Urban road moving source intelligent monitoring method and system based on multi-source data coupling

Abstract

The invention discloses an urban road mobile source intelligent monitoring method and system based on multi-source data coupling, and relates to the technical field of environment monitoring and intelligent traffic. Utilizing a machine learning algorithm to construct a pollutant emission, carbon emission and energy consumption prediction model; acquiring vehicle inventory data in the city, and calculating the pollutant emission, energy consumption and carbon emission of the whole city; the method comprises the following steps of: constructing a dynamic distribution diagram of automobile emission in a city by using real-time position data of vehicles and urban road network information, introducing an atmospheric diffusion model, simulating pollutant migration by combining urban geographic information and environmental data, comparing and verifying a simulated migration result with actually measured data of a national control site, and optimizing parameters of the atmospheric diffusion model; and storing the result into a real-time database, and displaying the pollutant distribution, emission, energy consumption and carbon emission of the urban road network in real time through a visual interface. According to the invention, the temporal-spatial resolution and prediction precision of data are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of environmental monitoring and intelligent transportation technology, in particular to a method and system for intelligent monitoring of mobile sources on urban roads based on multi-source data coupling. Background Art

[0002] With the acceleration of urbanization and the rapid growth of motor vehicle ownership, mobile source pollutant emissions have become a significant source of urban air pollution. Traditional pollutant emission monitoring methods, which primarily rely on fixed monitoring stations and regular emission testing, struggle to dynamically reflect the spatiotemporal distribution of urban mobile source pollutant emissions in real time, and are unable to accurately reflect the actual impact of urban traffic operating conditions on pollutant emissions. Traditional emission factor models are often based on average operating condition assumptions, making it difficult to accurately reflect the impact of actual vehicle operating conditions on pollutant emissions. Furthermore, existing research focuses primarily on the total amount of pollutant emissions and lacks dynamic simulation of the spatiotemporal distribution and migration trends of pollutants.

[0003] These issues make it difficult to make scientific and precise decisions about urban mobile source pollution control. Existing technologies are particularly limited in their ability to rapidly respond to and accurately predict pollutant distribution during unexpected traffic incidents or under unusual weather conditions. Therefore, a technical solution is urgently needed that can acquire real-time vehicle data, combine urban geographic information with weather conditions, and achieve real-time prediction and migration simulation of pollutant emissions. Summary of the Invention

[0004] In order to overcome the above problems existing in the prior art, the present invention proposes an urban road mobile source intelligent monitoring method and system based on multi-source data coupling.

[0005] The technical solution adopted by the present invention to solve the technical problem is: an intelligent monitoring method for mobile sources on urban roads based on multi-source data coupling, comprising the following steps: Step 1: Collect urban road mobile source data, establish a vehicle emission characteristics database, and process the database data; Step 2: Based on the feature database obtained in step 1, a machine learning algorithm is used to build a pollutant emission, carbon emission and energy consumption prediction model to predict the vehicle's real-time emissions, energy consumption and carbon emissions; Step 3: Obtain vehicle ownership data within the city, expand the per-vehicle emission prediction results obtained in Step 2 to the city scale, and calculate the city's overall pollutant emissions, energy consumption, and carbon emissions; Step 4: Use real-time vehicle location data and urban road network information to construct a dynamic distribution map of automobile emissions in the city. Introduce an atmospheric diffusion model, combine urban geographic information and environmental data to simulate pollutant migration, and compare and verify the simulated migration results with measured data from national monitoring stations to optimize the atmospheric diffusion model parameters. Step 5: Store the results obtained in steps 1-4 into a real-time database and display the pollutant distribution, emissions, energy consumption, and carbon emissions of the urban road network in real time through a visual interface.

[0006] In the above-mentioned intelligent monitoring method for mobile sources on urban roads based on multi-source data coupling, the data collection of mobile sources on urban roads in step 1 specifically includes vehicle identification information data collection, vehicle diagnostic system parameter collection, and annual inspection data collection.

[0007] In the above-mentioned intelligent monitoring method for mobile sources on urban roads based on multi-source data coupling, the pollutant emission, carbon emission and energy consumption prediction model in step 2 uses vehicle identification information data collection and on-board diagnostic system parameter collection as model input, and uses the vehicle's real-time emissions, energy consumption and carbon emissions as output, and is optimized using historical annual inspection data; an online learning framework for the emission prediction model is constructed, and the model parameters are dynamically corrected using a transfer learning algorithm through deviation analysis between the actual annual inspection data regularly updated by the vehicle management office and the prediction results.

[0008] In the above-mentioned intelligent monitoring method for mobile sources on urban roads based on multi-source data coupling, in step 1, the time series database InfluxDB is used for high-frequency, time-series real-time streaming data, taking advantage of its efficient time window aggregation and streaming processing capabilities; static structured data is managed through the relational database MySQL to ensure transaction consistency and support for complex queries; for spatial information data, spatial indexing and geographic operations are implemented based on the PostgreSQL extension component PostGIS.

[0009] In the above-mentioned intelligent monitoring method for mobile sources on urban roads based on multi-source data coupling, the pollutant emission, carbon emission and energy consumption prediction model architecture constructed in step 2 is specifically as follows: a convolutional neural network is used to analyze the spatial characteristics of the road network topology and the layout of surrounding buildings, a long short-term memory network is combined to capture the long-term dependency of time series signals, an attention mechanism is introduced to dynamically adjust the weights of multi-source features, and the coordinated optimization of spatial-temporal-environmental features is achieved. The information flow is controlled through a gating mechanism to avoid a single feature dominating the prediction results.

[0010] In the above-mentioned intelligent monitoring method for mobile sources on urban roads based on multi-source data coupling, the process of simulating pollutant migration in step 4 specifically includes: using a multi-scale coupling method to realize dynamic simulation of pollutant transmission, establishing a three-dimensional simplified model based on computational fluid dynamics, and calculating the turbulence effect and bypass path of pollutants in short distance diffusion by discretizing boundary conditions and combining real-time wind speed and direction and temperature stratification data; at the same time, for large-scale regional pollution, integrating a Gaussian smoke plume diffusion model or an AERMOD system, using the pollutant source intensity output by the emission prediction model as input, and generating a concentration spatiotemporal distribution matrix by parameterizing meteorological fields and terrain elevation data; The pollutant concentration field is spatially aligned with the road network and population density layers, and heat map rendering technology is used to display the pollution diffusion trend. The timeline can be dragged to playback the historical 24-hour concentration evolution process, and a threshold warning function is embedded. The spatial interpolation algorithm is combined to fill the monitoring blind spot data, while providing multi-layer overlay comparison and supporting interactive parameter adjustment.

[0011] An intelligent monitoring system for mobile sources on urban roads based on multi-source data coupling specifically includes a multi-source data acquisition and standardization module, a data preprocessing and storage module, an emission prediction and dynamic correction module, a city-level pollutant migration simulation module, a closed-loop verification and quality control module, and a visualization and decision support module. The multi-source data acquisition and standardization module includes a vehicle static data acquisition submodule, a dynamic OBD data real-time access submodule, an annual inspection and historical data integration submodule, and an environmental and geographic data acquisition submodule; the data preprocessing and storage module includes a data cleaning and feature engineering submodule, and a hierarchical storage and data lake architecture submodule; the emission prediction and dynamic correction module includes a triple coupling modeling engine submodule for constructing a storage prediction model, and a model online correction submodule focusing on long-term prediction stability; the city-level pollutant migration simulation module includes a diffusion dynamics calculation submodule for pollutant transmission simulation, and a pollution tracing and early warning submodule for realizing emission hotspot identification and risk prediction; the closed-loop verification and quality control module includes a national control site data comparison submodule and a safety and compliance audit submodule; the visualization and decision support module includes a real-time three-dimensional visualization submodule and a decision support analysis submodule.

[0012] In the above-mentioned intelligent monitoring system for mobile sources on urban roads based on multi-source data coupling, the data cleaning and feature engineering submodule implements data quality control through a multi-level rule engine, dynamically loads preset parameter thresholds based on vehicle models, and uses sliding windows for dynamic interpolation to fill signal loss; high-order indicators are constructed through feature derivation algorithms, and finally a standardized feature matrix is ​​output for model training; the hierarchical storage and data lake architecture submodule adopts a three-level storage strategy, and real-time data is stored in the original OBD stream through Kafka Topic partitions and synchronously written to Apache Parquet columnar files; high-frequency access data is stored in the ClickHouse cluster, and its MPP architecture and columnar compression are used to achieve sub-second response; low-frequency historical data is archived to HDFS, and cold backup access is provided through the MinIO object storage interface.

[0013] The present invention offers significant advantages in the field of real-time online monitoring of intelligent transportation and carbon emissions. First, through the coordinated application of multi-source data (including vehicle identification numbers (VINs), real-time operating parameters from on-board diagnostic systems (OBDs), and static information from annual inspections), it enables dynamic and comprehensive monitoring and analysis of urban mobile pollution sources. This approach, based on the fusion of three data sources, effectively overcomes the incomplete data coverage and inaccuracy inherent in existing technologies resulting from a single data source.

[0014] Secondly, in terms of data analysis and processing capabilities, the system incorporates machine learning algorithms to model and predict multi-dimensional data. Compared to traditional methods that rely solely on passive monitoring at fixed monitoring stations or simple OBD data analysis, this significantly improves the accuracy and real-time nature of emissions predictions. The system dynamically captures changes in vehicle operating characteristics and accurately calculates each vehicle's carbon emissions through intelligent prediction models, providing more precise data support for urban air quality management and traffic optimization.

[0015] Furthermore, in terms of monitoring methods and application effectiveness, a visual monitoring platform has been established. By dynamically displaying the distribution of pollutants within the urban area, it enables the rapid location and source tracking of vehicles with abnormal emissions. This capability is currently unavailable in existing technologies and significantly improves the efficiency of pollution source supervision. Furthermore, the system can provide real-time optimization suggestions to traffic management departments (such as adjusting traffic light timing and restricting high-emission vehicles), forming a closed-loop management mechanism of "monitoring-analysis-feedback."

[0016] Overall, this invention, through the innovative combination of multi-source data collaboration, intelligent modeling, and real-time monitoring technology, has achieved breakthrough progress in improving the timeliness, accuracy, and intelligence of urban mobile pollution source supervision. Its combined effectiveness significantly outperforms existing technologies, providing strong technical support for precise pollution control and intelligent traffic management. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic flow chart of the monitoring method of the present invention; Figure 2 It is a schematic diagram of the module composition of the monitoring system of the present invention. DETAILED DESCRIPTION

[0018] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1 As shown, this embodiment discloses an intelligent monitoring method for mobile sources on urban roads based on multi-source data coupling, comprising the following steps: Step 1: Collect urban road mobile source data, establish a vehicle emission characteristics database, and process the database data.

[0020] First, access vehicle VIN information through the vehicle administration office database interface, including key registration information such as brand, engine model, fuel type, or electric parameters, for vehicle identity verification and historical data analysis. Second, deploy on-board terminal equipment (such as OBD-II devices that support protocols such as SAE J1939 and ISO 15765) or collaborate with automakers to utilize 4G / 5G networks to transmit dynamic parameters such as engine speed and battery current and voltage in real time. Edge computing technology is also used to optimize data real-time performance and storage efficiency. Finally, connect to the vehicle administration office's annual inspection system API to regularly obtain mileage and emissions test results. By comparing OBD-reported data, the risk of vehicle tampering can be verified, and emission monitoring models can be calibrated to ensure data authenticity and compliance.

[0021] Environmental and geographic data integration covers both meteorological and spatial information. On the one hand, high-precision real-time meteorological data such as temperature, humidity, wind speed, and direction are accessed through the Meteorological Bureau's API, and combined with vehicle emission parameters to construct pollution diffusion models, such as analyzing the impact of low temperatures on NOx emissions. On the other hand, GIS platforms (such as ArcGIS) are used to obtain road network topology (number of lanes, speed limits, slopes, etc.) and three-dimensional building models (including height and facade details), supporting traffic route optimization and pollution heat map generation. For example, low-emission driving routes can be planned using road network data, or vehicle trajectories can be overlaid with air quality data to identify urban pollution hotspots, providing data support for intelligent transportation and urban planning.

[0022] The data of national monitoring stations are accessed in real time through the standardized interface opened by the environmental protection department to obtain PM 2.5 , NOx, and other air quality monitoring data, combined with vehicle and environmental data for in-depth correlation analysis. Using spatiotemporal alignment technology, the GPS coordinates of vehicles are matched with monitoring values ​​from nearby nationally monitored stations to construct a comprehensive "vehicle-environment" dataset, which is used to assess the contribution of individual vehicle emissions to regional air quality. Furthermore, combined with real-time pollutant concentration changes, OBD emissions data from vehicles passing through a specific time period is backtracked to quickly locate abnormal pollution sources (such as high-emission vehicles or congested roads), providing a precise basis for decision-making in environmental supervision and pollution control.

[0023] The data processing process includes data cleaning and fusion, as well as database design. During this stage, we first deeply process the raw data through the development of customized ETL (extraction, transformation, loading) tools, focusing on addressing anomalies such as OBD signal loss and sensor noise. For example, we use interpolation to fill missing data and filtering algorithms to smooth noisy data, ensuring the reliability of subsequent analysis. Furthermore, we construct a unified indexing system based on spatiotemporal characteristics for multi-source heterogeneous data (such as vehicle trajectories, ambient temperature and humidity, and road network topology). This system associates timestamps and geographic locations (such as GPS coordinates) with road network nodes, enabling dynamic matching and efficient querying across data dimensions. For example, we can precisely map sudden braking events to specific road slopes or congested sections, supporting multi-factor causal analysis.

[0024] In database design, a tiered storage architecture was adopted based on data type and access requirements. For high-frequency, time-sensitive real-time OBD streaming data (such as vehicle speed and engine speed), the time-series database InfluxDB was used, leveraging its efficient time window aggregation and stream processing capabilities. Static structured data, such as vehicle records and annual inspection records, was managed using the relational database MySQL, ensuring transaction consistency and supporting complex queries. For spatial information such as road network vector data and geofences, spatial indexing and geographic operations (such as path planning and regional collision detection) were implemented using the PostgreSQL extension component PostGIS. This architecture balances real-time performance, stability, and spatial analysis capabilities, providing multi-dimensional data support for upper-level applications.

[0025] Step 2: Based on the feature database obtained in step 1, a machine learning algorithm is used to build a pollutant emission, carbon emission and energy consumption prediction model to predict the vehicle's real-time emissions, energy consumption and carbon emissions.

[0026] In step 2, the pollutant emission, carbon emission and energy consumption prediction model uses vehicle identification information data collection and on-board diagnostic system parameter collection as model input, and real-time vehicle emissions, energy consumption and carbon emissions as output, and is optimized using historical annual inspection data; an online learning framework for the emission prediction model is constructed, and the model parameters are dynamically corrected using a transfer learning algorithm through deviation analysis between the actual annual inspection data regularly updated by the vehicle management office and the prediction results.

[0027] Predictive model building includes feature engineering, model architecture, and model training and optimization.

[0028] During the feature engineering stage, the system constructed a multi-dimensional feature system to accurately characterize the factors affecting emissions: static features cover the inherent attributes of the vehicle (such as brand, engine displacement, battery capacity, etc.), dynamic features extract real-time operating parameters (engine load timing fluctuations, battery charge and discharge power curves, transient changes in vehicle speed), and environmental features integrate external conditions (road slope based on GIS calculation of the vehicle's real-time position, pollutant diffusion coefficient driven by wind speed sensor data). Dimensional differences are eliminated through normalization and standardization, and principal component analysis (PCA) is used to reduce the redundancy of high-dimensional features to form an input vector with dynamic spatiotemporal linkage.

[0029] During the model architecture phase, a convolutional neural network (CNN) is first used to analyze the spatial characteristics of the road network topology and surrounding building layout (e.g., the cumulative effect of frequent start-stop traffic at an intersection on emissions). A long short-term memory (LSTM) network is then incorporated to capture the long-term dependencies of temporal signals such as engine operating conditions and battery health (e.g., the gradual change in engine load during continuous hill climbing). Finally, an attention mechanism is introduced to dynamically adjust the weights of multiple source features (e.g., automatically increasing the decision weight of battery thermal management efficiency features in high-temperature environments), achieving coordinated optimization of spatial, temporal, and environmental features. This hybrid model uses a gating mechanism to control information flow, preventing a single feature from dominating the prediction results.

[0030] Model training and optimization are conducted within a distributed computing framework, using PySpark or Flink for parallel preprocessing and feature concatenation of massive amounts of vehicle data. A parameter server architecture is employed to accelerate gradient updates. A transfer learning mechanism is introduced, using historical annual inspection emissions data as a pretraining benchmark. The model is fine-tuned in real time through online learning, integrating the latest OBD data (e.g., dynamic parameter updates based on Kalman filtering). Adversarial training is also incorporated to enhance generalization capabilities under extreme driving conditions. Model validation utilizes a sliding window strategy, using MAPE (mean absolute percentage error) and RMSE as core metrics. A Bayesian optimization algorithm is integrated for automatic parameter adjustment to ensure robust prediction results in dynamic traffic scenarios.

[0031] Step 3: Obtain data on the number of vehicles in the city, expand the emission prediction results of individual vehicles obtained in Step 2 to the city scale, and calculate the overall pollutant emissions, energy consumption, and carbon emissions of the city.

[0032] Step 4: Use real-time vehicle location data and urban road network information to construct a dynamic distribution map of automobile emissions in the city. Introduce an atmospheric diffusion model, combine urban geographic information and environmental data to simulate pollutant migration, and compare and verify the simulated migration results with the actual measured data from national monitoring stations to optimize the atmospheric diffusion model parameters.

[0033] Pollutant migration simulation includes diffusion model construction and visualization overlay.

[0034] A multi-scale coupling approach is used to simulate the dynamic transport of pollutants. First, a simplified three-dimensional model is established based on computational fluid dynamics (CFD) (e.g., the RANS equations combined with the k-ε turbulence model). By discretizing boundary conditions such as urban building layout and surface roughness, and integrating real-time wind speed, direction, and temperature stratification data, the turbulent effects and flow paths of pollutants over short distances (e.g., eddy retention of NOx within street canyons) are calculated. Furthermore, for large-scale regional pollution, a Gaussian plume dispersion model or the AERMOD system is integrated. Pollutant source intensity output from emission prediction models (e.g., vehicle exhaust CO2 and PM2.5 emission rates) is used as input. By parameterizing meteorological fields and terrain elevation data, a concentration spatiotemporal distribution matrix with a resolution of 100 meters is rapidly generated, enabling cross-scale concentration field prediction from microscopic street scenes to macroscopic urban agglomerations.

[0035] Use GIS engines (such as ArcGIS or open source QGIS) to spatially align pollutant concentration fields with road networks and population density layers, adopt heat map rendering technology (such as dynamic shading based on WebGL) to intuitively display pollution diffusion trends, support timeline dragging to replay the historical 24-hour concentration evolution process, and embed threshold warning functions (such as automatic highlighting of areas exceeding the standard); combine spatial interpolation algorithms (such as Kriging method) to fill in monitoring blind spot data, and provide multi-layer overlay comparison (such as simultaneous display of predicted values ​​and actual values ​​measured by ground monitoring stations), support interactive parameter adjustment (such as real-time re-rendering of diffusion paths after modifying wind speed assumptions), and provide a high-precision visualization foundation for pollution source tracing and traffic control strategy verification.

[0036] To ensure the model's prediction accuracy and reliability in practical applications, a closed-loop verification mechanism was established. Through designed comparative experiments, the model's predicted pollutant emissions (such as CO2 and NOx concentrations) were spatiotemporally aligned with measured data from national environmental monitoring stations. Statistical metrics (such as R² and NRMSE) were used to quantify deviations. When the error exceeded a preset threshold (e.g., daily average error >15%), a calibration module was automatically triggered. Diffusion model parameters (such as turbulent diffusion coefficient and sedimentation rate) were dynamically adjusted using a Bayesian optimization algorithm. The weight allocation strategy for the emission prediction model was incrementally updated (e.g., retraining the attention layer weights) using an online learning mechanism. The verification process incorporated data quality verification (e.g., removing sensor outliers). Timestamp synchronization and spatial interpolation ensured the spatiotemporal consistency of the compared datasets, forming a closed loop of "prediction-monitoring-feedback-iteration."

[0037] The monitoring platform utilizes a layered architecture to achieve multi-dimensional pollution control. The front-end features a visual cockpit built with the Vue.js framework, integrating WebGL and the Mapbox engine to render 3D urban pollution heat maps (such as building-level PM2.5 concentration distribution). This includes dynamic timeline replay and multiple overlays (traffic flow and weather data). The back-end provides a RESTful API through Spring Cloud microservices, employing Kafka message queues to parallelize real-time OBD data streams (peak throughput >100,000 messages / second), and leverages Flink for streaming detection of emissions violations. The alarm module sets multiple thresholds (e.g., if regional CO concentration exceeds 80 ppm for 10 minutes), triggering SMS / email notifications and automatically generating emergency control recommendations (such as adjustments to restricted areas). The system pre-filters invalid data using edge computing nodes (such as the vehicle's ECU) and utilizes GPU acceleration to improve visualization rendering efficiency, ensuring end-to-end latency from data ingestion to decision-making and response is less than 500ms.

[0038] Step 5: Store the results obtained in steps 1-4 into a real-time database and display the pollutant distribution, emissions, energy consumption, and carbon emissions of the urban road network in real time through a visual interface.

[0039] This monitoring method was put into use. During the small-scale validation phase, pilot systems were prioritized for deployment in high-emission-sensitive areas, such as major traffic arteries and industrial zones. Edge computing nodes collected real-time vehicle On-Board Device (OBD) data, pollutant concentrations at roadside monitoring stations, and meteorological parameters. These data were then compared with model predictions at the minute level (e.g., NOx prediction error rate ≤ 12%). Simultaneously, A / B testing was conducted to compare the overlap between traditional monitoring methods and the system's output of road section emissions heat maps, verifying key metrics such as spatial resolution (e.g., 10-meter grid accuracy) and time response latency (<30 seconds). During the pilot period, a fault-fusing mechanism was established, and stress testing was conducted for abnormal scenarios such as roadside unit (RSU) communication interruptions and data packet loss to ensure the system's robustness in complex environments.

[0040] Through user surveys and behavior log analysis, the interactive design of the monitoring platform is optimized - for example, a speed adjustment slider for the pollutant migration animation, a drag-and-drop timeline zoom function, and the introduction of LOD (level of detail) technology to achieve smooth rendering under large-scale data; on the technical level, vehicle compatibility is expanded, and a dedicated emission factor model is developed for hydrogen fuel vehicles. Through transfer learning, the existing CNN-LSTM framework is adapted to hydrogen consumption prediction (such as fine-tuning the feature extraction layer based on proton exchange membrane fuel cell operating data), and new sensor data streams such as hydrogen storage tank pressure and stack efficiency are integrated to achieve unified management of multi-energy vehicle models.

[0041] During the city-wide deployment phase, the system's collaborative capabilities will be strengthened, with deep integration with the intelligent traffic signal control system. Real-time data on intersection queue lengths and signal phases will be obtained through APIs, and dynamic emission optimization strategies will be generated in combination with reinforcement learning models. For example, during peak congestion periods, detour suggestions will be pushed to the navigation platform based on real-time pollution concentrations and vehicle composition (such as the proportion of diesel vehicles) (reducing regional emissions by 30%), or traffic lights will be linked to extend the green wave band to reduce idling time. At the same time, a dynamic update mechanism for regional emission inventories will be established, and the road network-level emission data output by the system will be automatically imported into the urban air pollution source analysis model, providing environmental protection departments with a basis for traceability and control (such as accurately identifying the behavior of heavy trucks illegally passing through low-emission zones), forming a full-chain smart governance closed loop of "monitoring-prediction-control-evaluation".

[0042] In subsequent operation and maintenance, the high availability of the system is guaranteed through intelligent monitoring and elastic architecture: based on Prometheus+Grafana, the CPU / memory / disk health status of the server cluster is monitored in real time, and automatic scaling policies are set in combination with Kubernetes (such as automatically expanding computing nodes when CPU utilization exceeds 70%) to cope with sudden traffic peaks (such as a 200% surge in data access during peak hours in the morning and evening); full data lifecycle management is implemented, and the "3-2-1" backup principle (3 copies, 2 types of media, 1 copy off-site) is adopted to perform daily incremental backups of the time series database and spatial data warehouse, and a hierarchical data security emergency plan is formulated (such as switching to an isolated disaster recovery cluster within 30 minutes in the event of a ransomware attack). Disaster recovery drills are conducted every quarter to verify the key indicators of RTO (recovery time objective) < 2 hours and RPO (recovery point objective) < 15 minutes.

[0043] Continuous model learning relies on a data-driven closed loop to achieve dynamic evolution: a feedback loop of "annual inspection data collection → feature engineering reconstruction → online incremental training" is constructed. For example, the exhaust gas test results in the vehicle annual inspection are compared with the historical prediction values ​​of the model, and a labeled data set is generated in real time through Flink streaming processing, triggering PyTorch lightweight fine-tuning (freezing the underlying CNN convolution kernel and only updating the LSTM-Attention layer parameters), shortening the model iteration cycle from months to days; when releasing an updated version of the model every quarter, transfer learning technology is used to adapt to the emission characteristics of new models (such as hydrogen fuel cell vehicles), and the prediction accuracy of the new models is verified through A / B testing (MAPE≤8%). Based on the CI / CD pipeline, a rolling upgrade and version rollback mechanism of the model image is implemented to ensure service continuity.

[0044] Based on the above monitoring method, this embodiment also discloses an urban road mobile source intelligent monitoring system based on multi-source data coupling, such as Figure 2As shown, the monitoring system includes six modules: multi-source data acquisition and standardization module, data preprocessing and storage module, emission prediction and dynamic correction module, city-level pollutant migration simulation module, closed-loop verification and quality control module, and visualization and decision support module.

[0045] (1) The multi-source data acquisition and standardization module mainly includes four sub-modules: vehicle static data acquisition sub-module, dynamic OBD data real-time access sub-module, annual inspection and historical data integration sub-module, and environmental and geographic data acquisition sub-module.

[0046] M1. In the vehicle static data collection submodule, the system uses the VIN (Vehicle Identification Number) to accurately map vehicle identity and performance parameters. Connecting to the vehicle management office's core database or the automaker's open API (such as the Tesla Developer Platform), the system parses digits 4-9 of the VIN to obtain key attributes such as engine model and fuel type (gasoline / diesel / electric). Regular expressions are used to extract the exhaust after-treatment system model (such as the SCR catalyst version). For new energy vehicles, the system integrates model libraries from mainstream battery manufacturers such as Contemporary Amperex Technology (CATL), matching parameters such as high-voltage battery pack capacity (such as a 100kWh ternary lithium battery) and electric motor peak power (such as a 200kW permanent magnet synchronous motor) to create standardized parameter tables. During the data cleansing phase, a rules engine is used to verify logical inconsistencies (such as triggering an anomaly flag when labeling battery capacity for fuel vehicles), ensuring consistency between static data and physical entities.

[0047] M2. The dynamic OBD data real-time access submodule focuses on capturing the vehicle's real-time operating status. The onboard terminal integrates a 4G / 5G dual-mode communication module (Quectel EC20 series). Compatible with the SAE J1939 protocol (commercial vehicle CAN bus) and ISO 15031 protocol (passenger vehicle K-line communication) via the OBD-II interface, it decodes over 500 signal items in real time, including engine speed (RPM), air-fuel ratio (λ value), and battery cell voltage. Data transmission complies with the GB / T 32960.3 standard, encapsulating signal values, timestamps, and geolocation into lightweight JSON messages, which are uploaded to the cloud via the MQTT protocol at a 1Hz frequency. The cloud access layer utilizes the Flink stream processing engine for millisecond-level data parsing and performs outlier filtering (e.g., triggering alerts when engine speed exceeds the redline threshold) based on sliding windows (e.g., 10-second windows).

[0048] The M3. Annual inspection and historical data integration submodule builds a data chain for the entire vehicle lifecycle. It accesses structured annual inspection reports through the vehicle management office's dedicated network interface, analyzes mileage and exhaust gas test values ​​(such as CO concentration 1.2g / km and NOx limit compliance), and integrates them with 4S dealership maintenance records (such as DPF regeneration times and oxygen sensor replacement records). For unstructured PDF annual inspection reports, it uses OCR (Tesseract engine) and NLP (BERT entity recognition) to extract key fields and stores them in a Neo4j graph database to form a "vehicle-component-test event" association network. This module provides a baseline correction basis for emission prediction models, for example, by calibrating OBD mileage sensor drift errors based on historical mileage data.

[0049] M4. The environmental and geographic data acquisition submodule integrates multi-dimensional spatial data. Real-time wind speed, temperature, and humidity data with a 1km×1km resolution are acquired through the China Meteorological Administration's Smart Grid API. LoRa IoT micro-meteorological stations (with an accuracy of ±0.5m / s) are deployed in key areas to compensate for local deviations in macro-data. The geographic information processing end loads city CIM (City Information Model) data and uses PostGIS extensions to analyze road network topology (e.g., road slope gradient calculations) and 3D building outline models (for CFD diffusion simulations of wind resistance coefficients). A spatiotemporal encoder dynamically aligns meteorological data with GIS coordinates, providing high-precision input for pollutant transport models. All heterogeneous data is uniformly mapped to the ISO 8601 time standard and WGS84 coordinate system after ETL processing to ensure spatiotemporal consistency across multiple data sources.

[0050] (2) The data preprocessing and storage module mainly includes a data cleaning and feature engineering submodule and a hierarchical storage and data lake architecture submodule.

[0051] N1. In the data cleaning and feature engineering submodule, the system implements data quality control through a multi-level rule engine. Preset parameter thresholds are dynamically loaded based on vehicle model (e.g., engine speed threshold 0-8000 RPM for fuel vehicles, battery temperature range -30°C to 60°C for new energy vehicles). Dynamic interpolation using a sliding window compensates for signal loss (e.g., when the OBD signal is interrupted, vehicle speed data is filled in based on the average of the preceding and following 5-second windows). Furthermore, feature-derived algorithms are used to construct high-level metrics. For example, instantaneous fuel consumption is calculated based on intake air volume, injection pulse width, and speed (formula: fuel flow rate = injection pulse width × injector flow coefficient × number of cylinders × speed / 2), or battery health (SOH = current maximum capacity / initial capacity × 100%) is estimated by combining battery cycle count and full charge capacity decay rate. Finally, a standardized feature matrix is ​​output for model training. Data lineage tracking is used to mark the processing path of the cleaned data to ensure traceability.

[0052] N2. The tiered storage and data lake architecture submodule adopts a three-tiered "hot-warm-cold" storage strategy. Real-time data is stored in Kafka topic partitions (retained for seven days to facilitate Flink's real-time calculation of emissions violations) and simultaneously written to Apache Parquet columnar files. High-frequency data (such as vehicle trajectories over the past three years and emission forecasts) is stored in a ClickHouse cluster, leveraging its MPP architecture and columnar compression to achieve sub-second response times (e.g., fast retrieval by VIN + timestamp combined index). Low-frequency historical data (annual inspection report PDFs, track points over 10 years old) is archived to HDFS, with cold backup access provided via the MinIO object storage interface. The data lake architecture uses a unified metadata catalog (such as Apache Atlas) to manage cross-tier data lineage and optimizes time travel queries (e.g., backtracking to pollution status at a specific point in time) based on the Iceberg table format.

[0053] (3) The emission prediction and dynamic correction module includes a triple coupling modeling engine submodule and a model online correction submodule.

[0054] U1. In the triple-coupled modeling engine submodule, multimodal deep learning achieves accurate predictions. Based on a 50m×50m grid of road network topology (such as road curvature and slope) and three-dimensional building layout data, a CNN (such as the ResNet-18 architecture) is used to extract spatial features (such as the cumulative emission effects caused by frequent starts and stops at intersections). Simultaneously, time series data such as engine speed and battery SOC are fed into a bidirectional LSTM network using a 60-second sliding window to capture the dynamic evolution of operating conditions (such as the time series pattern of a sharp increase in pollutant concentration during a cold start). Finally, a multi-head attention mechanism is introduced to dynamically adjust feature weights based on environmental parameters (temperature and traffic conditions). For example, in high-temperature environments, the weight of battery cooling energy consumption is increased by 40%. The system outputs minute-by-minute predictions for CO2, NOx, and energy consumption, achieving triple-coupled spatial, temporal, and environmental modeling. This reduces prediction error by 18%-25% compared to a single model.

[0055] U2. The model online correction submodule focuses on long-term prediction stability. By comparing model-predicted emissions with actual annual inspection exhaust gas data (e.g., measured CO 1.5g / km vs. predicted 1.7g / km), it triggers a Bayesian update algorithm (e.g., the NUTS sampler) to calibrate the LSTM unit bias and attention layer weights, eliminating the impact of OBD sensor drift (e.g., air-fuel ratio false alarms caused by oxygen sensor aging) on ​​the model. Furthermore, a vehicle-specific model library is built, extracting historical data for specific models (e.g., a certain brand of plug-in hybrid vehicles) to train dedicated submodels (e.g., adjusting the CNN convolution kernel size to suit its powertrain characteristics). Federated learning aggregates fleet-level features to achieve "global-individual" collaborative optimization, improving emission prediction accuracy by 32% for high-mileage vehicles (>150,000 kilometers).

[0056] (4) The city-level pollutant migration simulation module mainly includes a diffusion dynamics calculation submodule and a pollution source tracing and early warning submodule.

[0057] V1. In the diffusion dynamics calculation submodule, the system drives high-precision pollutant transport simulations based on real-time meteorological and spatial data. It integrates real-time wind direction and speed (updated every 10 minutes), a 3D building BIM model (including aerodynamic roughness parameters), and a road network-level emission heat map (50-meter resolution). A lightweight transient solver (using the k-ε turbulence model and finite volume method discretization) is built using the open-source CFD platform OpenFOAM to dynamically simulate pollutant diffusion paths (such as the vortex retention effect of PM2.5 within building canyons). The solver is optimized to iterate the diffusion field every second. GPU acceleration (CUDA parallel computing) reduces the latency of full 3D simulations for a typical 5 km x 5 km area to under 800 milliseconds, enabling minute-level response for traffic control decisions.

[0058] V2. The pollution source tracing and early warning submodule identifies emission hotspots and predicts risks. It uses the DBSCAN spatial clustering algorithm to analyze diffusion simulation results and integrates road network traffic data to identify high-emission clusters (e.g., identifying an intersection where NOx concentrations exceed the standard by 1.8 times due to a surge in morning and evening rush hour traffic). It also couples the WRF-CMAQ numerical forecasting model to provide a rolling forecast of pollutant concentrations for the next hour (100-meter grid resolution). When the predicted PM2.5 value in a region exceeds 75 μg / m³ (80% of the national daily limit of 75 μg / m³), a multi-level alert is triggered. The hotspot is highlighted in red on the GIS platform, and traffic management authorities are contacted to initiate emergency evacuation measures (e.g., dynamically adjusting traffic light timing). The source tracing function, combined with the inverse trajectory model (HYSPLIT), analyzes the contribution of pollution sources (e.g., identifying an industrial zone 3 kilometers away as the primary source of SO2 in the current area), supporting targeted governance.

[0059] (5) The closed-loop verification and quality control module includes a national control site data comparison submodule and a safety and compliance audit submodule.

[0060] X1. In the national monitoring station data comparison submodule, the system ensures prediction accuracy through spatiotemporal alignment. Pollutant concentration predictions output by the platform (such as hourly averages of CO) are matched with measured data from the Ministry of Ecology and Environment's national monitoring stations. Spatial interpolation (inverse distance weighted method) and a time window sliding average (30-minute granularity) are performed based on the WGS84 coordinate system and UTC timestamps. MAE (mean absolute error) and R² (goodness of fit) are calculated. If the MAE exceeds a threshold for three consecutive days (e.g., NOx error >15%), the system automatically triggers model retraining (e.g., using transfer learning to fine-tune CNN-LSTM model parameters). Kalman filter data assimilation is used to fuse measured and simulated values, increasing the spatial resolution of the diffusion model from 100 meters to 50 meters and dynamically correcting for bias in emission source intensity estimates.

[0061] X2. The security and compliance audit submodule establishes a comprehensive data governance system. Sensitive data (such as VIN numbers and vehicle trajectories) is encrypted and stored using the national SM4 algorithm. Combined with role-based multi-level access control (RBAC), decryption is authorized only to environmental protection departments using dedicated hardware keys. All data operations (such as queries and exports) are recorded in blockchain audit logs (retention period ≥ 6 years), supporting traceability of operator, time, and content. An automated compliance engine also verifies compliance with the Cybersecurity Law and the Personal Information Protection Law in real time (e.g., anonymizing trajectory data and recording cross-border transmission approval records), ensuring the system passes Level 3 security certification.

[0062] (6) The visualization and decision support module includes a real-time three-dimensional visualization submodule and a decision support analysis submodule.

[0063] Y1. In the real-time 3D visualization submodule, a highly immersive web-based pollution situation map is built based on the Cesium.js engine. Using WebGL, a 3D model of urban buildings and a road network-level emissions heat map (e.g., CO2 concentration gradient from blue to red) are rendered. Dynamic animations of pollutant diffusion particles (e.g., PM2.5 flow patterns with wind direction) are displayed, and building shielding effects are simulated (e.g., visualizing the blocking effect of high-rise buildings on pollution transmission). The interactive interface supports selecting any geographic area (e.g., a 2km radius circular selection), automatically aggregating statistical indicators such as total emissions and vehicle type contribution within the area, and generating customized reports (including time trend charts and spatial comparison matrices) that can be exported to PDF / Excel, helping managers quickly locate key pollution areas.

[0064] Y2. The decision support and analysis submodule deeply integrates prediction models and optimization algorithms. For traffic control scenarios, a multi-level coupled simulation model combining traffic restrictions, emissions, and diffusion is constructed to predict the percentage decrease in pollutant concentrations after banning fuel vehicles from specific road sections (e.g., a 23% reduction in NOx during the morning rush hour). A genetic algorithm-based route recommendation engine for prioritizing new energy vehicles is developed. This engine comprehensively considers real-time road conditions, the distribution of charging stations (e.g., the number of available charging stations within 1 km), and energy consumption constraints (e.g., remaining battery power ≥ 30%) to generate a globally optimal set of routes (reducing regional emissions by 15%). This engine also integrates with traffic signal control systems through an API, dynamically adjusting green wave parameters to achieve dual-objective optimization of emissions and traffic efficiency.

[0065] The above embodiments are merely exemplary embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art may make various modifications or equivalent substitutions to the present invention within the spirit and scope of protection of the present invention, and such modifications or equivalent substitutions shall also be deemed to fall within the scope of protection of the present invention.

Claims

1. An intelligent monitoring method for mobile sources on urban roads based on multi-source data coupling, characterized in that: The steps include: Step 1: Collect urban road mobile source data, establish a vehicle emission characteristics database, and process the database data; Step 2: Based on the feature database obtained in step 1, a machine learning algorithm is used to build a pollutant emission, carbon emission and energy consumption prediction model to predict the vehicle's real-time emissions, energy consumption and carbon emissions; Step 3: Obtain vehicle ownership data within the city, expand the per-vehicle emission prediction results obtained in Step 2 to the city scale, and calculate the city's overall pollutant emissions, energy consumption, and carbon emissions; Step 4: Use real-time vehicle location data and urban road network information to construct a dynamic distribution map of automobile emissions in the city. Introduce an atmospheric diffusion model, combine urban geographic information and environmental data to simulate pollutant migration, and compare and verify the simulated migration results with measured data from national monitoring stations to optimize the atmospheric diffusion model parameters. Step 5: Store the results obtained in steps 1-4 into a real-time database and display the pollutant distribution, emissions, energy consumption, and carbon emissions of the urban road network in real time through a visual interface.

2. The intelligent monitoring method for mobile sources on urban roads based on multi-source data coupling according to claim 1 is characterized in that: The urban road mobile source data collection in step 1 specifically includes vehicle identification information data collection, vehicle diagnostic system parameter collection, and annual inspection data collection.

3. The method for intelligent monitoring of mobile sources on urban roads based on multi-source data coupling according to claim 2 is characterized in that: In step 2, the pollutant emission, carbon emission and energy consumption prediction model uses vehicle identification information data collection and on-board diagnostic system parameter collection as model input, and uses the vehicle's real-time emissions, energy consumption and carbon emissions as output, and is optimized using historical annual inspection data; an online learning framework for the emission prediction model is constructed, and the model parameters are dynamically corrected using a transfer learning algorithm through deviation analysis between the actual annual inspection data regularly updated by the vehicle management office and the prediction results.

4. The intelligent monitoring method for mobile sources on urban roads based on multi-source data coupling according to claim 2 is characterized in that: In step 1, for high-frequency, time-series real-time streaming data, use the time series database InfluxDB, leveraging its efficient time window aggregation and streaming processing capabilities; Static structured data is managed through the relational database MySQL to ensure transaction consistency and complex query support; for spatial information data, spatial indexing and geographic operations are implemented based on the PostgreSQL extension component PostGIS.

5. The method for intelligent monitoring of mobile sources on urban roads based on multi-source data coupling according to claim 1 is characterized in that: The specific architecture of the pollutant emission, carbon emission and energy consumption prediction model constructed in step 2 is as follows: using a convolutional neural network to analyze the spatial characteristics of the road network topology and the surrounding building layout, combining a long short-term memory network to capture the long-term dependencies of time series signals, introducing an attention mechanism to dynamically adjust the weights of multi-source features, and realizing the coordinated optimization of spatial-temporal-environmental features. The information flow is controlled through a gating mechanism to avoid a single feature dominating the prediction results.

6. The method for intelligent monitoring of mobile sources on urban roads based on multi-source data coupling according to claim 1 is characterized in that: The process of simulating pollutant migration in step 4 specifically includes: using a multi-scale coupling method to achieve dynamic simulation of pollutant transmission, establishing a three-dimensional simplified model based on computational fluid dynamics, and calculating the turbulence effect and bypass path of pollutants in short-distance diffusion by discretizing boundary conditions and combining real-time wind speed and direction and temperature stratification data; at the same time, for large-scale regional pollution, integrating a Gaussian plume diffusion model or an AERMOD system, using the pollutant source intensity output by the emission prediction model as input, and generating a concentration spatiotemporal distribution matrix by parameterizing meteorological fields and terrain elevation data; The pollutant concentration field is spatially aligned with the road network and population density layers, and heat map rendering technology is used to display the pollution diffusion trend. The timeline can be dragged to playback the historical 24-hour concentration evolution process, and a threshold warning function is embedded. The spatial interpolation algorithm is combined to fill the monitoring blind spot data, while providing multi-layer overlay comparison and supporting interactive parameter adjustment.

7. An intelligent monitoring system for mobile sources on urban roads based on multi-source data coupling, characterized in that: Specifically, it includes a multi-source data acquisition and standardization module, a data preprocessing and storage module, an emission prediction and dynamic correction module, a city-level pollutant migration simulation module, a closed-loop verification and quality control module, and a visualization and decision support module. The multi-source data acquisition and standardization module includes a vehicle static data acquisition submodule, a dynamic OBD data real-time access submodule, an annual inspection and historical data integration submodule, and an environmental and geographic data acquisition submodule; the data preprocessing and storage module includes a data cleaning and feature engineering submodule, a hierarchical storage and data lake architecture submodule; the emission prediction and dynamic correction module includes a triple coupling modeling engine submodule for building a storage prediction model, and a model online correction submodule focusing on long-term prediction stability; the city-level pollutant migration simulation module includes a diffusion dynamics calculation submodule for pollutant transmission simulation, and a pollution tracing and early warning submodule for emission hotspot identification and risk prediction; the closed-loop verification and quality control module includes a national control site data comparison submodule and a safety and compliance audit submodule; the visualization and decision support module includes a real-time three-dimensional visualization submodule and a decision support analysis submodule.

8. The urban road mobile source intelligent monitoring system based on multi-source data coupling according to claim 7 is characterized in that: The data cleaning and feature engineering submodule implements data quality control through a multi-level rule engine, dynamically loads preset parameter thresholds based on vehicle models, and uses sliding windows for dynamic interpolation to fill signal loss; constructs high-order indicators through feature derivation algorithms, and finally outputs a standardized feature matrix for model training; the hierarchical storage and data lake architecture submodule adopts a three-level storage strategy, and real-time data is stored in the original OBD stream through Kafka Topic partitions and synchronously written to Apache Parquet columnar files; high-frequency access data is stored in the ClickHouse cluster, and its MPP architecture and columnar compression are used to achieve sub-second response; low-frequency historical data is archived to HDFS, and cold backup access is provided through the MinIO object storage interface.

Images