameters in the formula , and Equivalently replaced by virtual vehicle extreme values , and .

[0131] Calculate the scene-based collision time The processor 120 then compares it with a preset set of time thresholds. This set of time thresholds typically includes a primary alarm threshold, for example configured to be 5 seconds, and a critical alarm threshold, for example configured to be 3 seconds, to drive different response levels of subsequent actuators.

[0132] S402 executes the state machine's monotonic latching and degradation release mechanism.

[0133] The warning control state machine inside processor 120 includes at least four states: normal monitoring, primary alarm, severe alarm, and extreme intervention. The normal monitoring state outputs basic speed limit warnings or keeps the alarm module 160 in sleep mode; the primary alarm state prompts the vehicle to slow down and avoid obstacles; the severe alarm state outputs high-intensity audible and visual warnings; and the extreme intervention state directly outputs the highest-level warning when the local link degrades, the AI ​​road risk factor fails, or the collision time falls below the severe threshold. The state machine allows for instantaneous transitions from lower to higher-level states, but transitions from higher to lower states must meet hysteresis release conditions.

[0134] When processor 120 determines the instantaneous scalar speed of the target vehicle Speed ​​exceeding the dynamic recommended speed limit or scenario-based collision time When the time threshold is less than the preset threshold, the processor 120 triggers an alarm command of the corresponding level. Subsequently, the processor 120 controls the internal state machine to enter a one-way latching cycle.

[0135] During the one-way latching cycle, the processor 120 sends a high-level warning data frame to the display module 150, such as forcibly displaying highlighted characters for deceleration and avoidance, and outputs a high-duty-cycle pulse width modulation signal to the alarm module 160 to activate high-frequency audible and visual flashing. During this period, the state machine executes filtering logic, directly ignoring the differential deceleration fluctuations caused by the segmented braking of the target vehicle.

[0136] As long as the target vehicle speed If the strict release conditions are not met, processor 120 will forcibly maintain or adjust for a shorter time. The calculation results upgrade the current alarm intensity, limit the system to downgrade, or cut off the signal output of alarm module 160.

[0137] The processor 120 is internally configured with hysteresis comparator logic to address the boundary constraints for degrade removal. The alarm removal condition is not set to the target vehicle speed. Return to dynamic recommended speed limit Instead, it is set as Subtract a preset hysteresis bias ,For example The value is taken as 5 km / h. During numerical calculations performed internally by processor 120, if the speed unit output by radar module 110 is... storage units and If the configuration units are different, the processor 120 will first convert the above speed quantities into meters per second before making comparisons.

[0138] The speed of the target vehicle is determined only when the processor performs multiple consecutive calculation cycles, such as 5 consecutive seconds. Substantial reduction and below, and the latest calculated scene-based collision time If the speed exceeds the upper limit of the preset time threshold group and there is no risk of insufficient braking space, the risk of exceeding the limit is eliminated, and the monotonous latching state is released, restoring the silent or basic speed limit display in the normal monitoring mode.

[0139] Furthermore, when radar module 110 continuously No target vehicle was detected during the detection cycle. Alternatively, processor 120 can determine the target vehicle based on continuous distance changes. When the target has moved far from the blind zone entrance and the opposing node has not returned a valid risk target, the processor 120 can also determine that the current alarm object has left the risk area and release the monotonic latching state of the corresponding target. The number of consecutive detection cycles required to determine the loss of a target can be configured from 3 to 10, depending on the detection frame rate of the radar module 110, the speed of road vehicles, and the tolerance for false detection.

[0140] See attached document Figure 7 Step S50 is mainly used to capture the discrepancy between the road risk coefficient output by AI and the actual local vehicle response, and to trigger forced overloading and cross-level defense intervention of local physical limit parameters. Specifically, it may include the following steps:

[0141] S501 performs real-time theoretical braking space boundary derivation.

[0142] In conventional physics, a vehicle's braking distance is directly proportional to the square of its initial velocity and inversely proportional to the frictional deceleration provided by the road surface. Based on this principle, during the system's detection cycle, the processor 120 acquires the target vehicle data measured by the radar module 110 in real time. instantaneous scalar velocity .

[0143] When instantaneous scalar velocity The speed exceeds the system's current dynamic suggested limit. At that time, processor 120 calls the currently stored dynamic road surface adhesion coefficient in memory. and target vehicle Corresponding braking weight coefficient The processor 120, based on a rigid body kinematics physical model, calculates the target vehicle's speed in real time. Using the maximum physical braking efficiency provided by the current road surface adhesion conditions, reduce the speed from Reduce to The shortest theoretical braking distance required .

[0144] If instantaneous scalar velocity Less than or equal to The processor 120 determines that the current vehicle status is within the speed safety range and calculates the shortest theoretical braking distance. It is directly assigned the value of zero and subsequent boundary deduction calculations are skipped.

[0145] Calculate the shortest theoretical braking distance The corresponding mathematical derivation formula is as follows:

[0146] ;

[0147] In the formula, This is the scalar value for the shortest theoretical braking distance obtained by the processor 120, in meters; The acceleration due to gravity is constant, with a value of 9.8 m / s². 2 ; For AI-based road surface risk coefficient The calculated dynamic road surface adhesion coefficient.

[0148] S502 performs AI road risk coefficient and local micro-vehicle response divergence event capture.

[0149] In each computational scheduling cycle, the processor 120 will process the target vehicles measured in real time by the radar module 110. The instantaneous spatial distance between the probe and the detection surface, and the shortest theoretical braking distance obtained in step S501. Numerical comparisons are performed. This comparison process is used to determine whether the vehicle has reached the spatial critical line where extreme physical braking must be applied.

[0150] When processor 120 determines the instantaneous spatial distance Less than or equal to the shortest theoretical braking distance At this point, the system confirms that the target vehicle has crossed the spatial critical boundary. Under this critical state, the processor 120 further monitors the microscopic deceleration characteristics of the target vehicle.

[0151] If, within a consecutive preset time period after crossing this boundary, for example, three consecutive radar detection frames, the instantaneous scalar velocity of the target vehicle... Still exceeds the dynamic recommended speed limit If the actual deceleration value calculated by the velocity derivative is less than the system's preset deceleration threshold, it indicates that the target vehicle has a risk of insufficient actual deceleration response. When the actual deceleration value is greater than the velocity of the previous detection frame, the processor 120 divides the difference between the velocity of the previous detection frame and the velocity of the current detection frame by the time interval between adjacent detection frames to obtain a non-negative value; when the velocity of the current detection frame is greater than or equal to the velocity of the previous detection frame, the processor 120 sets the actual deceleration value to zero.

[0152] The deceleration threshold is derived by multiplying the theoretical maximum braking deceleration by a tolerance factor. The formula can be:

[0153] ;

[0154] The tolerance coefficient can be configured between 0.5 and 0.7 according to the system sensitivity requirements. Based on this, processor 120 determines the road surface risk coefficient of the target road segment as output by AI. The risk of inconsistency with local micro-vehicle response, and presumed based on conservative control principles, is considered. Possibly low, due to The converted The value may be too high, which could trigger the underlying failure arbitration logic.

[0155] S503 executes local physical limit parameter forced overload and over-level intervention.

[0156] After the failure arbitration logic is triggered, in order to address the risk of algorithm delay caused by the overly optimistic road risk coefficient output by the AI, the processor 120 suspends the use of newly received data from the wide-area module 130. It participates in risk calculation, but is still allowed to continue receiving and caching subsequent road risk parameter data packets for use in subsequent recovery verification.

[0157] The processor 120's memory manager actively clears the current compute cache. , and It also forces the recall of a preset local extreme environment limit adhesion coefficient from the protected region of non-volatile memory. Line memory overwrite. This limiting adhesion coefficient. It is used to characterize the frictional properties of a road surface under completely icy conditions, and its value is fixed between 0.1 and 0.2.

[0158] After overwriting the memory parameters, processor 120 forcibly sets the dynamic road surface adhesion coefficient to:

[0159] ;

[0160] And according to Recalculate dynamic recommended speed limit :

[0161] ;

[0162] exist During the lock period, processor 120 As a current dynamic recommendation, speed limits are advised. The processor 120 participates in the speed limit warning, minimum theoretical braking distance calculation, and alarm state machine judgment of the display module 150. Replace the original formula The task of re-executing the scenario-based collision time calculation instruction is then executed. Due to the significant reduction in the base friction coefficient, the re-calculated collision time will be significantly reduced, thereby directly meeting the triggering condition of the underlying severe alarm threshold.

[0163] Based on this, the processor 120 issues a bypass intervention command, which bypasses the step-by-step jump sequence of the conventional state machine and directly controls the display module 150 to emit light at the preset highest brightness level and output the warning character of the highest danger level. At the same time, it drives the alarm module 160 to output a continuous high-frequency audible and visual warning at the set maximum pulse duty cycle.

[0164] After entering the local physical limit parameter forced reload mode, processor 120 maintains... At least one preset locking period. After the locking period expires, if no actual insufficient deceleration response risk event occurs again within several consecutive radar detection periods, and the latest road risk parameter data packet cached by the wide-area module 130 passes timestamp verification, cyclic redundancy check, and numerical range verification, then the processor 120 releases the lock. Locked state, restore effective cloud-based delivery And recalculate and If the risk of insufficient actual deceleration response continues to occur, processor 120 will continue to maintain... And maintain a state of extreme intervention.

[0165] To enable those skilled in the art to more clearly understand the purpose, technical solution, and advantages of this invention, the present invention will be further described in detail below with reference to specific application embodiments, real experimental test data, and corresponding drawings. It should be noted that the embodiments described herein are only for explaining the present invention and are not intended to limit the scope of protection of the present invention.

[0166] I. Specific Application Examples

[0167] The application scenario is a series of sharp downhill curves in a mountainous area. After a rain in early winter, the temperature suddenly drops below 0 degrees Celsius at night, and the road surface gradually changes from wet and slippery to locally icy.

[0168] S10, Static boundary initialization:

[0169] During system deployment, technicians input static spatial geometric boundary parameters. This includes the radius of curvature of the curve. m, road superelevation cross slope The physical arc length of the blind zone between the first node 100 and the second node 200 m, longitudinal adhesion coefficient of static reference pavement (Corresponds to dry asphalt pavement).

[0170] Processor 120 maps the calculation results to a static safe passage speed reference value using an internal scaling factor. m / s (approximately 54 km / h).

[0171] S20, cloud-based downlink and dynamic rate limiting updates:

[0172] At 8:00 PM, the cloud server 300 obtained the rainfall data for this section of the road. mm / h, ambient temperature ℃. The AI ​​environmental risk identification model outputs the road surface risk coefficient. .

[0173] After receiving the data, processor 120 calculates the dynamic road surface adhesion coefficient. :

[0174] ;

[0175] According to this The processor 120 calculates and updates dynamic recommendations for rate limiting. (Approximately 36 km / h).

[0176] S30-S40, High-frequency detection and preliminary collision solution:

[0177] At 20:15, a heavy truck (target vehicle) As it approaches the first node 100, radar module 110 determines its instantaneous spatial distance. m, instantaneous scalar velocity m / s. Because its radar cross-section is greater than the threshold, a braking weighting coefficient is assigned. .

[0178] Simultaneously, local module 140 receives information from the second node 200 regarding the oncoming light car (target vehicle). Data: Instantaneous spatial distance m, instantaneous scalar velocity m / s, braking weight coefficient .

[0179] Processor 120 extracts the above data and substitutes it into the scene-based collision time formula:

[0180] ;

[0181] ;

[0182] The result obtained at this time Since it is less than the primary alarm threshold (5.0s) and slightly greater than the critical alarm threshold (3.0s), the state machine triggers the primary alarm, and the display module 150 outputs deceleration and avoidance.

[0183] S50, Microphysical Deviation Capture and Extreme Overload:

[0184] The driver applied the brakes after seeing the initial warning. In the next dispatch cycle, the heavy truck entered a localized area of ​​black ice.

[0185] Processor 120 performs theoretical braking space boundary derivation:

[0186] m;

[0187] The instantaneous spatial distance measured at this time m, obviously The vehicle is still outside the spatial safety boundary and will continue to be monitored routinely.

[0188] However, as the vehicle continued driving, when it was only a few steps away from the blind spot entrance... At m, due to slippery black ice, the truck's speed was still as high as m / s (not reduced to) (within). Recalculated at this point. m, but oncoming vehicles were also approaching at high speed. It dropped rapidly.

[0189] Assuming another extreme loss of control scenario for a single vehicle, the truck is far from the node. m, speed up m / s (severe speeding).

[0190] At this point, the theoretical shortest braking distance is:

[0191] m;

[0192] When the vehicle entered At m time, m, officially crosses the critical boundary of space.

[0193] Processor 120 initiates micro-deceleration characteristic monitoring. Within three consecutive detection frames, the radar measured vehicle speeds as 16.0 m / s, 15.9 m / s, and 15.8 m / s, respectively. The frame interval is 0.2 s, and the actual deceleration values ​​are... m / s 2 .

[0194] The theoretical maximum braking deceleration is:

[0195] m / s 2 ;

[0196] With a preset tolerance coefficient of 0.6, the deceleration determination threshold is: m / s 2 .

[0197] because The system determined that the AI-generated road risk coefficient (macro-meteorological forecast) was out of sync with the local micro-physical reality.

[0198] Processor 120 Forced Heavy Load Limit Adhesion Coefficient , replacement .

[0199] Re-execute the scenario-based collision time calculation:

[0200] s;

[0201] Corrected collision time The alarm time dropped to 1.06s, far below the 3.0s critical alarm threshold. The state machine bypassed the normal sequence and directly drove the alarm module 160 to output the highest level of audible and visual suppression at full load.

[0202] II. Experimental Verification and Effect Comparison

[0203] To verify the overall effectiveness of the AI-based intelligent vehicle passing warning sentinel system of this invention, particularly the effectiveness of the aforementioned multidimensional nonlinear dynamics solution and underlying state machine logic, this study constructed a hardware-in-the-loop simulation test environment. The injected test dataset contains 1000 valid vehicle passing events, covering complex extreme conditions such as dryness, rainfall, localized black icing, and local communication loss.

[0204] Comparison group settings:

[0205] Experimental group: The scheme of this invention (algorithm C) fully runs all the control logic from S10 to S50 above, including dynamic road surface adhesion coefficient penalty factor, asymmetric degradation defense, monotonic latching of the underlying state machine, and forced reload mechanism of local physical limit parameters.

[0206] Control group:

[0207] Traditional algorithm group (Algorithm A): relies solely on traditional linear kinematics formulas ( ,in For time, For distance, The collision time is calculated based on vehicle speed, without incorporating road friction decay and heavy vehicle braking weights, and there is no monotonic latching or physical limit heavy load logic.

[0208] Static feedforward group (Algorithm B): It uses the road risk coefficient issued by the cloud to correct the speed limit and collision time, but it lacks local micro-physical response deviation verification (missing step S50) and degradation intervention mechanism.

[0209] The test results are as follows:

[0210] according to Figure 8 It can be seen that, under extreme attenuation conditions such as vehicles passing each other on dark ice, the solution of this invention has significant safety redundancy and backup capability.

[0211] During the wet and slippery phase from 0 to 2 seconds, the TTC curve output by the control group (Algorithm A) showed a linear and slow decline, failing to reflect environmental risks. However, the experimental group (Algorithm C) introduced a dynamic road adhesion coefficient ratio and a heavy vehicle braking weight penalty term into its underlying calculation formula, resulting in a significantly lower collision time than Algorithm A. This allowed the system to reach the 5.0s primary alarm threshold as early as 1.5 seconds earlier (approximately 2.8 seconds earlier than Algorithm A).

[0212] More critically, at the 4-second mark, when the target vehicle entered the dark ice area and experienced physical skidding, the control group algorithm B failed to detect this underlying physical change. Its TTC curve continued to decline smoothly as expected by the macroscopic parameters in the cloud, causing the system to miss a critical warning opportunity in this dangerous situation. In contrast, the experimental group (algorithm C) successfully captured the actual deceleration (0.5 m / s²) at 4.2 seconds. 2 It is far below the theoretical deceleration threshold (2.11 m / s²). 2 The microscopic physical deviation event immediately triggered the invalidation arbitration logic. For example... Figure 8 As shown by the solid line, the curve of algorithm C drops vertically here, instantly falling below the critical alarm threshold (3.0s), forcibly activating the highest level of audio-visual suppression intervention, and achieving a last-ditch defense at the physical limit edge.

[0213] according to Figure 9 It can be seen that when the driver implements micro-driving disturbances, the present invention effectively shields invalid oscillations, ensuring the smoothness and monotonicity of alarm output.

[0214] During the test, the driver performed three consecutive high-frequency braking maneuvers due to poor visibility or nervousness, causing drastic fluctuations in the speed derivative on the input side. In the control group (Algorithm A), the alarm level frequently jumped between 0 (no alarm) and 1 (primary alarm) with the speed fluctuations; this intermittent effect easily caused driver visual fatigue and decision-making confusion. However, in the experimental group (Algorithm C) using the present invention, after the primary alarm was first triggered at 1.2 seconds, the microcontroller immediately entered a monotonic latching cycle. Figure 9 As shown in the thick solid line step diagram, during this period, algorithm C directly ignored the high-frequency alternation of deceleration caused by point braking, and the alarm status remained monotonically at level 1; until the 6.5th second, when the target vehicle speed substantially dropped below the safety boundary and met the hysteresis delay condition, the alarm level smoothly dropped back to 0, eliminating the critical oscillation.

[0215] Comprehensive experimental statistics show that, under complex and severe weather conditions combined with impaired underlying communication, the false negative rate of severe collision events in this invention is significantly reduced from 14.2% in the traditional algorithm to 0.3%; the average lead time for effective primary warnings is increased by 2.1 seconds; and the number of invalid system oscillations caused by microscopic driving disturbances is reduced by 96%. These results fully verify the technical superiority and high reliability of this invention in one-way network outage defense, failure correction, and disturbance resistance of the underlying state machine.

[0216] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based intelligent vehicle-passing warning sentinel system, characterized in that, include: The first node and the second node are arranged at the ends of the road with limited visibility. The first node and the second node are physically symmetrically deployed at both ends of the target curve or narrow blind spot, and are used to monitor vehicles entering from the two opposite lanes respectively. Both the first and second nodes include a radar module, a processor, a wide-area module, a local-area module, a display module, and an alarm module. And cloud servers independently located on the external network; The radar module is used to extract the multidimensional kinematic features of the target vehicle; The wide-area module is used to establish a low-frequency asynchronous data channel between the processor and the cloud server; The local module is used to establish a high-frequency local synchronization loop between the first node and the second node; The cloud server is used to identify the current road surface risk status and send road surface risk parameters to the processor via the wide area module; The processor establishes electrical connections with the radar module, wide-area module, local-area module, display module, and alarm module, respectively. The processor is used to receive the multi-dimensional kinematic features of the target vehicle extracted by the radar module, obtain the state dataset of the oncoming vehicle through the local module, perform a scenario-based collision time calculation in combination with the road risk parameters, and drive the display module and alarm module to execute an alarm based on the solution results.

2. The AI-based intelligent vehicle meeting early warning sentry system according to claim 1, characterized in that, The high-frequency local synchronization loop includes a first loop and a second loop; The first loop is a low-frequency wide-area asynchronous confirmation loop consisting of the cloud server, the wide-area module, and the processor. It is used for the cloud server to send a road risk parameter data packet containing the road risk coefficient to the processor, and for the processor to send parameter reception confirmation information back to the cloud server. The second loop is a high-frequency local bidirectional synchronous loop composed of the first node, the second node and their respective local modules, used for bidirectional exchange of vehicle kinematic state data and confirmation of link health status within the bidirectional blind zone.

3. The AI-based intelligent vehicle meeting early warning sentry system according to claim 2, characterized in that, The cloud server is internally deployed with an AI environmental risk identification model, which is implemented using a decision tree model. The decision tree model includes a data preprocessing module, a feature construction module, a decision tree inference module, and an output mapping module. The decision tree model is used to receive current precipitation, ambient temperature, relative humidity, historical road surface adhesion coefficient, and historical vehicle deceleration response data of the target road area, convert them into feature vectors, perform node judgments level by level in the tree structure, and generate the road surface risk coefficient corresponding to the current road segment based on the leaf node results.

4. The AI-based intelligent vehicle meeting early warning sentry system according to claim 3, characterized in that, The processor is configured with a serial port receive interrupt triggering mechanism and a parameter parsing task branch; when the road risk parameter data packet is received, the processor responds to the interrupt and stores the data in the receive buffer, and the parameter parsing task branch performs timestamp verification, cyclic redundancy verification and numerical range verification. After the verification is passed, the processor calls the mutex lock mechanism or temporarily disables the global interrupt, performs an overwrite operation on the specified memory address storing the basic calculation parameters, and calculates the dynamic road surface adhesion coefficient and dynamic suggested speed limit based on the updated road surface risk coefficient. If the verification fails, the data expires, or the value goes out of bounds, the processor maintains the original parameters in memory or rolls back to the preset conservative road risk coefficient.

5. The AI-based intelligent vehicle meeting early warning sentry system according to claim 4, characterized in that, The radar module is used to extract the instantaneous spatial distance, instantaneous scalar velocity, and target radar cross-section of the target vehicle; The processor is used to classify targets into light and heavy vehicles based on the target radar cross-section and assign corresponding braking weight coefficients; when the target radar cross-section is greater than or equal to the classification threshold, it is determined to be a heavy vehicle and a heavy vehicle braking weight coefficient is assigned; when the target radar cross-section is less than the classification threshold, it is determined to be a light vehicle and a standard braking weight is assigned. The processor generates a micro kinematic state dataset from the instantaneous spatial distance, instantaneous scalar velocity, and braking weight coefficient.

6. The AI-based intelligent vehicle meeting early warning sentry system according to claim 5, characterized in that, The processor is used to count the number of heartbeat packets that were not successfully received within the internal sliding time window in real time and to calculate the local link packet loss rate. When the local link packet loss rate is greater than or equal to the communication failure threshold, the processor forces the switch to asymmetric degradation defense mode; the processor's memory manager intercepts the listening thread, virtually constructs a virtual vehicle model under extreme physical boundary conditions in the data structure reserved for oncoming vehicles, and forces the dataset of the virtual vehicle model into the bidirectional collision calculation cache to make up for the input loss.

7. The AI-based intelligent vehicle meeting early warning sentry system according to claim 6, characterized in that, The extreme physical boundary conditions include: setting the spatial position of the virtual vehicle to the physical vertex at the other end of the blind zone, and setting the instantaneous spatial distance to zero; setting the instantaneous scalar speed of the virtual vehicle to reach the maximum theoretical safe passage speed benchmark value determined based on the longitudinal adhesion coefficient of the static reference road surface for the current road segment; and setting the braking weight coefficient of the virtual vehicle to be the braking weight coefficient of the heavy vehicle.

8. The AI-based intelligent vehicle meeting early warning sentry system according to claim 5, characterized in that, The processor has a scenario-based collision time solution algorithm built inside it. The scenario-based collision time calculation algorithm uses the physical arc length of the blind spot, the instantaneous spatial distance of the local target vehicle, and the instantaneous spatial distance of the oncoming vehicle to determine the equivalent relative interval distance, and uses the instantaneous scalar velocity of the two-way vehicles to determine the relative velocity; and uses the ratio of the dynamic road surface adhesion coefficient to the static reference road surface longitudinal adhesion coefficient, and the average value of the braking weight coefficient of the two-way vehicles as a penalty factor, and performs nonlinear scaling on the basic theoretical collision time calculated from the relative distance and relative velocity through algebraic operations, and outputs the scenario-based collision time to participate in the early warning decision.

9. The AI-based intelligent vehicle meeting early warning sentry system according to claim 1, characterized in that, The processor is internally configured with an early warning control state machine and a hysteresis comparator logic; when an alarm command is triggered, the early warning control state machine enters a one-way latching cycle to forcibly maintain or upgrade the alarm intensity. The processor must meet the following conditions to release the monotonic latch state: the target vehicle speed is measured to drop to the value of the dynamic suggested speed limit minus the preset hysteresis bias within multiple consecutive operation cycles, the latest calculated scenario-based collision time is greater than the upper limit of the preset time threshold group, and the risk of insufficient braking space is not triggered.

10. The AI-based intelligent vehicle meeting early warning sentry system according to claim 1, characterized in that, The processor is used to capture the divergence event between AI road risk coefficient and local micro-vehicle response. The processor calculates in real time the shortest theoretical braking distance required for the target vehicle to reduce its speed to the dynamically recommended speed limit; when it determines that the instantaneous spatial distance is less than or equal to the shortest theoretical braking distance, the processor monitors the micro-deceleration characteristics of the target vehicle. If the target vehicle speed exceeds the dynamic suggested speed limit within a consecutive preset time period, and the actual deceleration value is less than the system's preset deceleration judgment threshold, the processor triggers the failure arbitration logic to execute the local physical limit parameter forced reload, clears the road risk coefficient, dynamic road adhesion coefficient and dynamic suggested speed limit in the calculation cache, and forcibly calls the preset local extreme harsh environment limit adhesion coefficient for memory overwrite.

Images