An artificial intelligence-based vehicle emission identification system and method

Abstract

The application discloses a vehicle emission identification system and method based on artificial intelligence, and relates to the technical field of vehicle emission detection; the snapshot identification module is provided with a multispectral camera array and a synchronous controller, combines a laser radar and a millimeter wave radar to dynamically adjust imaging parameters, and supports multi-type vehicle identification; an AI algorithm processing module adopts a heterogeneous computing power architecture to distribute resources; a vehicle anomaly determination module constructs a multi-dimensional anomaly feature chain, calls OBD data to research and judge anomalies, and identifies blacklisted vehicles; a data storage and calling module adopts block chain storage and hierarchical storage; a mobile law enforcement adaptation module supports end-edge-cloud collaboration; a multi-source data fusion module realizes data space-time alignment; and a dynamic model self-updating module optimizes a model based on federated learning. The application improves the comprehensiveness of vehicle identification and anomaly determination, guarantees the legality and credibility of law enforcement data, relieves the computing power pressure of mobile terminals, supports model continuous adaptation technology iteration, helps efficient mobile law enforcement, and provides strong support for environmental protection law enforcement.

Description

Technical Field

[0001] This invention relates to the field of vehicle emission detection technology, and in particular to a vehicle emission identification system and method based on artificial intelligence. Background Technology

[0002] Vehicle emissions testing is a crucial step in ensuring air quality and enforcing environmental regulations. Traditional emissions testing relies heavily on large equipment at fixed stations, limiting testing to specific areas and failing to meet the needs of mobile enforcement. Furthermore, the testing process requires vehicles to be parked, resulting in low efficiency. While portable testing devices have emerged to meet the growing demand for mobile enforcement, these devices often employ single-spectrum imaging or simple data acquisition methods. Their ability to capture multi-dimensional information such as vehicle appearance, exhaust composition, and powertrain thermal characteristics is insufficient, limiting the accuracy of basic identification processes like license plate recognition and emissions stage determination. This is particularly problematic under complex lighting conditions and high-speed driving, where recognition rates are difficult to guarantee. Simultaneously, vehicle emissions cheating methods are becoming increasingly sophisticated, such as CALID / CVN rewriting and tampering with exhaust aftertreatment devices. Existing equipment lacks the ability to analyze and correlate multi-dimensional abnormal behaviors, making it difficult to accurately identify these covert cheating methods and leading to missed or false detections.

[0003] Furthermore, existing mobile emission detection equipment lacks flexibility in its computing resource allocation, often employing a single processor architecture. This results in slow processing speeds for computationally intensive tasks such as multispectral image analysis and deep feature extraction, failing to meet real-time enforcement needs. Regarding data storage and retrieval, traditional storage methods lack blockchain technology, making data susceptible to tampering and unreliable as legally valid enforcement evidence. Data sharing between cross-regional enforcement departments is also difficult, impacting enforcement collaboration. In terms of AI models, most devices use factory-preset recognition models, unable to dynamically update based on new vehicle technologies and cheating methods, leading to a gradual decline in recognition accuracy over long-term use. Simultaneously, the "edge-cloud" collaboration mechanism is incomplete. Mobile enforcement terminals have limited computing power, making it difficult to independently handle complex data processing, while high latency between the cloud and the terminal further restricts enforcement efficiency and recognition capabilities. These issues collectively result in significant shortcomings in recognition accuracy, enforcement efficiency, and technological adaptability of existing vehicle emission recognition technologies in mobile enforcement scenarios. Summary of the Invention

[0004] This invention proposes an artificial intelligence-based vehicle emission identification system and method to solve the problems mentioned in the prior art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a vehicle emission identification system based on artificial intelligence, comprising:

[0006] The image capture and recognition module is equipped with a multispectral camera array consisting of three visible light cameras, two infrared thermal imaging cameras, and one ultraviolet spectral camera. Each camera unit is synchronized via a synchronous trigger controller. Based on real-time vehicle 3D contour data acquired by LiDAR scanning, the module dynamically adjusts multispectral imaging parameters. The visible light cameras focus on vehicle exterior and license plate details, the infrared thermal imaging cameras capture the thermal radiation characteristics of the vehicle's powertrain and exhaust gases, and the ultraviolet spectral camera collects spectral information of characteristic pollutants in the exhaust gases. The module performs synchronous capture of multispectral images, supporting license plate recognition for vehicles meeting China VI, China V, China IV, China III, and new energy vehicle standards. In typical sunny scenarios, the recognition rate for typical license plates is ≥99.9%. The module uses onboard radar to acquire real-time vehicle speed and, combined with an exhaust gas diffusion velocity model, adaptively adjusts the image acquisition frame rate and multispectral exposure time. When an abnormal increase in exhaust gas concentration is detected, the exposure time of the ultraviolet spectral camera is automatically extended to enhance the acquisition of pollutant spectral signals.

[0007] AI algorithm processing module: It has a built-in heterogeneous computing power scheduling unit, which integrates one NPU and two FPGA chips to form a heterogeneous computing power architecture with NPU as the main body and FPGA acceleration; it dynamically allocates computing power resources for license plate recognition, emission spectrum analysis and three-dimensional contour matching; it performs deep convolution and feature extraction on captured images and data to generate a 1024-dimensional vehicle appearance feature vector, a 256-dimensional license plate feature vector and a 512-dimensional exhaust emission spectrum feature vector.

[0008] Vehicle anomaly detection module: Constructs an abnormal behavior feature chain that includes OBD data consistency, exhaust gas composition matching degree, power system thermal characteristic deviation degree, and control unit communication characteristics, containing a total of 12 core abnormal features and 36 derivative features; Real-time access to OBD online monitoring data via the vehicle CAN bus, combined with exhaust gas composition spectral data and vehicle operating status data extracted by the AI ​​algorithm processing module;

[0009] Data storage and retrieval module: Adopts a fusion architecture of blockchain notarization and hierarchical storage; blockchain notarization is based on consortium blockchain technology, selecting law enforcement agencies, car manufacturers, and testing institutions as consensus nodes; in terms of hierarchical storage, data is divided into hot data layer, warm data layer, and cold data layer according to access frequency and importance. The hot data layer uses high-speed SSD storage; the warm data layer uses HDD storage; and the cold data layer uses tape library storage, supporting fast data retrieval, historical review, and data sharing among law enforcement agencies in different regions.

[0010] Mobile law enforcement adaptation module: Equipped with a thermal management battery pack, it adopts a thermal management method that combines liquid cooling and air cooling, integrates the Beidou positioning system and 5G / Edge computing unit, and combines with the law enforcement APP to realize end-edge-cloud collaboration. The on-site law enforcement terminal is responsible for data collection and preliminary processing, the edge node undertakes parallel computing and temporary storage of data, and the cloud law enforcement platform performs big data analysis and global scheduling.

[0011] Data fusion module: Equipped with a clock synchronization unit, it performs spatiotemporal alignment on the multispectral image data from the capture and recognition module, the OBD and exhaust gas composition data from the vehicle anomaly judgment module, and the feature extraction data from the AI ​​algorithm processing module. It adopts an attention-based fusion algorithm to generate a 2048-dimensional multi-dimensional feature vector for vehicle emission recognition, providing a data foundation for subsequent judgment.

[0012] Dynamic Model Self-Update Module: Based on the federated learning framework, an iterative process of local training, parameter upload, global aggregation, and model distribution is designed. Each law enforcement terminal acts as a client of federated learning, training the AI ​​recognition model locally using newly added law enforcement data. After training, only the updated part of the model parameters is uploaded to the cloud aggregation server. The cloud server uses a weighted aggregation algorithm to generate a globally updated model and distributes it to each terminal, realizing continuous optimization of the AI ​​recognition model and the abnormal behavior feature chain.

[0013] Furthermore, in the capture and recognition module, the dynamic adjustment of multispectral imaging parameters satisfies the condition that the visible light exposure time t... v Infrared exposure time t i With UV exposure time t u The collaborative relationship is ,in For visible light weighting coefficients, For infrared weighting coefficients, is the ultraviolet weighting coefficient, and T is the threshold for total exposure time of multispectral imaging; through the cooperative relationship, multispectral images can capture vehicle appearance details, thermal radiation characteristics and exhaust pollutant spectra under different lighting and emission scenarios.

[0014] Furthermore, the heterogeneous computing power scheduling unit of the AI ​​algorithm processing module allocates computing power to meet the following requirements: , where P NPU To allocate computing power to the NPU, D NPU D represents the computational power density required by the NPU for the current subtask. FPGA P represents the FPGA's computational power demand density for the current subtask, α represents the NPU's computational efficiency coefficient, β represents the FPGA's computational efficiency coefficient, and P... total Total available computing power for AI algorithm processing modules.

[0015] Furthermore, the calculation of the multi-dimensional abnormal behavior feature chain and the abnormal correlation degree R of the vehicle anomaly determination module satisfies... Where m is the number of abnormal features involved in the association, and wjk is the association weight between the j-th feature and the k-th feature. Let xj be the Pearson correlation coefficient between the j-th eigenvector xj and the k-th eigenvector xk.

[0016] Furthermore, during the blockchain notarization process of the data storage and retrieval module, the data block generation time interval... satisfy ,in The initial block generation interval is defined as k, which is the block generation adjustment coefficient, and N is the amount of data currently stored. As the amount of stored data increases, the block generation interval is dynamically shortened.

[0017] Furthermore, in the end-edge-cloud collaboration of the mobile law enforcement adaptation module, the edge computing task offloading amount U satisfies... Where D(t) is the terminal computing task at time t, C(t) is the edge node computing power at time t, B(t) is the terminal remaining computing power at time t, t0 is the task unloading start time, and t1 is the task unloading end time.

[0018] Furthermore, in the federated learning parameter update of the dynamic model self-updating module, the local model parameter update amount... satisfy ,in For learning rate, Let be the gradient of the local data of the i-th terminal with respect to the model parameters θ. The momentum coefficient, This represents the parameter update amount from the last time.

[0019] A method for applying the aforementioned AI-based vehicle emissions identification system includes:

[0020] Multispectral collaborative capture steps: The multispectral camera array of the capture and recognition module acquires the vehicle's three-dimensional contour data in real time through LiDAR scanning. Combined with the vehicle's driving speed acquired by the onboard radar, the exposure time and acquisition frame rate of the visible light, infrared, and ultraviolet imaging units are dynamically adjusted. The visible light camera focuses on the front and rear of the vehicle, the infrared thermal imaging camera is aimed at the vehicle's engine compartment and exhaust emission area, and the ultraviolet spectral camera is pointed at the exhaust gas diffusion path to complete the synchronous capture of multispectral images and simultaneously acquire the ultraviolet spectral information of the vehicle's exhaust emissions.

[0021] Heterogeneous computing power scheduling and feature extraction steps: The heterogeneous computing power scheduling unit of the AI ​​algorithm processing module dynamically allocates computing power resources of NPU and FPGA according to the current task type; after preprocessing the multispectral capture image, the deep convolutional neural network is used to extract vehicle appearance feature vector, license plate feature vector and exhaust emission spectrum feature vector.

[0022] Data fusion and preliminary identification steps: The feature vectors of the captured images, the OBD online monitoring data retrieved by the vehicle anomaly judgment module, and the vehicle operating status data are spatiotemporally aligned by the clock synchronization unit. An attention mechanism fusion algorithm is used to generate a multi-dimensional fusion feature vector with a dimension of 2048. Based on the vector, the AI ​​algorithm processing module uses a pre-trained classification model to achieve a preliminary determination of the vehicle emission stage, with an accuracy rate of ≥96%.

[0023] Anomaly correlation analysis steps: The vehicle anomaly determination module retrieves OBD online monitoring data, exhaust gas composition spectrum data, and vehicle operating status data to construct an abnormal behavior feature chain and calculate the anomaly correlation degree; at the same time, it writes abnormal behavior to CALID / CVN and calculates the anomaly confidence degree; when the anomaly correlation degree or anomaly confidence degree exceeds the preset threshold, it determines that the vehicle has abnormal behavior and completes the identification and warning of blacklisted vehicles.

[0024] Blockchain evidence storage and hierarchical storage steps: The data storage and retrieval module generates blockchain blocks to store the captured images, recognition results, and anomaly detection data; at the same time, the data is classified and stored in the hot data layer, warm data layer, or cold data layer according to the data access frequency and importance.

[0025] Edge-cloud collaborative law enforcement steps: The mobile law enforcement adaptation module dynamically adjusts the system's operating power based on the remaining battery power and computing load; through the 5G / Edge computing unit, some data processing tasks are offloaded to edge nodes; Beidou positioning data and law enforcement APP are integrated to achieve real-time synchronization between on-site law enforcement data and the cloud law enforcement platform, and the cloud platform can remotely retrieve historical data and global law enforcement statistics;

[0026] Federated learning model update steps: The dynamic model self-update module initiates the federated learning process. Each law enforcement terminal updates the parameters of the AI ​​recognition model locally using the newly added law enforcement data. Only the updated model parameters are uploaded to the cloud aggregation server. The cloud uses a weighted aggregation algorithm to generate a globally updated model and then distributes it to each terminal to achieve continuous model optimization.

[0027] Furthermore, in the multispectral collaborative capture step, the dynamic adjustment of multispectral imaging parameters also satisfies the following: when the vehicle exhaust emission diffusion velocity is vg, the spatial sampling interval d of the multispectral camera array satisfies , where K is the exhaust gas feature spatial resolution coefficient, and f is the image acquisition frame rate.

[0028] Furthermore, in the anomaly correlation analysis step, for the identification of CALID / CVN write anomalies, the anomaly confidence level C is... cheat satisfy ,in Let be the matching coefficient between the brushing behavior features and the standard features at time t. Let t be the deviation between the OBD data at time t and the vehicle's factory reference data. a t represents the start time for abnormal behavior monitoring. b This is the end time for abnormal behavior monitoring.

[0029] Compared with existing technologies, the beneficial effects of this invention are:

[0030] By employing multispectral collaborative capture and dynamic imaging parameter adjustment, comprehensive information on vehicle appearance, thermal radiation, and exhaust pollutant spectra can be captured, providing richer and more accurate foundational data for subsequent identification and effectively improving the completeness and clarity of vehicle recognition in complex scenarios. Leveraging heterogeneous computing power scheduling and multi-source data fusion technology, the advantages of different computing units are fully utilized to accelerate feature extraction and analysis. Simultaneously, multi-dimensional data is integrated to generate high-dimensional fusion features, significantly improving the accuracy of vehicle emission stage determination and abnormal behavior identification, and reducing missed and false detections.

[0031] The combination of multi-dimensional anomaly correlation analysis and blockchain evidence storage technology can not only accurately identify various abnormal behaviors, including covert data spoofing, but also ensure the authenticity and legal validity of law enforcement data, providing a reliable data foundation for cross-regional law enforcement collaboration. The "end-edge-cloud" collaborative mechanism and federated learning model update method alleviate the computing power pressure on mobile terminals, improve law enforcement response speed, and allow the AI ​​recognition model to continuously adapt to vehicle technology iterations and new cheating methods, ensuring long-term recognition accuracy and technological adaptability, and promoting the development of mobile vehicle emission enforcement towards intelligence, efficiency, and collaboration. Overall, this invention comprehensively improves the comprehensiveness, accuracy, enforcement efficiency, and technological adaptability of vehicle emission identification, providing stronger technical support for environmental law enforcement. Attached Figure Description

[0032] Figure 1 This is a schematic block diagram of a vehicle emission identification system based on artificial intelligence proposed in this invention;

[0033] Figure 2 This is a schematic diagram of a vehicle emission identification method based on artificial intelligence proposed in this invention;

[0034] Figure 3 This is a schematic diagram illustrating the relationship between the heterogeneous computing power allocation ratio and task type in an artificial intelligence-based vehicle emission identification method proposed in this invention.

[0035] Figure 4 This diagram illustrates the relationship between the anomaly detection accuracy and feature dimension of the vehicle emission identification method based on artificial intelligence proposed in this invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0038] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0039] Reference Figures 1 to 4 A vehicle emission identification system based on artificial intelligence includes a capture and identification module, an AI algorithm processing module, a vehicle anomaly detection module, a data storage and retrieval module, a mobile law enforcement adaptation module, a multi-source data fusion module, and a dynamic model self-updating module.

[0040] The image capture and recognition module is equipped with a multispectral camera array consisting of at least three visible light cameras, two infrared thermal imaging cameras, and one ultraviolet spectral camera. Each camera unit achieves millisecond-level time synchronization via a synchronization trigger controller. Based on real-time vehicle 3D contour data acquired by LiDAR scanning, the multispectral imaging parameters are dynamically adjusted. The visible light cameras focus on vehicle exterior and license plate details, the infrared thermal imaging cameras capture the thermal radiation characteristics of the vehicle's powertrain and exhaust gases, and the ultraviolet spectral camera collects spectral information of characteristic pollutants in the exhaust gases. This module has an image acquisition capability of over 8 megapixels and can simultaneously capture multispectral images within 0.1 seconds. It supports license plate recognition for various vehicle types, including those meeting China VI, China V, China IV, China III, and new energy vehicles. Under typical sunny conditions, the recognition rate for typical license plates is ≥99.9%. By acquiring vehicle speed in real time through onboard millimeter-wave radar and combining it with a gas diffusion velocity model from the exhaust outlet, the module adaptively adjusts the image acquisition frame rate and multispectral exposure time. When the vehicle speed exceeds 60 km / h, the frame rate is increased from the basic 30fps to 60fps. When an abnormal increase in exhaust emission concentration is detected, the exposure time of the ultraviolet spectral camera is automatically extended by 2-5 times to enhance the acquisition effect of pollutant spectral signals.

[0041] The AI ​​algorithm processing module features a built-in heterogeneous computing power scheduling unit, integrating a high-performance NPU with a computing power of 20 TOPS and two FPGA chips, each with a computing power of 5 TOPS, forming a heterogeneous computing power architecture of "NPU-led + FPGA-accelerated". It dynamically allocates computing resources for different sub-tasks of vehicle recognition, such as license plate recognition, emission spectrum analysis, and 3D contour matching; it performs deep convolution and feature extraction on captured images and multi-source data, generating a 1024-dimensional vehicle appearance feature vector, a 256-dimensional license plate feature vector, and a 512-dimensional exhaust emission spectrum feature vector, providing a foundation for subsequent recognition and judgment.

[0042] The vehicle anomaly detection module constructs a multi-dimensional abnormal behavior feature chain, including "OBD data consistency," "exhaust gas composition matching degree," "powertrain thermal characteristic deviation," and "control unit communication characteristics," encompassing 12 core anomaly features and 36 derived features. It retrieves OBD online monitoring data in real time via the vehicle's CAN bus, combining it with exhaust gas composition spectral data extracted by the AI ​​algorithm processing module and vehicle operating status data. This allows for multi-dimensional correlation analysis to determine whether the vehicle exhibits abnormal behaviors such as CALID / CVN rewriting, exhaust aftertreatment device tampering, or emission control system failure. The anomaly behavior recognition rate exceeds 98%. Simultaneously, it incorporates a blacklist vehicle database containing 100,000 records, supporting sub-second rapid identification and audible / visual warnings for blacklisted vehicles.

[0043] Data storage and retrieval module: Adopts a fusion architecture of "blockchain notarization + layered storage". Blockchain notarization is based on consortium blockchain technology, selecting law enforcement agencies, car manufacturers, and testing institutions as consensus nodes; for layered storage, data is divided into hot data layer, warm data layer, and cold data layer according to access frequency and importance. The hot data layer uses high-speed SSD storage with read / write latency <1ms; the warm data layer uses enterprise-grade HDD storage; and the cold data layer uses tape library storage, supporting fast data retrieval, historical backtracking, and data sharing across law enforcement agencies in different regions, with a cross-regional data retrieval response time of <2s.

[0044] Mobile law enforcement adaptation module: Equipped with a 60Ah intelligent thermal management battery pack, employing a combined liquid and air cooling thermal management method, supporting over 12 hours of continuous operation in environments ranging from -20℃ to 60℃. It integrates a BeiDou high-precision positioning system and a 5G / Edge computing unit, with a 5G communication rate ≥1Gbps and an Edge computing unit computing power of 8TOPS. Combined with a dedicated law enforcement APP, it achieves "end-edge-cloud" collaboration. On-site law enforcement terminals are responsible for data collection and preliminary processing, edge nodes handle parallel computing and temporary storage of complex data, and the cloud-based law enforcement platform performs big data analysis and global scheduling.

[0045] Multi-source data fusion module: Equipped with a clock synchronization unit with a time synchronization accuracy of 10μs, it performs spatiotemporal alignment on the multispectral image data of the capture and recognition module, the OBD and exhaust gas composition data of the vehicle anomaly judgment module, and the feature extraction data of the AI ​​algorithm processing module. It adopts a fusion algorithm based on the attention mechanism to generate a multi-dimensional feature vector of vehicle emission recognition with a dimension of 2048, providing a unified and high-dimensional data foundation for subsequent accurate judgment.

[0046] The dynamic model self-updating module, based on the federated learning framework, designs an iterative process of "local training - parameter upload - global aggregation - model distribution". Each law enforcement terminal acts as a client of federated learning, training the AI ​​recognition model locally using newly added law enforcement data. After training, it only uploads the updated model parameters to the cloud aggregation server. The cloud server uses a weighted aggregation algorithm to generate a globally updated model and distributes it to each terminal, enabling continuous optimization of the AI ​​recognition model and abnormal behavior feature chain to adapt to the iteration of vehicle technology and the evolution of emissions cheating methods.

[0047] In this invention, the dynamic adjustment of multispectral imaging parameters in the capture and recognition module satisfies: visible light exposure time t v Infrared exposure time t i With UV exposure time t u The collaborative relationship is ,in For visible light weighting coefficients, For infrared weighting coefficients, , where is the ultraviolet weighting coefficient, and T is the threshold for total exposure time in multispectral imaging. Through this synergistic relationship, under different lighting and emission scenarios, the multispectral images are ensured to fully capture vehicle exterior details, thermal radiation characteristics, and exhaust pollutant spectra. This enables subsequent AI algorithm processing modules to extract richer and more accurate features from the images, improving the comprehensiveness and accuracy of vehicle identification and emission determination.

[0048] According to claim 1, the vehicle emission identification system based on artificial intelligence is characterized in that the heterogeneous computing power scheduling unit of the AI ​​algorithm processing module allocates computing power in accordance with... , where P NPU To allocate computing power to the NPU, D NPU D represents the computational power density required by the NPU for the current subtask. FPGA P represents the FPGA's computational power demand density for the current subtask, α represents the NPU's computational efficiency coefficient, β represents the FPGA's computational efficiency coefficient, and P... total This allocation method ensures the total available computing power for the AI ​​algorithm processing module. By fully leveraging the advantages of the NPU in floating-point operations and the FPGA in parallel acceleration, the overall algorithm efficiency is improved, providing computing power support for real-time emission identification.

[0049] In this invention, the multi-dimensional abnormal behavior feature chain of the vehicle anomaly determination module, and the calculation of the anomaly correlation degree R, satisfy the following... Where m is the number of abnormal features involved in the association, and wjk is the association weight between the j-th feature and the k-th feature. Let xj be the Pearson correlation coefficient between the j-th feature vector xj and the k-th feature vector xk. By calculating the correlation between multiple anomalous features, we can effectively identify covert cheating behaviors where "a single feature is normal but multiple features are linked to be abnormal," avoiding misjudgments based on a single feature and improving the accuracy and comprehensiveness of anomalous behavior identification.

[0050] In this invention, during the blockchain notarization process of the data storage and retrieval module, the data block generation time interval is... satisfy ,in The initial block generation interval is defined as k, the block generation adjustment coefficient is k, and the amount of data currently stored is N. As the amount of stored data increases, the block generation interval is dynamically shortened. This ensures the efficiency of blockchain consensus while increasing the difficulty of data tampering, ensuring the authenticity and legal validity of law enforcement data, and providing a credible foundation for cross-regional mutual recognition of law enforcement data.

[0051] In this invention, in the "end-edge-cloud" collaboration of the mobile law enforcement adaptation module, the edge computing task offloading amount U satisfies Where D(t) is the terminal's computing task load at time t, C(t) is the edge node's computing power at time t, B(t) is the terminal's remaining computing power at time t, t0 is the task unloading start time, and t1 is the task unloading end time. Through this integral calculation, the edge computing task unloading ratio is dynamically adjusted, significantly reducing the terminal's computing load and improving the efficiency and response speed of mobile law enforcement.

[0052] In this invention, the local model parameter update amount is included in the federated learning parameter update of the dynamic model self-updating module. satisfy ,in For learning rate, Let be the gradient of the local data of the i-th terminal with respect to the model parameters θ. The momentum coefficient, This represents the parameter update amount from the previous iteration. By introducing a momentum term, the direction of parameter updates becomes more consistent, accelerating the convergence speed of model parameters in federated learning and ensuring continuous optimization and adaptation of the model.

[0053] This invention also discloses a method for applying the aforementioned artificial intelligence-based vehicle emission identification system, comprising the following steps:

[0054] Multispectral collaborative capture steps: The multispectral camera array of the capture and recognition module acquires the vehicle's three-dimensional contour data in real time through LiDAR scanning. Combined with the vehicle's driving speed acquired by the onboard millimeter-wave radar, the exposure time and acquisition frame rate of the visible light, infrared, and ultraviolet imaging units are dynamically adjusted. The visible light camera focuses on the front and rear of the vehicle, the infrared thermal imaging camera is aimed at the vehicle's engine compartment and exhaust emission area, and the ultraviolet spectral camera is pointed at the exhaust gas diffusion path. The multispectral image is captured synchronously within 0.1 seconds, while simultaneously acquiring the ultraviolet spectral information of the vehicle's exhaust emissions. When the vehicle speed exceeds 60 km / h, the image acquisition frame rate is increased from 30 fps to 60 fps. When an abnormal increase in exhaust emission concentration is detected, the exposure time of the ultraviolet spectral camera is extended to 3 times the original time to ensure the complete acquisition of multi-dimensional images and spectral information.

[0055] Heterogeneous computing power scheduling and feature extraction steps: The heterogeneous computing power scheduling unit of the AI ​​algorithm processing module dynamically allocates computing resources of NPU and FPGA according to the current task type; after preprocessing the multispectral captured images, a deep convolutional neural network is used to extract vehicle appearance feature vectors, license plate feature vectors, and exhaust emission spectral feature vectors, with each feature vector extraction taking ≤100ms. Multi-source data fusion and preliminary identification steps: The multi-source data fusion module uses a clock synchronization unit to perform spatiotemporal alignment of the captured image feature vectors, OBD online monitoring data retrieved by the vehicle anomaly judgment module, and vehicle operating status data, and uses an attention mechanism fusion algorithm to generate a multi-dimensional fusion feature vector with a dimension of 2048; based on this vector, the AI ​​algorithm processing module uses a pre-trained classification model to achieve a preliminary determination of the vehicle emission stage, with an accuracy rate ≥96%;

[0056] Multi-dimensional anomaly correlation analysis steps: The vehicle anomaly determination module retrieves OBD online monitoring data, exhaust gas composition spectrum data, and vehicle operating status data to construct a multi-dimensional abnormal behavior feature chain and calculate the anomaly correlation degree. Simultaneously, it writes abnormal behavior to CALID / CVN and calculates the anomaly confidence level. When the anomaly correlation degree or anomaly confidence level exceeds a preset threshold, the vehicle is determined to have abnormal behavior, and sub-second identification and warning of blacklisted vehicles are completed. Blockchain notarization and hierarchical storage steps: The data storage and retrieval module generates blockchain blocks to notarize captured images, recognition results, and anomaly determination data. Simultaneously, based on data access frequency and importance, data is categorized and stored in hot data layers, warm data layers, or cold data layers to ensure efficient data storage and long-term preservation.

[0057] Edge-cloud collaborative law enforcement steps: The mobile law enforcement adaptation module dynamically adjusts the system's operating power based on the remaining battery power and computing load; through the 5G / Edge computing unit, some data processing tasks are offloaded to edge nodes; Beidou high-precision positioning data and law enforcement APP are integrated to achieve real-time synchronization between on-site law enforcement data and the cloud law enforcement platform, and the cloud platform can remotely retrieve historical data and global law enforcement statistics;

[0058] Federated learning model update steps: The dynamic model self-update module initiates the federated learning process. Each law enforcement terminal updates the parameters of the AI ​​recognition model locally using the newly added law enforcement data. Only the updated model parameters are uploaded to the cloud aggregation server. The cloud uses a weighted aggregation algorithm to generate a globally updated model and then distributes it to each terminal to achieve continuous model optimization. A full system model update is completed once every quarter.

[0059] In this invention, during the multispectral collaborative capture step, the dynamic adjustment of the multispectral imaging parameters also satisfies the following: when the vehicle exhaust emission diffusion velocity is vg, the spatial sampling interval d of the multispectral camera array satisfies Where K is the spatial resolution coefficient of exhaust gas features, and f is the image acquisition frame rate. This spatial sampling interval ensures the spatial resolution of the spectral imaging of pollutants in the exhaust gas, enabling the subsequent AI algorithm processing module to accurately extract the spatial distribution characteristics of exhaust gas components and improve the ability to identify low-concentration emission anomalies.

[0060] In this invention, during the multi-dimensional anomaly correlation analysis step, the anomaly confidence level C is used to identify CALID / CVN write anomalies. cheat satisfy ,in Let be the matching coefficient between the brushing behavior features and the standard features at time t. Let t be the deviation between the OBD data at time t and the vehicle's factory reference data. a t represents the start time for abnormal behavior monitoring. b This represents the end time of abnormal behavior monitoring. This integral calculation fully considers the cumulative effect of abnormal behavior over time. Compared with the traditional "single-point judgment," it can more accurately identify short-duration but highly harmful write behaviors, improving the accuracy and timeliness of abnormal identification.

[0061] A specific implementation of an artificial intelligence-based vehicle emissions identification system and method:

[0062] (Example 1: Mobile law enforcement scenario on urban roads)

[0063] This embodiment addresses the mobile law enforcement needs during morning and evening rush hours on urban main roads. The enforcement vehicles are new energy electric vehicles (range ≥300km). The system is adaptable to urban ambient temperatures ranging from -10℃ to 40℃ and needs to handle traffic congestion (speed 0-30km / h), rapid passage (speed 30-60km / h), and complex lighting conditions (strong sunlight on sunny days, weak sunlight on cloudy days). It supports the identification of all types of vehicles from China III to China VI emission standards, as well as new energy vehicles. The specific implementation is as follows:

[0064] I. Hardware Deployment Details

[0065] The image capture and recognition module is equipped with a multispectral camera array. The visible light camera is a Hikvision DS-2CD6A20UVF (8 megapixels, focal length 8-32mm), the infrared thermal imaging camera is a FLIRA X8 (resolution 640×512, temperature range -20℃ to 150℃), and the ultraviolet spectral camera is an OceanOptics Flame-T (spectral range 200-1100nm, resolution 0.3nm). The lidar is a RoboSense RS-LIDAR-M1 (scanning frequency 10Hz, ranging range 0.1-200m), and the millimeter-wave radar is a Continental ARS408 (ranging range 0.5-250m, speed accuracy ±0.1km / h). The synchronous trigger controller is an NICDAQ-9178, which enables multi-camera timing synchronization within 0.5ms.

[0066] AI algorithm processing module: The heterogeneous computing unit adopts the architecture of "Horizon Journey 5 NPU (20 TOPS) + two Xilinx Zynq UltraScale + FPGA (5 TOPS each)", and is equipped with the Linux operating system. AI model deployment is accelerated by TensorRT, and the model inference latency is ≤80ms.

[0067] Vehicle anomaly detection module: OBD data is retrieved via ELM327 Bluetooth adapter (communication rate 115200bps), the blacklist database is stored on local SSD (capacity 128GB), supports offline query, and is updated once a day (incremental packages are downloaded via 5G).

[0068] Data storage and retrieval module: The blockchain node adopts Huawei Cloud ECS (4 cores 8G). In the tiered storage architecture, the hot data layer uses Samsung 990 Pro SSD (2TB, read and write speed 7450MB / s), the warm data layer uses Seagate Exos X18HDD (18TB, 7200rpm), and the cold data layer is temporarily stored locally (uploaded to the municipal law enforcement cloud tape library every week).

[0069] Mobile law enforcement adaptation module: The battery pack uses CATL lithium iron phosphate battery (60Ah, voltage 3.2V), and the thermal management adopts air cooling (fan model NidecU60T24M). The Beidou positioning module uses Hexin Xingtong UM220-IV (positioning accuracy 1m), and the 5G module uses Huawei ME909s-821 (download speed 1.6Gbps).

[0070] Multi-source Data Fusion and Dynamic Model Self-update Module: The clock synchronization unit selects Symmetricom XLi (synchronization accuracy 10 μs), the federated learning client is deployed on the edge computing unit of the law enforcement vehicle (computing power 8 TOPS), the local training framework uses PyTorch, the learning rate η = 0.001, and the momentum coefficient γ = 0.9.

[0071] II. Full-process Implementation Steps

[0072] 1. Multi-spectral Cooperative抓拍

[0073] The law enforcement vehicle travels to the urban main road (such as the intersection of Jiefang Road and Renmin Road). The lidar scans the passing vehicles in real time to obtain the three-dimensional contour data of a certain car meeting the national stage V emission standard (vehicle length 4.8 m, vehicle height 1.5 m). The millimeter-wave radar detects its driving speed as 50 km / h. According to the vehicle speed, the capture and recognition module dynamically adjusts the parameters: the image acquisition frame rate is increased from the basic 30 fps to 45 fps; the multi-spectral exposure time is calculated according to the cooperative formula where λ v = 0.4 (visible light weight), λ i = 0.3 (infrared weight), λ u = 0.3 (ultraviolet weight), T = 0.05 s (total exposure threshold). Substituting these values, we get t v = 0.02 s (visible light), t i = 0.015 s (infrared), t u = 0.015 s (ultraviolet). The visible light camera focuses on the license plate (Beijing A12345), the infrared camera captures the engine compartment temperature (85 °C) and the exhaust gas thermal radiation (45 °C), and the ultraviolet camera collects the characteristic spectrum of NOx in the exhaust gas (wavelength 398 nm). Synchronized capture is completed within 0.1 seconds.

[0074] 2. Heterogeneous Computing Power Scheduling and Feature Extraction

[0075] The AI algorithm processing module receives the multi-spectral images. The heterogeneous computing power scheduling unit allocates the computing power according to the formula : The current task is license plate recognition (D NPU = 0.8, D FPGA = 0.5), α = 1.2 (NPU efficiency coefficient), β = 0.8 (FPGA efficiency coefficient), P total = 25 TOPS (total computing power). Calculating, we get P NPU = 16.8 TOPS, and the remaining 8.2 TOPS are allocated to the FPGA. The NPU uses ResNet-50 to extract a 1024-dimensional vehicle appearance feature vector (including body lines and headlight shapes). The FPGA accelerates the license plate character segmentation to generate a 256-dimensional license plate feature vector. The feature extraction takes 85 ms, and the license plate recognition result is "Beijing A12345" with a recognition rate of 100%.

[0076] 3. Multi-source data fusion and anomaly detection

[0077] The multi-source data fusion module, through a clock synchronization unit, spatiotemporally aligns the feature vector with OBD data (2000 rpm, 60 kPa intake pressure) and exhaust gas spectral data (800 ppm NOx concentration), and uses ScaledDot-ProductAttention to generate a 2048-dimensional vector. The vehicle anomaly detection module calculates the anomaly correlation degree. m=4 (OBD, spectral, thermal features, communication features), w12=0.3 (OBD and spectral correlation weights), ρ(x1,x2)=0.8 (correlation coefficients), and other weights and coefficients are set according to the actual scenario. The calculated R=0.72 (threshold 0.6) indicates an anomaly; further calculation of the CALID / CVN write confidence level is needed. ta=10:00:00, tb=10:00:05, ω(t)=0.9 (matching coefficient), δ(t)=0.7 (deviation), integral result=3.15, Ccheat=0.95 (threshold 0.8), finally determined that there is a writing anomaly, and at the same time queried the blacklist database (no match), and completed the warning within 200ms.

[0078] 4. Data storage and federated learning updates

[0079] The data storage and retrieval module generates blockchain blocks. Each block contains the hash value of the captured image, the judgment result, and a timestamp (10:00:05.123). These blocks are then sent to the Municipal Environmental Protection Bureau and the vehicle manufacturer's nodes for consensus via the consortium blockchain. The block generation time interval is as follows: Calculations: Δt0 = 10s, k = 0.1, N = 5000 (current data volume), Δt = 5.2s, evidence storage time 4.8s. During the federated learning phase, the model is trained locally using 1000 newly added law enforcement data entries (10 epochs), parameter update amount... The parameters are uploaded to the cloud in a size of 20MB. The cloud performs weighted aggregation (the weight is set according to the amount of data on the terminal) and then sends the updated model down. The local model update takes 15 minutes.

[0080] III. Data Characterization for Effectiveness Verification

[0081] Table 1: Performance Comparison of Mobile Law Enforcement Scenarios on Urban Roads

[0082] Evaluation indicators Traditional mobile inspection equipment This invention system Multi-type vehicle recognition rate (National III emission standard - New energy vehicles) 88% 99.5% CALID / CVN write anomaly detection accuracy 82% 98.2% Multispectral image feature extraction time 200ms 85ms Blockchain data storage time - 4.8s Cross-regional data retrieval response time 8s 1.5s

[0083] Traditional mobile detection equipment only carries a single visible light camera, lacking infrared and ultraviolet spectral data. Due to wear and tear on older vehicles meeting only the National III emission standard, and the unique license plate design of new energy vehicles, the recognition rate is only 88%. Anomaly detection relies on single OBD data, making it difficult to identify covert spoofing behavior, with an accuracy rate of only 82%. This invention acquires complete features through multispectral collaborative capture, combined with multi-source data fusion and anomaly correlation analysis, increasing the recognition rate to 99.5% and the anomaly detection accuracy to 98.2%. Heterogeneous computing power scheduling reduces feature extraction time by 57.5%, blockchain notarization ensures the legal validity of the data, and layered storage and 5G communication reduce cross-regional data retrieval time from 8 seconds to 1.5 seconds, fully meeting the efficiency and accuracy requirements of urban mobile law enforcement.

[0084] Example 2: Highway Fixed Monitoring Point Scenario

[0085] This embodiment addresses the fixed monitoring requirements at highway service area entrances. The monitoring point needs to cover four lanes in both directions (lane width 3.75m), be adaptable to extreme temperatures ranging from -20℃ to 50℃ (summer sun exposure, winter low temperatures), support rapid identification of highway vehicles (speed 60-120km / h), and simultaneously undertake the tasks of regional law enforcement data storage and model updates. The specific implementation is as follows:

[0086] I. Hardware Deployment Details

[0087] The image capture and recognition module consists of a multispectral camera array mounted on a 10m high monitoring pole (2m spacing). Each group includes two visible light cameras (Dahua DH-IPC-HFW5849E, 12MP, 12-40mm focal length), one infrared camera (Amap G113 (1280×1024 resolution, temperature range -40℃ to 500℃), and one ultraviolet camera (Avantes AvaSpec-ULS2048L, spectral range 190-1100nm). The lidar is a Hesai Pandar128 (20Hz scanning frequency, 0.3-200m ranging), the millimeter-wave radar is a Bosch LRR4 (0.1-250m ranging, -40 to 250km / h speed range), and the synchronization controller is an Advantech ADAM-6050 (0.1ms synchronization accuracy).

[0088] AI algorithm processing module: It adopts an industrial-grade computing cabinet, with two Horizon Journey 6 NPUs (40 TOPS each) and four Xilinx Alveo U280 FPGAs (10 TOPS each), for a total computing power of 120 TOPS. It is equipped with Ubuntu Server 22.04 and uses TensorRT 8.6 for model inference acceleration.

[0089] Data Storage and Retrieval Module: The blockchain node uses Alibaba Cloud ECS (16 cores, 32G). Hierarchical storage architecture: The hot data layer (data within the last 7 days) uses 10 Samsung 990 Pro 4TB SSDs (RAID5), the warm data layer (data from 1 to 3 months) uses 20 Seagate Exos 20TB HDDs (RAID6), and the cold data layer (data over 3 months) uses an IBM TS4500 tape library (capacity 1PB), supporting 10Gbps fiber optic communication.

[0090] Other Modules: For the clock synchronization unit of the multi-source data fusion module, Microsemi SyncServer S600 (synchronization accuracy 1μs) is selected; for the federated learning cloud server of the dynamic model self-update module, Huawei Cloud ModelArts (8 V100 cards) is used, and the local client is deployed on the edge server at the monitoring point (computing power 64 TOPS); the mobile law enforcement adaptation module is integrated into the terminal in the monitoring point duty room (equipped with a 60Ah battery, supporting 8 hours of offline operation).

[0091] II. Full Process Implementation Steps

[0092] 1. Multi-Spectral Cooperative抓拍

[0093] When the monitoring point detects a certain National VI heavy truck (speed 90 km / h) entering, the lidar scans its three-dimensional contour (vehicle length 12m, vehicle height 3.8m), and the millimeter-wave radar confirms the stable speed. The capture and recognition module adjusts the parameters according to the spatial sampling interval formula : K = 0.5 (tail gas resolution coefficient), vg = 15 m / s (tail gas diffusion speed), f = 60 fps (acquisition frame rate), and calculates d = 0.0056m (spatial sampling interval) to ensure the tail gas spectral imaging resolution; the multi-spectral exposure time is calculated according to the cooperative formula: λ v = 0.35, λ i = 0.35, λ u = 0.3, T = 0.04s, and obtains t v = 0.014s, t i = 0.014s, t u = 0.012s. The visible light camera captures the license plate (Ji A67890), the infrared camera detects the tail gas temperature (60°C), the ultraviolet camera collects the SO2 characteristic spectrum (wavelength 280nm), and the capture is completed within 0.08 seconds.

[0094] 2. Edge Computing and Data Fusion

[0095] The AI ​​algorithm processing module allocates computing power as follows: the NPU (60 TOPS) is responsible for determining the vehicle's appearance and emission stage, while the FPGA (40 TOPS) accelerates spectral data analysis and OBD data parsing. The NPU uses YOLOv8 to extract feature vectors and determines the vehicle as a "China VI heavy-duty truck" (99.8% accuracy); the FPGA parses OBD data (fuel consumption 35L / 100km, exhaust temperature 550℃) in 60ms. After aligning the data, the multi-source data fusion module generates a 2048-dimensional fusion vector. The offloading of edge computing tasks is calculated based on... Calculation: t0=14:30:00, t1=14:30:02, D(t)=1200task / s (terminal task volume), C(t)=900task / s (edge ​​computing power), B(t)=400task / s (remaining terminal computing power), integral result=1840task, unloaded volume accounts for 76.7%, terminal computing power load decreased from 80% to 30%.

[0096] 3. Anomaly Detection and Data Storage

[0097] The vehicle anomaly detection module calculates an anomaly correlation coefficient R=0.58 (threshold 0.6), indicating a high match between OBD data and spectral data, thus determining no anomaly; a query of the blacklist database yields no match. The data storage and retrieval module generates blockchain blocks containing truck identification results, OBD data, and spectral maps. The block generation time interval Δt=4.5s (Δt0=10s, k=0.1, N=8000), achieving consensus through the provincial environmental protection department and highway traffic police nodes, with a storage time of 4.2s. The hot data layer stores captured images and detection results (access frequency ≥10 times / day), the warm data layer stores historical OBD data (access frequency 1-5 times / day), and the cold data layer stores monthly statistical data (access frequency <1 time / day).

[0098] 4. Federated learning model update

[0099] Federated learning starts on the 1st of each month: monitoring points act as clients, using 5000 high-speed vehicle data points for local training (20 epochs), with parameter update frequency... Upload parameters (50MB) to the cloud; the cloud aggregates parameters from 10 monitoring points (weights are allocated according to data volume, with this monitoring point having a weight of 0.12), generates a global model, and then distributes it (80MB). After the local model is updated, the accuracy of identifying China VI trucks increases by 2.3%, and the false detection rate of anomalies decreases by 1.8%.

[0100] III. Data Characterization for Effectiveness Verification

[0101] Table 2. Performance Comparison of Fixed Monitoring Points on Highways:

[0102] Evaluation indicators Traditional fixed monitoring equipment This invention system High-speed vehicle (90-120km / h) recognition rate 90% 99.8% Edge computing task offloading percentage - 76.7% Blockchain data storage success rate - 100% Accuracy improved after federated learning model update 0.5% 2.3% Continuous working time under extreme temperatures 6 hours 12 hours

[0103] Traditional fixed monitoring equipment suffers from low frame rate (20fps) and lack of dynamic parameter adjustment for high-speed vehicles, resulting in blurred images of vehicles traveling at 90-120km / h and a recognition rate of only 90%. Furthermore, the lack of edge computing offloading leads to frequent terminal crashes due to prolonged high load operation. The absence of a hierarchical data storage architecture results in low retrieval efficiency and a lack of model update mechanisms, leading to a decline in accuracy over long-term use. This invention achieves a 99.8% high-speed vehicle recognition rate through high frame rate capture and dynamic parameter adjustment; edge computing offloading accounts for over 70%, significantly reducing terminal load; blockchain evidence storage has a 100% success rate, and the data can serve as cross-departmental enforcement evidence; federated learning improves model accuracy by 2.3%, adapting to technological iterations in high-speed vehicle scenarios; and intelligent thermal management batteries and wide-temperature hardware design double the continuous operating time under extreme temperatures, fully meeting the 24 / 7 uninterrupted monitoring needs of highways.

[0104] Figure 3 The table, based on a heterogeneous computing power allocation formula, quantifies the advantages of the "NPU-led + FPGA-accelerated" architecture. The computing power allocation varies depending on the characteristics of different tasks: license plate recognition relies on deep learning, with the NPU accounting for 70%, leveraging its floating-point operation advantages; exhaust gas spectral analysis requires parallel processing of multi-band data, with the FPGA accounting for 60%, utilizing its parallel computing capabilities; OBD data parsing involves rule-based operations, with the FPGA accounting for 70%, reducing the NPU load. In terms of processing time, the heterogeneous architecture's processing time is all below 100ms, while a single CPU requires 150-350ms, resulting in a 3-4 times efficiency improvement. This design solves the "computing power mismatch" problem of traditional single-processor architectures—avoiding NPU resource idleness in rule-based operations and FPGA performance limitations in deep learning, providing computing power support for real-time emission recognition and ensuring low-latency response in mobile law enforcement scenarios.

[0105] Figure 4 The core of this solution echoes multi-dimensional anomaly correlation analysis technology, revealing the crucial role of feature dimension enhancement in judgment performance. Traditional 1-dimensional judgment (using only OBD data) suffers from limited information, resulting in an accuracy of only 75% and a false negative rate as high as 18%, failing to identify covert cheating such as "normal OBD but abnormal spectrum." While 2-3 dimensions offer improvements, the lack of communication features (such as communication traces from CALID / CVN writes) still leads to a false negative rate exceeding 5%. This solution, through 4-dimensional feature correlation combined with anomaly confidence integral formulas, improves accuracy to 98% and reduces the false negative rate to 1.5%. Combined with blacklist matching, the accuracy reaches 98.5%, fully meeting industry standards. Although the judgment time increases from 30ms to 220ms, it is still far below the "sub-second response" requirement (≤500ms) for mobile law enforcement, achieving a balance between accuracy and timeliness. This effectively solves the dilemma of traditional solutions: either too many false negatives or too slow judgment.

[0106] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A vehicle emission identification system based on artificial intelligence, characterized in that, include: The image capture and recognition module is equipped with a multispectral camera array consisting of three visible light cameras, two infrared thermal imaging cameras, and one ultraviolet spectral camera. Each camera unit is synchronized via a synchronous trigger controller. Based on real-time vehicle 3D contour data acquired by LiDAR scanning, the module dynamically adjusts multispectral imaging parameters. The visible light cameras focus on vehicle exterior and license plate details, the infrared thermal imaging cameras capture the thermal radiation characteristics of the vehicle's powertrain and exhaust gases, and the ultraviolet spectral camera collects spectral information of characteristic pollutants in the exhaust gases. The module performs synchronous capture of multispectral images, supporting license plate recognition for vehicles meeting China VI, China V, China IV, China III, and new energy vehicle standards. In typical sunny scenarios, the recognition rate for typical license plates is ≥99.9%. The module uses onboard radar to acquire real-time vehicle speed and, combined with an exhaust gas diffusion velocity model, adaptively adjusts the image acquisition frame rate and multispectral exposure time. When an abnormal increase in exhaust gas concentration is detected, the exposure time of the ultraviolet spectral camera is automatically extended to enhance the acquisition of pollutant spectral signals. AI algorithm processing module: It has a built-in heterogeneous computing power scheduling unit, which integrates one NPU and two FPGA chips to form a heterogeneous computing power architecture with NPU as the main body and FPGA acceleration; it dynamically allocates computing power resources for license plate recognition, emission spectrum analysis and three-dimensional contour matching; it performs deep convolution and feature extraction on captured images and data to generate a 1024-dimensional vehicle appearance feature vector, a 256-dimensional license plate feature vector and a 512-dimensional exhaust emission spectrum feature vector. Vehicle anomaly detection module: Constructs an abnormal behavior feature chain that includes OBD data consistency, exhaust gas composition matching degree, power system thermal characteristic deviation degree, and control unit communication characteristics, containing a total of 12 core abnormal features and 36 derivative features; Real-time access to OBD online monitoring data via the vehicle CAN bus, combined with exhaust gas composition spectral data and vehicle operating status data extracted by the AI ​​algorithm processing module; Data storage and retrieval module: adopts a fusion architecture of blockchain evidence storage + layered storage; blockchain evidence storage is based on consortium blockchain technology, and law enforcement agencies, car manufacturers, and testing institutions are selected as consensus nodes; In terms of tiered storage, data is divided into hot data layer, warm data layer and cold data layer according to access frequency and importance. Hot data layer uses high-speed SSD storage; warm data layer uses HDD storage; cold data layer uses tape library storage, which supports fast data retrieval, historical review and data sharing across law enforcement agencies in different regions. Mobile law enforcement adaptation module: Equipped with a thermal management battery pack, it adopts a thermal management method that combines liquid cooling and air cooling, integrates the Beidou positioning system and 5G / Edge computing unit, and combines with the law enforcement APP to realize end-edge-cloud collaboration. The on-site law enforcement terminal is responsible for data collection and preliminary processing, the edge node undertakes parallel computing and temporary storage of data, and the cloud law enforcement platform performs big data analysis and global scheduling. Data fusion module: Equipped with a clock synchronization unit, it performs spatiotemporal alignment on the multispectral image data from the capture and recognition module, the OBD and exhaust gas composition data from the vehicle anomaly judgment module, and the feature extraction data from the AI ​​algorithm processing module. It adopts an attention-based fusion algorithm to generate a 2048-dimensional multi-dimensional feature vector for vehicle emission recognition, providing a data foundation for subsequent judgment. Dynamic Model Self-Update Module: Based on the federated learning framework, an iterative process of local training, parameter upload, global aggregation, and model distribution is designed. Each law enforcement terminal acts as a client of federated learning, training the AI ​​recognition model locally using newly added law enforcement data. After training, only the updated part of the model parameters is uploaded to the cloud aggregation server. The cloud server uses a weighted aggregation algorithm to generate a globally updated model and distributes it to each terminal, realizing continuous optimization of the AI ​​recognition model and abnormal behavior feature chain. In the capture and recognition module, the dynamic adjustment of multispectral imaging parameters satisfies the condition that the visible light exposure time t... v Infrared exposure time t i With UV exposure time t u The collaborative relationship is ,in For visible light weighting coefficients, For infrared weighting coefficients, The ultraviolet weighting coefficient is T, and the total exposure time threshold for multispectral imaging is T. Through collaborative relationships, multispectral images can be used to capture vehicle exterior details, thermal radiation characteristics, and exhaust pollutant spectra under different lighting and emission scenarios. The heterogeneous computing power scheduling unit of the AI ​​algorithm processing module allocates computing power to meet the following requirements: , where P NPU To allocate computing power to the NPU, D NPU D represents the computational power density required by the NPU for the current subtask. FPGA P represents the FPGA's computational power demand density for the current subtask, α represents the NPU's computational efficiency coefficient, β represents the FPGA's computational efficiency coefficient, and P... total Total available computing power for AI algorithm processing modules; The multi-dimensional abnormal behavior feature chain of the vehicle anomaly determination module, and the calculation of the anomaly correlation degree R, satisfy the following: Where m is the number of abnormal features involved in the association, and wjk is the association weight between the j-th feature and the k-th feature. Let xj be the Pearson correlation coefficient between the j-th eigenvector xj and the k-th eigenvector xk; During the blockchain notarization process of the data storage and retrieval module, the data block generation time interval is... satisfy ,in The initial block generation interval is defined as k, which is the block generation adjustment coefficient, and N is the amount of data currently stored. The block generation interval is dynamically shortened as the amount of stored data increases. In the end-edge-cloud collaboration of the mobile law enforcement adaptation module, the edge computing task offloading amount U satisfies Where D(t) is the terminal computing task at time t, C(t) is the edge node computing power at time t, B(t) is the terminal remaining computing power at time t, t0 is the task unloading start time, and t1 is the task unloading end time. In the federated learning parameter update of the dynamic model self-updating module, the local model parameter update amount satisfy ,in For learning rate, Let be the gradient of the local data of the i-th terminal with respect to the model parameters θ. The momentum coefficient, This represents the parameter update amount from the last time.

2. A method for applying the artificial intelligence-based vehicle emission identification system as described in claim 1, characterized in that, include: Multispectral collaborative capture steps: The capture and recognition multispectral camera array acquires the vehicle's three-dimensional contour data in real time through LiDAR scanning. Combined with the vehicle's driving speed acquired by the onboard radar, the exposure time and acquisition frame rate of the visible light, infrared, and ultraviolet imaging units are dynamically adjusted. The visible light camera focuses on the front and rear of the vehicle, the infrared thermal imaging camera is aimed at the vehicle's engine compartment and exhaust emission area, and the ultraviolet spectral camera is pointed at the exhaust gas diffusion path to complete the synchronous capture of multispectral images, while simultaneously acquiring the ultraviolet spectral information of the vehicle's exhaust emissions. Heterogeneous computing power scheduling and feature extraction steps: The heterogeneous computing power scheduling unit, which processes AI algorithms, dynamically allocates computing power resources of NPU and FPGA according to the current task type; after preprocessing the multispectral captured images, a deep convolutional neural network is used to extract vehicle appearance feature vectors, license plate feature vectors, and exhaust emission spectral feature vectors. Data fusion and preliminary identification steps: The feature vectors of the captured images, the OBD online monitoring data retrieved by the vehicle anomaly judgment module, and the vehicle operating status data are spatiotemporally aligned by the clock synchronization unit. An attention mechanism fusion algorithm is used to generate a multi-dimensional fusion feature vector with a dimension of 2048. Based on the vector, the AI ​​algorithm processing module uses a pre-trained classification model to achieve a preliminary determination of the vehicle emission stage, with an accuracy rate of ≥96%. Anomaly correlation analysis steps: Vehicle anomaly determination retrieves OBD online monitoring data, exhaust gas composition spectrum data, and vehicle operating status data to construct an abnormal behavior feature chain and calculate the anomaly correlation degree; simultaneously, it writes abnormal behavior to CALID / CVN and calculates the anomaly confidence degree; when the anomaly correlation degree or anomaly confidence degree exceeds the preset threshold, it is determined that the vehicle has abnormal behavior, and at the same time, it completes the identification and warning of blacklisted vehicles. Blockchain evidence storage and hierarchical storage steps: Data storage and retrieval generate blockchain blocks to store captured images, recognition results, and anomaly detection data; at the same time, data is classified and stored in hot data layer, warm data layer, or cold data layer according to the data access frequency and importance. Edge-cloud collaborative law enforcement steps: Mobile law enforcement adaptation dynamically adjusts the system's operating power based on remaining battery power and computing load; some data processing tasks are offloaded to edge nodes through 5G / Edge computing units; Beidou positioning data and law enforcement APP are integrated to achieve real-time synchronization between on-site law enforcement data and the cloud law enforcement platform, and the cloud platform can remotely retrieve historical data and global law enforcement statistics; Federated learning model update steps: The dynamic model self-updates to initiate the federated learning process. Each law enforcement terminal updates the parameters of the AI ​​recognition model locally using the newly added law enforcement data. Only the updated model parameters are uploaded to the cloud aggregation server. The cloud uses a weighted aggregation algorithm to generate a globally updated model and then distributes it to each terminal to achieve continuous model optimization.

3. The method for a vehicle emission identification system based on artificial intelligence according to claim 2, characterized in that, In the multispectral collaborative capture step, the dynamic adjustment of multispectral imaging parameters also satisfies the following: when the vehicle exhaust emission diffusion velocity is vg, the spatial sampling interval d of the multispectral camera array satisfies , where K is the exhaust gas feature spatial resolution coefficient, and f is the image acquisition frame rate.

4. The method for a vehicle emission identification system based on artificial intelligence according to claim 2, characterized in that, In the anomaly correlation analysis step, for the identification of CALID / CVN write anomalies, the anomaly confidence level C is... cheat satisfy ,in Let be the matching coefficient between the brushing behavior features and the standard features at time t. Let t be the deviation between the OBD data at time t and the vehicle's factory reference data. a t represents the start time for abnormal behavior monitoring. b This is the end time for abnormal behavior monitoring.

Images