A dynamic monitoring and analysis method and system for a smart community vehicle

Abstract

The present application relates to the technical field of community management, and particularly relates to a dynamic monitoring and analysis method and system for a smart community vehicle, comprising: collecting multi-source data through a sensor, generating structured perception data using a lightweight OCR algorithm; according to the structured perception data, combining environmental parameters and equipment states, calculating the real-time weight of each sensor using an adaptive dynamic weight algorithm, calculating the fusion confidence through a feature confidence fusion formula, eliminating unqualified data, and generating high-precision perception data; constructing a community digital twin scene, judging abnormalities through a dynamic threshold algorithm, positioning faults through a three-level mechanism and generating a configuration update instruction, and generating system state data; inputting the high-precision perception data, system state data and vehicle historical violation records into a game theory model to calculate the vehicle violation probability, and generating a differentiated management strategy according to the vehicle violation probability. The present application realizes the precise management of community vehicles through multi-source fusion and hierarchical association.

Description

Technical Field

[0001] This invention relates to the field of community management technology, and in particular to a method and system for dynamic monitoring and analysis of vehicles in a smart community. Background Technology

[0002] Current community parking management faces multiple pain points, and traditional models are ill-suited to the growing demand for vehicle ownership. At the data collection level, single sensors have significant limitations: cameras are susceptible to backlighting and shadows, leading to inaccurate license plate recognition; ultrasonic sensors are affected by temperature interference, reducing distance accuracy; and geomagnetic sensors are prone to false triggering due to electromagnetic interference. Furthermore, the lack of a synchronous fusion mechanism for multi-sensor data often results in data mismatch and difficulty in quantifying confidence levels, hindering the formation of a reliable perception chain.

[0003] In terms of equipment management, reliance on manual inspections leads to delayed fault response. Anomaly judgment uses fixed thresholds without considering dynamic factors such as equipment usage frequency and weather, which easily results in false alarms and missed alarms. Fault diagnosis lacks a systematic mechanism, making it difficult to distinguish between software and hardware problems. Repair relies on manual experience, which cannot be iteratively optimized, resulting in a high equipment failure rate.

[0004] In vehicle behavior management, traditional strategies rely on fixed rules for control without taking into account vehicle history of violations or regional risk levels (such as the difference between core areas at entrances and exits and ordinary parking spaces). There is no quantitative model for violation risk, relying on subjective judgment. High-risk responses are not timely, while low-risk responses are over-controlled. Furthermore, strategy optimization lacks a data loop, making it difficult to balance governance and resource consumption.

[0005] Furthermore, data is fragmented across different stages, with a lack of coordination between data collection, fusion, decision-making, and execution. Perception data is not used for fault diagnosis, and fault results are not fed back for strategy optimization, resulting in low efficiency across the entire chain. Summary of the Invention

[0006] This invention achieves precise management of community vehicles through multi-source fusion and hierarchical association.

[0007] The technical solution proposed in this invention is: a dynamic monitoring and analysis method for vehicles in a smart community, the method comprising: Multi-source data is collected through sensors, and structured perception data is generated using a lightweight OCR algorithm; Based on structured sensing data, combined with environmental parameters and equipment status, an adaptive dynamic weighting algorithm is used to calculate the real-time weights of each sensor. The fusion confidence is calculated using a feature confidence fusion formula. Unqualified data is then removed based on the fusion confidence to generate high-precision sensing data. A community digital twin scenario is built based on high-precision sensing data. Anomaly detection is performed through a dynamic threshold algorithm. When an anomaly is detected, a three-level mechanism is used to locate the fault and generate a configuration update instruction. System status data is generated based on the fault handling results, digital twin scenario update information, and equipment operation status data. High-precision sensing data, system status data, and historical vehicle violation records are input into a game theory model to calculate the probability of vehicle violations, and differentiated management strategies are generated based on the probability of vehicle violations.

[0008] Preferably, the specific process for generating the structured sensing data is as follows: Vehicle image data is acquired via a camera, distance data within a specific range is acquired via an ultrasonic sensor, and the trigger status is output in real time via a geomagnetic sensor. The acquired data is preprocessed. The lightweight OCR algorithm of the edge computing box extracts features from the preprocessed image, generates feature maps of different resolutions, presets anchor frames corresponding to the license plate size, predicts the license plate frame coordinates and outputs the initial confidence score, and optimizes the confidence score through time consistency verification. After cropping the valid license plate frames that meet the confidence score, the character region is located, input into the classification model to recognize the characters, and the overall character recognition confidence score is calculated. The vehicle outline is extracted through an instance segmentation model, the aspect ratio is calculated, and the vehicle model code is mapped by combining a threshold. The presence of the vehicle is initially determined based on the geomagnetic trigger status and the calibrated distance. The positional deviation between the center of the license plate frame and the ultrasonic detection center is calculated, and data with deviations exceeding the threshold are discarded.

[0009] Preferably, the process of obtaining the high-precision sensing data is as follows: Based on historical equipment failure rates and environmental parameters, real-time trust weights for cameras, ultrasonic sensors, and geomagnetic sensors are calculated separately to ensure that the total weights are a specific value and that no single weight is lower than a minimum threshold. Any excess weights are deducted proportionally from the other weights. Weights are converted to meet different needs for coordinate fusion and confidence fusion. Coordinate fusion uses only the camera and ultrasonic sensor weights and normalizes them, while confidence fusion retains the original total weight system. The final vehicle coordinates are calculated using a weighted average. Based on the weights of each sensor, the fusion confidence is calculated by combining the product of the camera's positioning and character recognition confidence, the deviation ratio between the ultrasonic distance and the predicted distance, and the geomagnetic triggering state. Only data that meets the fusion confidence standard are retained.

[0010] Preferably, the specific process for anomaly detection is as follows: Based on the equipment's rated parameters, operating frequency, and adjustment coefficient, calculate the dynamic upper and lower limit thresholds for parameters such as equipment voltage and network signal strength. Adjust the network signal threshold based on the average daily number of transmission failures. Monitor equipment voltage and network signal strength in real time, recording whether they exceed the threshold, the duration of the exceedance, and the fluctuation amplitude within the threshold. If a parameter exceeds the threshold for a specific consecutive period, it is judged as abnormal. If a single period exceeds the threshold but does not reach a continuous period, record the fluctuation and increase the monitoring frequency. If the fluctuation amplitude within the threshold exceeds the rated value proportion, mark it as unstable and trigger self-calibration.

[0011] Preferably, the specific process for obtaining the configuration update instruction is as follows: The system calls the preset IF-THEN rule engine to initially screen for software faults. If no match is found, a deep neural network model outputs the fault probability distribution, which is then cross-validated using an expert rule base to confirm the software configuration issue to which the fault belongs. Key information is extracted from the fault diagnosis data to identify the names of the parameters that need to be adjusted, and the target values ​​of the parameters are determined based on the cause of the fault and historical repair cases. The execution time is determined by combining the real-time working status of the equipment and the scope of the fault's impact. If the faulty equipment is in a low-load period, it is set to execute immediately; if the equipment is processing core business, it is delayed until the off-peak period to avoid interrupting normal service due to configuration updates.

[0012] Preferably, the specific calculation process for the probability of vehicle violations is as follows: Clearly define the strategies of vehicle owners and the management system, and define the benefits and costs of both parties under different strategy combinations. Based on the regional coefficient, the probability of missed detection, and the additional monitoring cost, construct a non-zero sum benefit matrix for both vehicle owners and the system to quantify the benefit values ​​of both parties under different strategies. Let the probability of vehicle owner violations and the probability of enhanced monitoring by the system be variables. By calculating the expected benefits of both parties and setting their partial derivatives to 0, solve for the equilibrium violation probability under the Nash equilibrium state.

[0013] Preferably, the specific process of the differentiated management strategy is as follows: By combining regional risk coefficients, equilibrium violation probabilities solved using game theory, and real-time standardized speed and acceleration characteristics of vehicles, a weighted calculation is performed to obtain a vehicle violation risk score. If the vehicle violation risk score is greater than or equal to the high-risk threshold, an emergency interception is triggered at the gate, and a warning notification containing evidence of violation is generated simultaneously. If the score is in the medium-risk range, the gate allows normal passage but the vehicle is included in key monitoring, and a prompt for proper driving is sent. If the score is less than the low-risk threshold, regular management is maintained, and only vehicle behavior data is recorded. Through a reinforcement learning model, the strategy selection logic is iteratively optimized with a decrease in violation rate as an effect reward, strategy execution cost as a penalty, and speed fluctuation as a stability reward, and the reward weights are temporarily adjusted according to the real-time scenario.

[0014] The present invention also provides a method for dynamic monitoring and analysis of vehicles in a smart community, wherein the system is used to execute the method for dynamic monitoring and analysis of vehicles in a smart community.

[0015] The present invention also provides a computer-readable storage medium storing a computer program, which is executed by a processor to implement the aforementioned method for dynamic monitoring and analysis of vehicles in a smart community.

[0016] The beneficial effects of this invention are: 1. The multi-source data fusion mechanism achieves a leap in data quality through three steps: First, a preprocessing process is designed for different sensor characteristics, such as grayscale conversion, median filtering, and contrast enhancement optimization of camera images; ultrasonic data is combined with temperature calibration; and geomagnetic signals are verified in real time, reducing noise in the raw data from the source. Second, an adaptive dynamic weighting algorithm is introduced, combining historical equipment failure rates (such as the proportion of failure time in the past 30 days) and environmental parameters (light intensity, weather type) to calculate the trust weight of each sensor, ensuring that high-reliability data accounts for a higher proportion in the fusion. Finally, through spatial consistency fusion (unifying coordinate units and weighting vehicle positions) and feature confidence fusion (combining positioning, character recognition, distance deviation, and geomagnetic state), high-precision perception data with a fusion confidence score ≥0.9 is generated. Ultimately, this mechanism improves the license plate recognition accuracy from the initial 88% to over 96%, controls the distance data calibration error within ±2cm, and reduces the position deviation misjudgment rate by 60%, providing high-quality data support for subsequent digital twin modeling, violation risk calculation, and other stages, avoiding management decision-making errors caused by unreliable data.

[0017] 2. The digital twin and intelligent self-healing system upgrades equipment management through dual innovation: On the one hand, it constructs a community digital twin scenario based on high-precision sensing data, mapping vehicle location and equipment operating status (voltage, network signal) in real time, and uses a dynamic threshold algorithm to calculate the reasonable range of equipment parameters (such as adjusting voltage thresholds according to usage frequency), improving the accuracy of anomaly detection to 92%. On the other hand, it establishes a three-level fault diagnosis mechanism. Over 200 preset rules can quickly locate common software faults (such as incorrect parameter configuration), and a deep neural network model (inputting 12-dimensional fault features) can identify complex hardware problems (such as power module aging). After diagnosis, it issues software configuration update commands via the MQTT protocol (confirmed within 30 seconds) or generates priority work orders, reducing the average repair time for hardware faults to within one hour. In practical applications, this system reduces the equipment failure rate from 15% to below 5%, reduces manual intervention by 70%, and increases the system's annual fault-free operation time to over 340 days, significantly reducing maintenance costs and ensuring continuous and stable parking management services.

[0018] 3. The hierarchical management mechanism achieves refined management through risk quantification, differentiated control, and iterative optimization: First, a two-person non-zero-sum model is constructed based on game theory to quantify the benefits (time value saved) and costs (fines) of vehicle owners' violations, and the benefits (avoiding social costs) and costs (additional expenses) of system monitoring. The equilibrium violation probability is solved, and combined with regional risk coefficients (0.8 for core areas, 0.5 for secondary core areas, and 0.3 for ordinary areas) and real-time vehicle behavior (standardized speed and acceleration), a violation risk score of 0-100 is generated. Second, risk levels are divided into high (≥80 points), medium (60-79 points), and low (<60 points) risk levels according to the score, and corresponding differentiated strategies are triggered to avoid resource misallocation. Finally, reinforcement learning is introduced for iterative optimization, with the decrease in violation rate as the effect reward, the strategy execution cost as the penalty, and speed fluctuation as the stability reward. Feature weights are dynamically adjusted (e.g., increasing the weight of speeding features in core areas) to ensure that the management strategy continuously adapts to the actual scenario. Application data shows that this mechanism has increased the accuracy of violation processing in the core area from 68% to 94%, stabilized the average response time within 1.2 seconds, improved vehicle traffic efficiency by 30% during high-traffic periods, and reduced monitoring resource consumption by 25%, achieving multiple goals of "precise governance, efficient traffic, and low-cost operation and maintenance". Attached Figure Description

[0019] Figure 1 This is a flowchart of a dynamic monitoring and analysis method for vehicles in a smart community according to the present invention; Figure 2 This is a flowchart illustrating the monitoring process of a dynamic monitoring and analysis method for vehicles in a smart community according to the present invention. Detailed Implementation

[0020] The following description is intended to disclose the present invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art. The basic principles of the invention defined in the following description can be applied to other embodiments, modifications, improvements, equivalents, and other technical solutions that do not depart from the spirit and scope of the invention.

[0021] It is understood that the term "a" should be understood as "at least one" or "one or more," that is, in one embodiment, the number of an element can be one, while in another embodiment, the number of the element can be multiple, and the term "a" should not be understood as a limitation on the number.

[0022] like Figure 1 and Figure 2As shown, after the system starts, a regular camera, ultrasonic sensor, and geomagnetic sensor simultaneously collect vehicle data. The camera captures vehicle images at a resolution of 1920×1080 and a frame rate of 25fps, providing a clear visual foundation for subsequent license plate and vehicle model recognition; the ultrasonic sensor collects distance data within a range of 0.1-5m at 0.5-second intervals (accuracy ±2cm) to determine the actual distance between the vehicle and the sensor; the geomagnetic sensor outputs the trigger status in real time (0 for no trigger, 1 for trigger, denoted as M), intuitively reflecting whether the vehicle has entered the detection area.

[0023] In the raw data preprocessing stage, the camera image is first converted to grayscale, compressing the RGB three-color channel information into a single-channel grayscale value to reduce color interference; then, a median filtering algorithm is used to take the median of a 5×5 pixel neighborhood to effectively remove salt-and-pepper noise (such as isolated white spots caused by sensor noise); finally, a contrast enhancement algorithm is used to stretch the dynamic range of the grayscale histogram (from... Mapped to [0,255]), this solves the problem of blurred license plates in scenarios such as backlighting and shadows. Raw distance data from the ultrasonic sensor. It needs to be considered in conjunction with the ambient temperature. (Unit: °C) Calibration is performed using the following formula: This is because the speed of sound in air varies significantly with temperature (the speed of sound is 331.4 m / s under standard conditions, and increases by 0.607 m / s for every 1°C increase in temperature). For example, at a raw distance of 1.8 m in a 28°C environment, an uncalibrated distance will produce an error of 0.29 cm. After calibration, the distance accuracy can be controlled within ±2 cm. Data from the three types of sensors are aligned through a timestamp synchronization mechanism to ensure an error of ≤0.1 seconds, avoiding the problem of "image and distance data mismatch" caused by clock deviation.

[0024] The preprocessed data is transmitted to an edge computing box, where a lightweight OCR algorithm (based on a simplified version of YOLOv8n) begins deep processing. The object detection model first extracts features from the image, generating feature maps at three resolutions: 8×8, 16×16, and 32×32. Three anchor boxes are preset for the license plate sizes of common vehicles in the community: small cars (440×140 pixels), large cars (520×160 pixels), and new energy vehicles (480×150 pixels). The top-left corner coordinates of the license plate frame are predicted through bounding box regression. and the coordinates of the bottom right corner At the same time, output the initial confidence level. ( There is a probability that the target exists. For each category probability, all are mapped to the [0,1] interval using the sigmoid function. Considering that the speed of vehicles in the community is ≤20km / h, the same vehicle may appear in 3 consecutive frames, so a temporal consistency check is introduced for optimization. : ,in, , , These are the initial confidence levels for the current frame and the two previous frames, respectively. Dynamic weight (the change in the position of the license plate frame over three consecutive frames) 5 pixels (position stability threshold) )hour, =1.1; otherwise =0.9). For example, if the initial confidence levels of three consecutive frames are 0.88, 0.90, and 0.92, and the positional change is 3 pixels each, then... .

[0025] During the license plate character recognition stage, (Effective threshold for license plate location) The valid license plate frame is cropped, and the character regions are located using horizontal projection (7 digits for small cars / new energy vehicles, 8 digits for large vehicles). Each region is input into the MobileNetV2 lightweight classification model. The confidence score for individual character recognition is... Take the highest probability that the character belongs to all candidate categories. Overall character recognition confidence score. Weighted calculation is used: 7-digit license plate (small car / new energy vehicle): ; 8-digit license plate (large vehicle): ; The first two characters (Chinese characters and letters) have higher weight because they are more crucial for uniquely matching license plates.

[0026] Vehicle model recognition extracts the vehicle outline using an instance segmentation model and calculates the aspect ratio. (width The maximum horizontal distance of the outline, height (the maximum distance in the vertical direction), and grouped according to vehicle type classification thresholds. Mapped to vehicle model code : when (satisfy hour, (Small car, 7-digit license plate); when (satisfy )hour, (SUV, 7-digit license plate); when (satisfy )hour, (Large vehicle, 8-digit license plate).

[0027] Among the above thresholds, , , , , , used to distinguish different vehicle models and adapt to the corresponding number of license plate characters.

[0028] The real-time verification process adopts the "multi-sensor mutual verification" mechanism: when the geomagnetic state and the calibrated distance (there is a vehicle distance threshold ), it is preliminarily determined that the vehicle exists (to avoid false alarms of a single sensor. For example, when the geomagnetic is mis-triggered by electromagnetic interference, if the ultrasonic detection distance ≥ , it is excluded); then calculate the Euclidean distance between the center coordinates of the license plate frame and the center coordinates of the vehicle detected by ultrasonic: , (unit: m). When (the position deviation threshold ), the data is excluded (to eliminate position misalignment caused by camera perspective deviation or ultrasonic reflection angle).

[0029] The verified data is integrated into structured perception data , where the new parameter is the number of license plate characters (7 or 8), determined by the vehicle model code ( when, , when ), and other parameters include: is the recognized license plate number (such as "Hu B56789" for a small vehicle and "Hu B123456" for a large vehicle); is the center coordinates of the license plate frame; is the data acquisition timestamp; , are the positioning and character recognition confidence levels respectively; is the calibrated distance; is the geomagnetic state.

[0030] Taking the actual scenario of a community entrance as an example: the environmental temperature °C, a truck (large vehicle) drives into the detection area. The camera captures the license plate frame coordinates (150, 200, 350, 300), with initial confidence levels of 0.85, 0.87, and 0.89 for 3 consecutive frames and a position change of 4 pixels (less than ), and it is calculated that The confidence scores for the eight characters are 0.92, 0.88, 0.84, 0.86, 0.90, 0.87, 0.85, and 0.89, respectively. The calculated... Vehicle length-to-width ratio (less than) ),determination ,correspond Original distance of ultrasound After calibration (less than) Geomagnetic state Center of the license plate frame Ultrasonic Testing Center Calculated The verification passed. The final generated structured sensing data is as follows: This data will be transmitted to subsequent modules for in-depth analysis.

[0031] To continuously optimize recognition performance, the system collects false detection data from manual review weekly and labels the actual license plates. Model and number of characters Construct an incremental training set. Update the OCR model parameters using the Adam optimizer. , with loss function: To optimize the objective, For cross-entropy, The mean squared error is used. Through iterative training, the model is adapted to the license plate features of different vehicle types. The actual test shows that after three rounds of iteration, the license plate recognition accuracy can be improved from the initial 88% to more than 96%.

[0032] Based on structured sensing data High-precision sensing data is used to calculate environmental parameters and equipment environmental parameters. Environmental parameters include time characteristics. ( The hour range is 0-23, obtained through the system clock; (Weekly status, 0 = weekday / 1 = weekend, calculated from the date) and weather status. ( Weather type is 0 = sunny / 1 = rain / 2 = fog, obtained through an external meteorological sensor; Light intensity (unit: lux, collected via the camera's built-in light sensor module) and device status parameters include historical failure rates. and real-time status codes (0 = normal / 1 = downgraded / 2 = fault, obtained through equipment self-test signals).

[0033] Among them, the proportion of camera malfunctions in the past 30 days The calculation process is as follows: Let the total duration of camera malfunctions in the past 30 days be... (Unit: hour) The results are limited to the range of 0-1; ultrasonic sensor failure rate and geomagnetic sensor failure rate The calculation method and Consistent, respectively , ,in , These represent the total fault duration of the corresponding sensors over the past 30 days, calculated from the start and end times of the faults recorded in the equipment logs.

[0034] The first step in processing is to calculate the real-time trust weights of each sensor. ,satisfy Furthermore, the lower limit for the weight of each item is 0.1, among which the weight of the camera is... Ultrasonic weighting ( This is a weather coefficient, corresponding to Take 1 at a time Take 0.8 at the time. (Take 0.6 at time); Geomagnetic weight If the calculated result is lower than 0.1, it will be forcibly set to 0.1, and the excess will be deducted from other weights proportionally to maintain the sum at 1.

[0035] Before performing spatial consistency fusion, a weight transformation module needs to be executed. This module is specifically used to convert the total weight system into the sub-weights required by each functional module. For coordinate fusion weight transformation, since the geomagnetic sensor does not participate in position measurement, only the weights of the camera and ultrasonic sensor are used for coordinate calculation. Therefore, it is necessary to... and Normalize to ensure the sum of the weights is 1, the formula is: , , When using confidence fusion weights, the original total weight system is used. Because of the geomagnetic state Participate in confidence level calculation.

[0036] Next, a coordinate system transformation is performed using a pre-calibrated transformation matrix. (A 3×3 matrix, obtained by placing markers with known coordinates in the detection area, collecting pixel coordinates, and then solving using the least squares method) This represents the camera pixel coordinates. Convert to metric coordinates The conversion formula is: Ensure that the metric coordinates are directly output by the ultrasonic sensor. The units are consistent.

[0037] Spatial consistency fusion uses a weighted average to calculate the final coordinates of the vehicle. The feature confidence fusion formula is: ,in This is a distance prediction value based on historical data from the same period, obtained by querying the same data from the previous 30 days. , , The distance data under the given conditions is calculated by averaging. To integrate confidence levels, only retain The data.

[0038] In handling misjudgment cases, when Or, discrepancies may be discovered during manual verification (incorrect license plate or incorrect position). When the value is 0.3m, it is marked as a misjudgment case. ( It is a real license plate. (The real-world locations are obtained through manual annotation). Misclassified cases are input into the training set weekly, and a normalized loss function is used. Update the fusion algorithm parameters, where the MSE term is normalized by dividing by the square of the maximum permissible error (range 0-1). When a sensor weight... If the value is less than 0.2 for 10 consecutive minutes, a maintenance warning is triggered (this is achieved by periodically polling the weight value and accumulating the duration).

[0039] Output high-precision sensing data middle, The process of obtaining (accurate license plate number) is as follows: When The initial recognition result is used directly. Otherwise, mark it as "pending review"; (Latitude and longitude coordinates) are obtained through Gauss-Kruger projection transformation to detect the latitude and longitude of the region's center point. Using metric coordinates as the reference, Convert to latitude and longitude offset, the formula is: , Conversion error ≤ 0.5m; To achieve the fusion confidence score, the score is directly taken from the feature confidence fusion result and is ≥0.9; and The timestamp and vehicle model code from the structured perception data are used respectively; 111319.49 is the total base duration within the statistical period determined based on actual production conditions, which can vary according to actual production conditions.

[0040] For example, in a community's smart parking system, at 17:00 on a certain weekday ( , The weather was rainy. , The camera's downtime over the past 30 days was 72 hours. ), the failure duration of the ultrasonic sensor in the past 30 days was 36 hours ( ), the geomagnetic sensor had no failure ( ), and the real-time status of all three was normal ( ).

[0041] Calculate the sensor weights: Camera weight ; Ultrasonic weight ; Geomagnetic weight ; Execute the weight conversion model, and normalize the coordinate fusion weights , .

[0042] The structured perception data shows the initial license plate recognition result (compact car, , ), the center pixel coordinates of the license plate frame recognized by the camera , the metric coordinates after conversion by the conversion matrix , the vehicle center coordinates detected by the ultrasonic wave , the license plate positioning confidence , the character recognition confidence , the calibrated ultrasonic distance , the geomagnetic status , the distance predicted according to historical data . .

[0043] The final vehicle coordinates ; Fusion confidence .

[0044] This data was marked as a misjudgment case. After manual review, it was found that the actual license plate was "Zhe A12346", and this case was included in the training set to update the model parameters. At the same time, since the camera weight continued to be lower than 0.2 within the next 15 minutes, the system triggered a camera maintenance warning to prompt the staff to check the equipment.

[0045] Based on the high-precision perception data and the device operating status (voltage, network signal strength, etc.), the system realizes digital twin modeling, fault self-healing, and hierarchical behavior management decision-making within the community through multi-module collaboration. The device operating status data is collected in real time by the built-in monitoring modules of each sensor. The voltage value is obtained through the sensor power management chip, and the network signal strength is obtained by converting the RSSI value of the communication module.

[0046] In the digital twin scenario construction link, with high-precision perception data Driven by the core mechanism, the system updates the location and status of vehicle models in the community's virtual scene every 0.5 seconds (e.g., "identified" or "pending payment"). For community parking lots with more than 100 spaces, the system automatically enables Level of Detail (LOD) technology to reduce rendering accuracy for distant vehicles; if an underground parking garage exists, a 3D spatial mapping is constructed based on BIM data. Simultaneously, device status parameters (e.g., camera voltage below 80% of rated value, ultrasonic sensor network signal strength <20dBm) are mapped to the digital twin device model in real time. When an anomaly occurs, the corresponding icon flashes red, and maintenance personnel can click to retrieve details.

[0047] The system uses a dynamic threshold algorithm to detect anomalies in equipment operating status. In voltage monitoring, the rated voltage of the equipment is assumed to be... (Standard operating voltage set by the factory), the average daily usage frequency of the equipment is... The equipment's baseline number of uses is (Baseline value for average usage frequency of community equipment), with an adjustment coefficient related to equipment usage frequency as follows: (Highly used entrance and exit cameras in the community) Low-frequency environmental sensors Then the dynamic adjustment coefficient Dynamically adjusted lower voltage threshold Dynamically adjusted upper voltage threshold Network signal strength is normally within the range of 20-80 dBm. Adjustments are made based on the average daily number of transmission failures. This applies when device voltage, network signal strength, and other status parameters are continuously... An abnormality is determined when a cycle exceeds the corresponding threshold range, including abnormal voltage at the entrance / exit gates. (10 seconds / cycle) Geomagnetic sensor signal abnormality (30 seconds / cycle). If the voltage exceeds the threshold within a single cycle but does not reach a continuous value... In each cycle, the system will record fluctuation events and increase the monitoring frequency (e.g., shortening from 10 seconds / cycle to 5 seconds / cycle); if the voltage is within the threshold range but the fluctuation amplitude exceeds 5% of the rated value (i.e., , The voltage of the current cycle. If the voltage is the same as the voltage of the previous cycle, it is marked as "voltage unstable" and the device self-calibration process is triggered.

[0048] Upon triggering an anomaly, the system employs a three-level troubleshooting mechanism. The system first invokes over 200 pre-defined rules for rapid diagnosis. These rules utilize an IF-THEN logic structure, for example: IF device voltage. And after restarting THEN is determined to be "temporary power fluctuation"; IF network signal strength And other equipment in the same area The result is determined to be a "single device network module failure". The rule engine supports dynamic updates, and operations personnel can add new rules through the low-code interface. If the rule matches successfully, the cause of the failure is directly output and the self-healing process begins; otherwise, the machine learning-based fine-tuning stage begins.

[0049] When the rules engine cannot find a match, the system invokes a deep neural network (DNN) fault diagnosis model for analysis. The model's input layer receives the current abnormal data vector. It contains 12 feature dimensions: ,in This refers to the voltage fluctuation amplitude. This represents the hourly interval, with a value range of 0-23. This indicates the day of the week, with a value ranging from 1 to 7. The weather type is coded as follows: 0 represents sunny, 1 represents rainy, and 2 represents snowy. Ambient light intensity, in lux; This refers to the total number of times the equipment has malfunctioned in the past 30 days; This refers to the cumulative operating time of the equipment since it was put into use, expressed in days. This refers to the number of core business modules directly associated with the faulty equipment, such as vehicle recognition, gate control, and payment systems. Non-core modules, such as environmental temperature and humidity monitoring, are not included in the calculation. This represents the ratio of the device's real-time workload to its rated load, ranging from 0 to 1. The model's hidden layers employ a 3-layer fully connected structure (with 64, 32, and 16 neurons respectively), using ReLU as the activation function. The output layer outputs 10 types of fault probability distributions using the Softmax function. The system randomly selects 1000 similar cases from the historical fault case database as a reference set and calculates the Euclidean distance-weighted similarity between the current case and the reference set. ; in: For Euclidean distance, For reference cases The probability of fault type is calculated, and the top 3 fault types with the highest similarity are selected as candidate results.

[0050] The system cross-validates the candidate results output by machine learning with an expert rule base. The rule base contains 50+ heuristic rules developed by domain experts, such as: IF candidate fault is "sensor drift" and device usage time... 365 days, THEN failure probability increases by 30%; IF multiple sensors report abnormalities simultaneously and the current weather is heavy rain ( Then, environmental interference factors are given priority. The confidence level of candidate faults is adjusted through rule-based weighted calculation, and the result with the highest confidence level is selected as the final diagnostic conclusion. If the highest confidence level is lower than 70%, the system will automatically generate a diagnostic work order containing the original data and intermediate calculation results, and push it to the expert panel for manual review.

[0051] Upon diagnosis, a self-healing process is initiated. Software malfunctions generate configuration update commands. Sending and confirming via MQTT within 30 seconds; generating work orders for hardware failures, with work order priority. Among them, the fault impact range coefficient The fault urgency coefficient is calculated based on the number of core business modules associated with the faulty equipment (adding 0.2 for each associated core module; core modules include vehicle identification, gate control, payment systems, etc.; non-core modules such as environmental temperature and humidity monitoring are not included in the calculation). Based on fault type (Level 1 = 1, Level 2 = 0.6, Level 3 = 0.3), the fault repair difficulty coefficient is... Based on the historical average repair time (add 0.1 for every hour exceeding the limit).

[0052] When optimizing the federated learning strategy, each community subsystem constructs a local training sample set. , where the feature vector ( Weather type For hourly segments, (Voltage during fault) Label For optimal handling measures (such as "rebooting the device" or "replacing the power module"), the subsystem uses local model training based on Gradient Boosting Tree (GBDT) to initialize global model parameters. In each training round: Each subsystem downloads the current global model. Calculate gradients using local data ( (where the cross-entropy loss function is used). Gradient encryption is performed using a secure aggregation algorithm: the masking mechanism is modified by adopting a "subsystem mask pair cancellation" strategy, whereby subsystems are randomly divided into several groups (each group has two or more subsystems), and the random masks generated by the subsystems within each group satisfy the following conditions: (e.g., subsystem 1 is generated) Subsystem 2 generation ), calculate encryption gradient Upload it while retaining the mask key; Aggregator server collects all Calculate the sum After the subsystem uploads the mask key, the server decrypts it to obtain... Update the global model ( For learning rate, (Number of subsystems) Repeat the iterations until the model converges (the accuracy on the validation set improves by less than 1% for three consecutive rounds).

[0053] For example, in a community on a rainy day ( In a scenario where an ultrasonic sensor experiences false alarms due to raindrop interference, the local model learns from 1000 historical cases and determines the optimal strategy as "increasing the signal threshold from 20dBm to 25dBm." The gradient parameters of the corresponding feature dimensions are encrypted and uploaded, then aggregated with similar data from other communities to form a global strategy. The system evaluates the strategy's effectiveness every 24 hours using a test set. If a strategy improves the fault handling success rate by more than 20% in three or more communities, it is automatically encoded into a standardized instruction executable by the rule engine (e.g., ...). ), synchronized to all community subsystems.

[0054] The abnormal and self-healing log records are as follows , used to optimize the model.

[0055] Final output system status data The process of obtaining it is as follows: Equipment operating status data The fault handling results are generated by collecting real-time data such as voltage, network signal strength, and device self-test status codes from each device, combined with the judgment results of device status during anomaly monitoring. The update is determined based on the confirmation signal returned by the software configuration update and the feedback after the hardware failure work order is processed; if it fails, the reason is recorded; digital twin scenario update information. The system generates updates on vehicle positions and statuses in virtual scenes based on high-precision sensing data, as well as mapping changes in equipment operating status within the digital twin model, ensuring real-time synchronization with the actual community situation.

[0056] For example, the rated voltage of the camera at the entrance and exit of a community parking lot. Average daily usage This (belongs to high-frequency equipment) Calculate the dynamic adjustment coefficient. Thus, the lower voltage threshold is obtained. Voltage upper limit threshold .

[0057] At 9:00 AM one day, the system detected that the camera's voltage was 11.2V, 11.1V, and 11.0V in three consecutive cycles (each cycle lasting 10 seconds), all below the acceptable level. This triggers an anomaly alarm. The rule engine initially diagnoses that no matching rule was found, so the system calls a deep neural network model to extract feature vectors. (in , It means 9 a.m. on Wednesday. It was a sunny day. Light intensity, , The input model outputs three candidate results: "Power module aging (probability 65%)", "Voltage regulator failure (probability 20%)", and "Poor line contact (probability 10%)". After expert rule verification, the "Aging Probability Increase" rule was triggered due to the equipment's 450-day usage time, ultimately increasing the confidence level of "Power module aging" to 85% and identifying it as the final cause of the failure.

[0058] The system generates a hardware fault work order and calculates the work order priority: This fault affects the core business module of vehicle recognition at the entrance / exit (related to 1 core module). This is a level two fault. The average repair time for similar historical faults is 2 hours. ),but The system pushed the work order to the maintenance personnel. After the maintenance personnel replaced the power module, the voltage was restored to 12.1V, which took 1.5 hours. The self-healing was successful and the log was recorded.

[0059] During the federated learning optimization phase, the community will analyze 100 "power module failure" cases (feature vectors containing...) The device type (e.g., camera, voltage at fault: 11.0V) is incorporated into the local training to calculate the gradient parameters for "replacing module A". (Based on device type, voltage, and time period characteristics), form a group with another community subsystem to generate paired masks. , Then upload the encrypted gradient. , The aggregation server collects encrypted gradients from five communities, decrypts them, and obtains the average gradient. Update global model parameters. After rollout, when other communities encounter the same fault, the system prioritizes recommending module A, reducing the average repair time from 2 hours to 1 hour and increasing the fault handling success rate from 75% to 90%. System status data. middle, The camera's current voltage is recorded as 12.1V, signal strength as 52dBm, and no fault codes are found. Mark as successful; Update the camera icon in the digital twin scene to the normal state and synchronize vehicle recognition data to the virtual scene.

[0060] In the community smart parking management system, hierarchical behavior management decision-making achieves dynamic control of vehicle behavior and continuous improvement of system performance by constructing a closed-loop system. Based on high-precision perception data, system status data, and historical vehicle violation records, violation risks are quantified through regional hierarchical classification and game theory models to generate differentiated management strategies. The decision-making model is then iteratively optimized using reinforcement learning, ultimately outputting management command data and forming a closed-loop feedback system.

[0061] The input data contains three core elements: high-precision sensing data and real-time acquisition of vehicle acceleration. (unit (Reflects the degree of acceleration and deceleration of the vehicle) and driving speed V (unit: V). Records vehicle real-time movement speed and latitude / longitude coordinates. (Location of vehicle in space), timestamp (Marking data acquisition time) and multi-sensor fusion results (e.g., data confidence level, valued between 0 and 1, to assess the reliability of the perceived data), updated 20 times per second to ensure trajectory capture accuracy; system status data integrates equipment operating parameters (camera voltage). Network signal strength ), Fault handling results (Marking successful / failed fault handling and the reason) and digital twin scenario update information (Synchronize vehicle location and equipment status changes); Store vehicle historical violation records for the past 6 months, including violation type codes. (1 = illegal parking, 2 = speeding, 3 = failure to pay toll), violation location code (1 = Core Area, 2 = Secondary Core Area, 3 = Ordinary Area) Processing Results and cumulative number of violations (Statistics on vehicle violation frequency).

[0062] The process is divided into four stages: First, regional classification and characteristic quantification are carried out, dividing the community into core areas (entrances / main roads, high pedestrian and vehicle traffic), secondary core areas (secondary roads, medium traffic), and ordinary areas (parking spaces, low traffic) using electronic fences, and assigning risk coefficients to each area. , , Standardized processing of vehicle behavior characteristics, speed characteristics ( This is a regional speed limit, such as 20 km / h in the core area; exceeding this will result in... Take 1) Acceleration characteristics (5 is the critical acceleration threshold; exceeding this threshold indicates rapid acceleration / deceleration.) Historical violation probability. (Adding 30 to the denominator creates a smoothing term to avoid excessive probability fluctuations due to a small number of violations).

[0063] Secondly, a game theory-based risk assessment model is constructed. Using a two-player non-zero-sum game framework, the benefits and costs of strategy choices for both the car owner and the management system are quantified to solve for the risk of violations in equilibrium. Among the game participants, the car owner's strategy set... ( "Violations", such as speeding and illegal parking; (For "compliance behavior"), management system policy set ( To "enhance monitoring," high-frequency sampling and real-time interception are implemented. (For "routine monitoring," standard sampling and post-event processing are performed.) Define the payoff function for both parties: the payoff for a vehicle owner who violates regulations but is not caught. ( To save time for violations (unit: minutes; v is the time value coefficient, taken as 2 yuan / minute), and the losses incurred if violations are detected. ( Basic fine; This is a regional coefficient, based on the code of the violation location. Confirmed: Core area 1.5, secondary core area 1.2, general area 1.0), compliant behavior benefits. The management system successfully prevented illegal gains under enhanced monitoring. ( The social cost of a single violation is set at 500 yuan (including traffic congestion, safety risks, etc.) and the losses incurred when no violation is detected under routine monitoring. ( (This represents the probability of missed detection; 0.3 for routine monitoring, 0.05 for enhanced monitoring). The additional cost of enhanced monitoring... (20 yuan / hour in the core area, 10 yuan / hour in other areas), routine monitoring has no additional cost. The final payoff matrix is ​​expressed as follows: ; Let the probability of the car owner choosing to violate the rules be... The probability of the management system selecting enhanced monitoring is By solving the equilibrium condition where the partial derivatives of the expected payoffs of both parties are zero, the equilibrium violation probability is obtained. (Core parameter, reflecting the likelihood of violation under rational decision-making) and equilibrium monitoring probability .Will By fusing real-time vehicle features, the final risk score is: .

[0064] The third stage generates dynamic management strategies: when (High risk, such as a vehicle speeding in the core area with a high probability of violation) triggers an emergency interception signal at the barrier gate, simultaneously generating a warning notification containing the time and location of the violation, along with screenshots of evidence (from camera sensing data); when (Medium risk, such as a vehicle showing a tendency to violate regulations but not actually violating them) When the barrier gate allows passage normally but adds the vehicle to the key monitoring list (sampling frequency increased from 20Hz to 50Hz), a driving safety reminder is sent (e.g., "You have entered the main road, current speed limit 20km / h"); when (In cases of low risk and compliant vehicle behavior) routine management is maintained, with only behavioral data recorded in the database.

[0065] In the reinforcement learning-based decision-making model optimization phase, the system constructs a complete "state perception-policy decision-reward feedback" mechanism, deeply integrating vehicle behavior characteristics with historical violation data to achieve dynamic iteration of management strategies. The state space S contains four core features, and standardization ensures comparability between features. Among real-time behavior features, standardized speed... Standardized acceleration reflects whether the vehicle is currently speeding and the extent of the speeding. To detect whether a vehicle has a tendency to drive dangerously; regarding historical violation characteristics, the probability of historical violations. This reflects the vehicle's habitual violations and its adaptability to regional violations. Quantify the vehicle's tendency to violate regulations in the current area; additionally, it includes the area type k and system health. This is used to reflect the reliability of the monitoring system.

[0066] Action space A is defined as three types of management strategies, which are clearly mapped to state characteristics. High-risk strategies ( Triggering the barrier gate to intercept, generating a warning notification containing evidence of violation, execution cost Medium-risk strategy ( The barrier gate allows passage normally, but including vehicles in the key monitoring system and sending out standardized reminders incurs implementation costs. Low-risk strategy ( Maintaining routine management without sending notifications incurs execution costs. .

[0067] The reward function is designed as a multi-objective optimization function, balancing the effectiveness of violation governance with system resource consumption. It corrects the standardization problem of stability rewards, ensuring that each indicator contributes weight within a unified dimension. The formula is as follows: Among them, the effect reward Cost penalty Strategy execution cost, stability reward (because It has been normalized. ), weighting coefficient , , The determination is made through grid search optimization. When the system detects an emergency scenario (such as a traffic jam in the core area with more than 5 vehicles), it will temporarily adjust the settings. Increased to 0.8 The weight is reduced to 0.05 to prioritize the effectiveness of violation management; during low-traffic periods, the original weight is restored to balance stability and resource consumption. The model employs the Deep Deterministic Policy Gradient (DDPG) algorithm, using an Actor-Critic architecture to optimize the policy. The Actor network outputs a policy selection probability distribution based on the current state S. Critic network evaluates state-action value. Through timing difference error Update parameters: After processing 500 decision data points, update the Actor network parameters using gradient ascent. This ensures that the parameters converge stably.

[0068] Output management command data Generated through three-level parsing. Barrier gate control signal. Direct mapping based on strategy type, with additional execution timing parameters; vehicle owner notification content. Dynamically populate feature data using natural language templates; optimize parameter update values. Record the adjustment amount of the feature weights in this iteration. These instruction data form a closed loop through a specific path: Feedback is sent to the multi-source data fusion module to adjust sensor weights; and The execution results are captured by high-precision perceptual data and used as input for the next round of reinforcement learning; the digital twin module is based on Optimize behavior prediction models in virtual scenarios to synchronize virtual decision-making with actual management results.

[0069] For example, vehicle C is in the core area ( The historical data is as follows: (All were speeding) , (All historical violations occurred in the core area). Current status: (Actual speed 16km / h, speed limit 20km / h) 0.2 0.95 (System status is good). Initial model output strategy. (Medium risk) Within 30 minutes of implementation, the violation rate in the area decreased by 12%. Receive rewards After 500 iterations, the model found... and The combination of [various factors] is more critical for violation prediction, so the weights of the two are increased from 0.2 and 0.3 to 0.25 and 0.35 respectively; when vehicle C re-enters the core area in the same state, the risk score calculated by the new model is increased from 72 (medium risk) to 81 (high risk), triggering [a certain event / condition]. The strategy (barrier interception + warning notification) reduced the vehicle's subsequent violation rate in the core area by 65%. Through this mechanism, the learning model continuously learns the correlation between features and strategies, upgrading management decisions from "experience-driven" to "data-driven." The accuracy rate of core area violation handling gradually increased from the initial 68% to 94%, and the average response time stabilized within 1.2 seconds.

[0070] The feedback optimization closed loop is achieved through a three-level linkage: the multi-source data fusion module receives... Then, the sensor weight matrix W is adjusted (e.g., if reinforcement learning discovers that a geomagnetic sensor in a certain area misjudged, leading to a false alarm of violation, its weight is reduced from 0.4 to 0.2, while the weight of the camera data is increased from 0.6 to 0.8); the digital twin module is based on Optimize fault diagnosis thresholds (e.g., dynamically adjust the abnormal current threshold of the barrier gate motor from 15A to 16A during high-traffic periods to reduce false alarms caused by instantaneous current fluctuations); high-precision perception data continuously provides a basis for verifying the effectiveness of the strategy (e.g., compare the changes in vehicle violations after implementing the new strategy; if the speeding violation rate in the core area drops from 15% to 8%, then retain the current model parameters; otherwise, further adjust them).

[0071] The processes described above with reference to the flowcharts in the embodiments disclosed in this invention can be implemented as computer software programs. The embodiments disclosed in this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by a central processing unit (CPU), it performs the functions defined in the methods of this application. It should be noted that the computer-readable medium described above in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wire segments, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to: wireless segments, wire segments, optical fibers, RF, etc., or any suitable combination thereof.

[0072] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0073] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the present invention. The purpose of the present invention has been fully and effectively achieved. The functions and structural principles of the present invention have been shown and explained in the embodiments. Without departing from the stated principles, the implementation of the present invention may have any variations or modifications.

Claims

1. A method for dynamic monitoring and analysis of vehicles in a smart community, characterized in that, The method includes: Multi-source data is collected through sensors, and structured perception data is generated using a lightweight OCR algorithm; Based on structured sensing data, combined with environmental parameters and equipment status, an adaptive dynamic weighting algorithm is used to calculate the real-time weights of each sensor. The fusion confidence is calculated using a feature confidence fusion formula. Unqualified data is then removed based on the fusion confidence to generate high-precision sensing data. A community digital twin scenario is built based on high-precision sensing data. Anomaly detection is performed through a dynamic threshold algorithm. When an anomaly is detected, a three-level mechanism is used to locate the fault and generate a configuration update instruction. System status data is generated based on the fault handling results, digital twin scenario update information, and equipment operation status data. High-precision sensing data, system status data, and vehicle historical violation records are input into a game theory model to calculate the probability of vehicle violations, and differentiated management strategies are generated based on the probability of vehicle violations. The specific process for anomaly detection is as follows: Based on the equipment's rated parameters, operating frequency, and adjustment coefficient, calculate the dynamic upper and lower limit thresholds for equipment voltage and network signal strength. Adjust the network signal threshold based on the average daily number of transmission failures. Monitor equipment voltage and network signal strength in real time, recording whether they exceed the threshold, the duration of the exceedance, and the fluctuation amplitude within the threshold. If the parameter exceeds the threshold for a specific consecutive period, it is judged as abnormal. If it exceeds the threshold for a single period but does not reach a continuous period, record the fluctuation and increase the monitoring frequency. If the fluctuation amplitude within the threshold exceeds the rated value proportion, mark it as unstable and trigger self-calibration. The specific process of obtaining the configuration update command is as follows: The system calls the preset IF-THEN rule engine to initially screen for software faults. If no match is found, the system outputs the fault probability distribution through a deep neural network model and cross-validates it with the expert rule base to confirm the software configuration problem to which the fault belongs. Key information is extracted from the fault diagnosis data to identify the names of the parameters that need to be adjusted, and the target values ​​of the parameters are determined based on the cause of the fault and historical repair cases. The execution time is determined by combining the real-time working status of the equipment and the scope of the fault impact. If the faulty equipment is in a low-load period, it is set to execute immediately; if the equipment is processing core business, it is delayed until the off-peak period to avoid interrupting normal service due to configuration updates. The specific calculation process for the probability of vehicle violations is as follows: Clearly define the strategies of vehicle owners and the management system, and define the benefits and costs of both parties under different strategy combinations; construct a non-zero sum benefit matrix for both vehicle owners and the system based on regional coefficients, missed detection probability, and additional monitoring costs, and quantify the benefit values ​​of both parties under different strategies; set the vehicle owner violation probability and the system enhanced monitoring probability as variables, and solve for the equilibrium violation probability under the Nash equilibrium state by calculating the expected benefits of both parties and setting their partial derivatives to 0. The specific process of the differentiated management strategy is as follows: By combining regional risk coefficients, equilibrium violation probabilities solved by game theory, and real-time standardized speed and acceleration characteristics of vehicles, a weighted calculation is performed to obtain a vehicle violation risk score. If a vehicle's violation risk score is greater than or equal to the high-risk threshold, the gate will be triggered for emergency interception, and a warning notification containing evidence of the violation will be generated simultaneously. If the score is in the medium-risk range, the gate will allow the vehicle to pass normally, but the vehicle will be included in key monitoring, and a prompt for proper driving will be sent. If the score is less than the low-risk threshold, regular management will be maintained, and only vehicle behavior data will be recorded. By using a reinforcement learning model, the strategy selection logic is iteratively optimized by using a decrease in violation rate as an effect reward, strategy execution cost as a penalty, and speed fluctuation as a stability reward, and the reward weight is temporarily adjusted according to the real-time scenario.

2. The method for dynamic monitoring and analysis of vehicles in a smart community according to claim 1, characterized in that, The specific process for generating the structured sensing data is as follows: Vehicle image data is collected by a camera, distance data within a specific range is collected by an ultrasonic sensor, and trigger status is output in real time by a geomagnetic sensor. The collected data is preprocessed. The lightweight OCR algorithm of the edge computing box extracts features from the preprocessed image, generates feature maps of different resolutions, presets anchor frames corresponding to the license plate size, predicts the license plate frame coordinates and outputs the initial confidence score, and optimizes the confidence score through time consistency verification. After cropping the valid license plate frame that meets the confidence level, locate the character area, input it into the classification model to identify the character and calculate the overall character recognition confidence level; extract the vehicle outline through the instance segmentation model, calculate the aspect ratio and combine it with the threshold to map the vehicle model code; preliminarily determine the presence of the vehicle based on the geomagnetic trigger state and the calibrated distance, calculate the positional deviation between the center of the license plate frame and the ultrasonic detection center, and remove data with deviations exceeding the threshold.

3. The method for dynamic monitoring and analysis of vehicles in a smart community according to claim 2, characterized in that, The process of obtaining the high-precision sensing data is as follows: Based on the historical failure rate of the equipment and environmental parameters, the real-time trust weights of the camera, ultrasonic sensor and geomagnetic sensor are calculated separately to ensure that the total weight is a specific value and that the weight of each item is not lower than the minimum threshold. Any excess weight is deducted from the other weights proportionally. The weights are converted according to the different needs of coordinate fusion and confidence fusion. Coordinate fusion only uses the weights of the camera and ultrasonic sensor and normalizes them, while confidence fusion retains the original total weight system. The final coordinates of the vehicle are calculated using a weighted average. Based on the weight of each sensor, the fusion confidence score is calculated by combining the product of the camera's positioning and character recognition confidence scores, the deviation ratio between the ultrasonic distance and the predicted distance, and the geomagnetic triggering state. Only data that meet the fusion confidence score criteria are retained.

4. A dynamic monitoring and analysis system for vehicles in a smart community, characterized in that, The system is used to execute the dynamic monitoring and analysis method for smart community vehicles as described in any one of claims 1-3.

5. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which is executed by a processor to implement the dynamic monitoring and analysis method for smart community vehicles as described in any one of claims 1-3.

Images