Campus low-carbon behavior recognition recording method based on AI image recognition

Abstract

The invention belongs to the technical field of artificial intelligence and block chain fusion application, and provides an AI image recognition-based campus low-carbon behavior recognition recording method, which comprises the following steps of: responding to a low-carbon behavior authentication request initiated by a user through a carbon popularity platform, and obtaining multi-modal verification data containing a field image; the pre-trained AI identification model is combined with Internet of Things sensor data and campus management system data to carry out cross validation, a verified low-carbon behavior record is generated, the record is quantified into standard carbon emission reduction data according to a carbon popularity methodology model, and a corresponding carbon point is issued to a user account. The system uploads carbon emission reduction data and image target information to a block chain to form a tamper-resistant evidence storage block, realizes logic binding with a carbon asset project package through an intelligent contract, and supports carbon asset aggregation declaration and traceability query. According to the invention, the identification precision, the anti-counterfeiting capability and the supervision compliance can be improved, and the integrated management of campus low-carbon behaviors is promoted.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and blockchain integration application technology, and in particular to a method for recording low-carbon behavior on campus based on AI image recognition. Background Technology

[0002] With the advancement of dual-carbon goals, universities, as densely populated social units, possess enormous potential in energy conservation and emission reduction. However, in actual campus management, the promotion of green behaviors largely relies on publicity and education, as well as manual supervision. It is difficult to accurately capture each student's low-carbon behavior, scientifically calculate the emission reduction effects of these behaviors, and sustainably incentivize them. This results in low enthusiasm among students to participate in low-carbon actions, low reliability of collected behavioral data, and difficulty in transforming scattered individual emission reduction behaviors into large-scale carbon assets.

[0003] Current carbon credit systems on the market still have many problems when used in scenarios like school campuses. For example, when collecting behavioral data, most rely solely on single methods like swiping cards or scanning codes, which cannot fully reflect students' actual behavior in different scenarios such as cafeterias, dormitories, and teaching buildings. Moreover, the verification process is rather subjective and lacks solid evidence, easily leading to disputes. In addition, the methods for calculating carbon emission reductions are also rather crude, failing to consider whether the behavior actually occurred or whether the on-site environment might affect the accuracy of the data. All of these factors make it difficult to meet regulatory requirements and gain market recognition when applying for carbon assets.

[0004] In recent years, artificial intelligence, especially computer vision technology, has developed rapidly, providing new ideas for automatic behavior recognition. However, its direct application in low-carbon campus scenarios still encounters many challenges. Campus lighting conditions vary greatly, such as the strong light in the cafeteria at noon and the dim light in the corridors in the evening. Furthermore, students may obstruct each other's view, and their movements vary in speed, leading to image blurring. All of these factors make image recognition results unstable. Moreover, relying solely on a single photo or video clip makes it easy for forgery and duplicate submissions to occur; verification must be combined with other data. More importantly, there is currently no mature method for transforming students' casual actions, such as garbage sorting and walking to and from get out of class, into carbon assets that meet national or even international standards and can be claimed. There are also privacy concerns; storing student behavior images on a central server could easily lead to information leaks.

[0005] Therefore, a technological system is needed that integrates AI image recognition, multi-data fusion, blockchain notarization, and carbon asset conversion. For common low-carbon behaviors on campus, such as finishing meals in the cafeteria, garbage sorting, using school buses, and trading secondhand goods, equipment capable of rapid data processing should be deployed in key areas. Simultaneously, a data recording system that is auditable, traceable, and tamper-proof should be established. Previous research has used AI for behavior recognition or blockchain to store environmental data, but most of these only addressed single-stage issues, failing to establish a complete process from behavior occurrence to carbon asset conversion. They also neglected details such as data uncertainty and abnormal behavior detection, making it difficult to meet the high data quality requirements of the carbon market.

[0006] Developing a low-carbon behavior recognition and recording method that is suitable for campus scenarios, accurate in identification, reliable in data, and flexible in expansion has become the key to promoting the digital transformation of green campuses. Summary of the Invention

[0007] This invention provides a method for identifying and recording low-carbon behaviors on campus based on AI image recognition. It aims to solve the technical problems existing in the current methods for identifying and recording low-carbon behaviors on campus, such as insufficient stability of behavior recognition, low data credibility, broken links in carbon assetization, and lack of privacy protection. By constructing a process of multimodal perception, intelligent verification, blockchain notarization, and carbon assetization, it achieves accurate identification, reliable notarization, and standardized asset transformation of low-carbon behaviors on campus, providing technical support for the green transformation of universities.

[0008] This invention provides a method for identifying and recording low-carbon behaviors on campus based on AI image recognition, comprising the following steps: S1: In response to a user's request for low-carbon behavior certification initiated through the carbon benefit platform, obtain multimodal verification data, including on-site images of the low-carbon behavior to be certified; S2: Use a pre-trained AI recognition model to identify multimodal verification data, and combine it with at least one of the IoT sensor data and campus management system data associated with the low-carbon behavior to be certified for cross-validation to generate verified low-carbon behavior records. S3: Based on the preset carbon inclusive methodology model, the verified low-carbon behavior records are quantified into standard carbon emission reduction data, and corresponding carbon credits are issued to users' carbon accounts; the carbon inclusive methodology model is constructed based on the preset low-carbon behavior emission reduction factor database and / or the behavior-driven carbon emission reduction quantitative assessment model trained based on historical data; the emission reduction factor database includes standard parameters for unit emission reduction of different low-carbon behavior categories; S4: Upload the target information of standard carbon emission reduction data and on-site images of low-carbon behaviors to be certified as evidence to the blockchain network to generate an immutable evidence block; the target information includes a verifiable digital digest, a searchable storage identifier, a timestamp, and a user anonymity identifier; S5: On the blockchain network, the identifier of the evidence storage block is logically bound to one or more carbon asset project packages that meet the carbon inclusive certification emission reduction declaration requirements through smart contracts; among them, the carbon asset project package is formed by the carbon inclusive platform operator by aggregating the records of similar low-carbon behaviors of multiple users and the standard carbon emission reduction data generated therefrom and then submitting them in a unified manner. S6: Provides a traceable query interface for low-carbon behavior records associated with a user's carbon account. This traceable query interface is used to display the evidence block information corresponding to the low-carbon behavior record through the blockchain network.

[0009] Furthermore, the method is implemented through a carbon inclusion platform, which is a comprehensive technology platform built for campus scenarios, integrating data collection, intelligent verification, and carbon asset management. Its system architecture, from bottom to top, includes: The data acquisition layer is used to connect to and standardize access to various low-carbon behavior-related data sources on campus, including intelligent waste sorting and recycling equipment, library management system, mobile health application step counting interface and image acquisition terminal. Among them, the image acquisition terminal is deployed in canteen, teaching building and dormitory, and is specifically used to capture on-site images of low-carbon behaviors. The data layer is used for unified cleaning, structured governance, persistent storage, and in-depth analysis of various low-carbon behavior-related data sources; The support layer is used to integrate IoT platforms to enable device connectivity and status monitoring, as well as to integrate AI platforms to support the deployment and inference optimization of pre-trained AI recognition models; The application layer includes a personal carbon ledger module, a low-carbon behavior verification module, and a low-carbon learning knowledge base. The personal carbon ledger module records the accumulation and usage of carbon credits in the user's carbon account. The low-carbon behavior verification module is used to realize the low-carbon behavior certification request and response and the multimodal verification data flow interaction. The low-carbon learning knowledge base is used to provide personalized green and low-carbon education content recommendations and learning outcome certification.

[0010] Furthermore, the image acquisition terminal is equipped with a lightweight model. This lightweight model is based on the cloud-based full recognition model deployed on a cloud server and is obtained after processing through one or more of the following methods: knowledge distillation, pruning, or quantization. It is deployed using a containerized encapsulation and edge collaborative scheduling architecture, supporting incremental updates. Specifically, knowledge distillation uses the cloud-based full recognition model as the teacher model to guide the lightweight student model in learning its output layer soft label distribution and intermediate layer feature representation. Pruning evaluates the importance of redundant convolutional kernels or channels in the cloud-based full recognition model and filters and removes low-contribution weights based on preset indicators, including L1 norm, gradient sensitivity, or Taylor expansion. Quantization converts the floating-point parameters in the cloud-based full recognition model into low-bit integers.

[0011] Furthermore, the low-carbon behaviors to be certified include at least one of the following: the "Clean Plate Campaign," proper sorting and disposal of recyclables, electricity conservation, water conservation, low-carbon walking, school bus travel, and second-hand goods trading.

[0012] Furthermore, in step S2, the AI ​​recognition model adopts a multi-task learning architecture to simultaneously complete the category recognition of the low-carbon behavior to be certified, the detection of the number of participating entities, the estimation of duration, and the labeling of environmental interference factors. Among them, environmental interference factors include changes in illumination, degree of occlusion, motion blur, and semantic similarity interference. An attention mechanism is introduced to adaptively weight and focus on key visual regions. Key visual regions are visual feature regions that are highly correlated with the category of low-carbon behavior to be certified. The output of the AI ​​recognition model includes multiple branch output heads, which respectively generate the category probability distribution, the object detection bounding box sequence, the temporal action integrity score, and the confidence interval. The multi-task learning architecture shares backbone network parameters during the AI ​​recognition model training phase and combines gradient balancing strategies to balance the convergence speed differences among sub-tasks.

[0013] Furthermore, after step S3, the method further includes: organizing the standard carbon emission reduction data that meets the preset conditions into a carbon asset declaration material package for review by the carbon inclusive platform; the preset conditions include and simultaneously meet the following: the cumulative carbon emission reduction value of a single verified low-carbon behavior record reaches the minimum carbon asset declaration unit threshold, the behavior of the verified low-carbon behavior record occurs within the carbon asset project cycle, and the completeness of the IoT sensor data associated with the low-carbon behavior in the verified low-carbon behavior record meets the carbon asset regulatory requirements.

[0014] Furthermore, in step S5, Logical binding is achieved through smart contracts deployed on a blockchain network built and maintained by the Carbon Inclusive platform; When a carbon asset project package is approved by the regulatory node of the carbon benefit platform and a unique serial number is issued to it, a smart contract pre-deployed in the blockchain network is automatically triggered to associate the serial number with the block hash or globally unique identifier of all source behavior evidence blocks that constitute the carbon asset project package. This creates a traceable and tamper-proof mapping relationship between the serial number and each source behavior evidence block. The source behavior evidence block is the evidence block corresponding to the low-carbon behavior of each user that constitutes the carbon asset project package.

[0015] Furthermore, in step S3, based on the preset carbon inclusiveness methodology model, the verified low-carbon behavior records are quantified into standard carbon emission reduction data, introducing uncertainty quantification, specifically including: Using the comprehensive confidence level corresponding to the verified low-carbon behavior records as prior information, and combined with standard carbon emission reduction data, the measurement error range of each input parameter is calculated as a characterization of observation uncertainty, and a Bayesian inference framework is constructed. Within this Bayesian inference framework, the posterior probability distribution of standard carbon emission reduction data is modeled using Markov chain Monte Carlo or variational inference methods to obtain the uncertainty distribution; Based on the uncertainty distribution, the expected value of the standard carbon emission reduction data is extracted as the central value of carbon emission reduction, and the uncertainty interval is determined by the coverage interval under a pre-set confidence level. The uncertainty interval is used as a component of the evidence information in step S4, and an error propagation model based on covariance propagation is used to mathematically synthesize the uncertainty intervals of the components of the carbon asset project package in step S5. The joint uncertainty distribution of the total carbon emission reduction of the carbon asset project package is calculated, and the aggregated total emission reduction center value and its corresponding total uncertainty interval are derived from it.

[0016] Furthermore, it also includes enhanced verification processing for abnormal low-carbon behavior authentication requests initiated by the carbon benefit platform. The specific process is as follows: Real-time monitoring of the frequency of low-carbon behavior authentication requests initiated by the same user, the same device, or the same geographical area within a unit time window, wherein the unit time window is a continuous 60-minute time interval; If the frequency exceeds the frequency threshold dynamically calculated based on historical baselines, and the confidence interval generated by the output of the AI ​​recognition model exceeds the preset threshold interval, the low-carbon behavior certification request will be marked as a potential anomaly, and enhanced verification will be initiated. Only after the enhanced verification is passed will the low-carbon behavior certification request be marked as normal and allowed to proceed to the next steps.

[0017] Furthermore, enhanced verification includes at least one of the following verification mechanisms: When the acquisition angle of the on-site images of the low-carbon behavior to be certified is insufficient to fully represent the low-carbon behavior to be certified, an interactive instruction is generated to request the user to provide supplementary video stream data of the on-site low-carbon behavior to be certified from different spatial perspectives. When there are time discontinuities or data sparsity issues in the environmental perception information in IoT sensor data, a data retrieval request is sent to the IoT sensor to obtain the relevant sensor data sequence covering a longer time window for time sequence continuity verification. When the location information in the campus management system data is not well matched or there is spatial overlap interference from multiple entities, a location cross-verification based on geofencing and GNSS / Bluetooth beacon fusion positioning is initiated.

[0018] Compared with the prior art, the present invention has the following advantages and beneficial effects: (i) Design a multi-task learning AI model and a multi-modal cross-validation mechanism. By acquiring multi-modal verification data of on-site images of low-carbon behaviors to be certified, and combining IoT sensor data and campus management system data for cross-validation, the impact of single data error is reduced and the reliability of AI recognition model output results is improved.

[0019] (ii) Uploading standard carbon emission reduction data and target information to the blockchain generates an immutable evidence block, ensuring the transparency and traceability of every emission reduction data; at the same time, the evidence block identifier is logically bound to the carbon asset project package through smart contracts, and the platform operator aggregates similar behavior records for unified declaration, so that the originally scattered individual small emission reductions can be aggregated into carbon asset projects that meet the verification requirements, realizing the effective connection from individual behavior records to large-scale carbon assets.

[0020] (iii) By modeling the uncertainty of standard carbon emission reduction data through methods such as Bayesian inference framework, the uncertainty interval is used as part of the evidence information, making the carbon emission reduction data more scientific and rigorous and in line with the normative requirements of carbon asset certification. At the same time, when storing evidence on the blockchain, only verifiable digital digests of the on-site images of low-carbon behavior to be certified and user anonymity identifiers are uploaded, rather than the original images or the user's real identity information. This achieves data traceability while maximizing the protection of user privacy and avoiding the problem of privacy protection deficiency.

[0021] (iv) It has built a fully functional carbon inclusive platform system architecture, providing comprehensive technical support for the identification and recording of low-carbon behaviors on campus. It has realized integrated functions from multi-source data access and governance analysis to AI model deployment, IoT device monitoring, personal carbon ledger management, behavior verification and low-carbon education.

[0022] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0023] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0024] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram illustrating the steps of a campus low-carbon behavior recognition and recording method based on AI image recognition. Figure 2 This is a schematic diagram of the carbon benefit platform system architecture; Figure 3 This diagram illustrates the steps involved in enhancing the verification process for abnormal low-carbon behavior authentication requests initiated by the Carbon Inclusive Platform. Detailed Implementation

[0025] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0026] This invention provides a method for recording and identifying low-carbon behaviors on campus based on AI image recognition, such as... Figure 1 As shown, it includes the following steps: S1: In response to a user's request for low-carbon behavior certification initiated through the carbon benefit platform, obtain multimodal verification data including on-site images of the low-carbon behavior to be certified; S2: Use a pre-trained AI recognition model to identify multimodal verification data, and combine it with at least one of the IoT sensor data and campus management system data associated with the low-carbon behavior to be certified for cross-validation to generate verified low-carbon behavior records. S3: Based on the preset carbon inclusive methodology model, the verified low-carbon behavior records are quantified into standard carbon emission reduction data, and corresponding carbon credits are issued to users' carbon accounts; the carbon inclusive methodology model is constructed based on the preset low-carbon behavior emission reduction factor database and / or the behavior-driven carbon emission reduction quantitative assessment model trained based on historical data; the emission reduction factor database includes standard parameters for unit emission reduction of different low-carbon behavior categories; S4: Upload the target information of standard carbon emission reduction data and on-site images of low-carbon behaviors to be certified as evidence to the blockchain network to generate an immutable evidence block; the target information includes a verifiable digital digest, a searchable storage identifier, a timestamp, and a user anonymity identifier; S5: On the blockchain network, the identifier of the evidence storage block is logically bound to one or more carbon asset project packages that meet the carbon inclusive certification emission reduction declaration requirements through smart contracts; among them, the carbon asset project package is formed by the carbon inclusive platform operator by aggregating the records of similar low-carbon behaviors of multiple users and the standard carbon emission reduction data generated therefrom and then submitting them in a unified manner. S6: Provides a traceable query interface for low-carbon behavior records associated with a user's carbon account. This traceable query interface is used to display the evidence block information corresponding to the low-carbon behavior record through the blockchain network.

[0027] The working principle of the above technical solution is as follows: After a user initiates low-carbon behavior certification through the campus carbon credit platform, the system will first enter the S1 stage to collect multimodal verification data, including on-site images. This data is not only photos or videos of behavior taken by mobile phones or fixed cameras, but also integrates auxiliary data such as geographical location, equipment information, and time at the time of shooting to form a preliminary chain of behavioral evidence, ensuring the authenticity and traceability of the original data from the source. In the S2 stage, the system calls a pre-trained AI model to analyze the collected images and identify the specific low-carbon behavior, such as garbage sorting or walking into the school. This AI model has been trained with a large amount of campus scene data and can accurately identify behaviors even in complex environments. It can also distinguish different types of subdivided behaviors. However, relying solely on image data is prone to errors, so cross-validation is added: the AI ​​recognition results are compared with IoT sensor data (such as geomagnetic sensors that detect bicycle parking and smart meters that record electricity consumption) and campus management system data (such as campus card swipe records and academic scheduling information) to ensure that the time, location, and logic are consistent. For example, if a student applies for garbage sorting, the system will simultaneously check whether the garbage bin's sensors have recorded weight changes and whether the access control system shows that the student entered the disposal area during that period, confirming from multiple dimensions that the behavior has actually occurred. During cross-validation, all data sources will include a microsecond-level timestamp synchronized by the platform, and location information will be uniformly converted to the coordinate system of the campus GIS map to avoid spatiotemporal discrepancies. In the event of data conflicts, physical sensor data will be given priority because it is not easy to falsify. For example, if the image shows that the light is off, but the current of the smart meter has not dropped to the standby state, the meter data will be used to determine that the energy-saving behavior was unsuccessful. The specific arbitration rules can be found in Table 1. Table 1 Arbitration Rules for AI Image Recognition Results Phase S3 primarily converts low-carbon behaviors into standard carbon emission reductions and awards carbon credits to users. The system has a pre-set carbon benefit methodology model and an authoritative database of emission reduction factors, such as standard parameters like 0.5 kg CO2e reduction per kilogram of recyclables and 0.18 kg CO2e reduction per kilometer of cycling. At the same time, the model is continuously optimized based on historical behavioral data on campus, such as adjusting emission reduction coefficients for the travel habits of students in different grades to make the calculation results more realistic. The system extracts information such as the type, frequency, and intensity of verified behaviors, inputs it into the model to calculate accurate carbon emission reductions, and then converts them into carbon credits and sends them to users' personal carbon accounts to incentivize participation. The focus of Phase S4 is ensuring data security and preventing tampering. The system will package the calculated carbon emission reductions, along with key information from on-site images (such as the SHA-256 hash value of the image to verify its integrity; a unique identifier from distributed storage for quick retrieval; a trusted timestamp to pinpoint the time of the action; and an anonymized user identifier to protect privacy while linking it to a specific action), and upload it to the blockchain network. Once on the chain, this information cannot be modified, forming legally valid electronic evidence. Phase S5 transforms scattered individual emission reductions into claimable carbon assets. The system deploys smart contracts on the blockchain to aggregate similar low-carbon behavior records from multiple users, forming a unified carbon asset project package. This package will meet national or local carbon credit certification requirements, including necessary elements such as total emission calculation, additionality justification, and monitoring reports. The smart contracts automatically handle data aggregation and compliance checks, and can also connect to the application interface, reducing manual operations and improving the efficiency of carbon asset development. Finally, in the S6 stage, each user can view the blockchain-based evidence information corresponding to their low-carbon behavior through the traceability query interface of their carbon account. Whether it is the user or the regulator, by entering the behavior record ID, they can see all the information such as the time of on-chain recording, storage location, and status of the associated carbon asset project package. From points to original evidence, everything can be traced transparently, which not only gives everyone peace of mind but also provides reliable data for subsequent carbon trading, auditing, and policy making.

[0028] The beneficial effects of the above technical solution are as follows: By adopting the solution provided in this embodiment, through a five-layer architecture of AI recognition, multi-source verification, scientific quantification, blockchain notarization and smart contract binding, the entire process of campus low-carbon behavior from perception to assetization is automated, trustworthy and standardized. It solves the key pain points in the traditional carbon inclusiveness model, such as weak evidence, easy to forge, difficult to trace and delayed incentives, and builds a low-carbon governance system that combines accuracy, security and sustainability.

[0029] In one embodiment, such as Figure 2 As shown, the method is executed through a carbon inclusion platform, which is a comprehensive technology platform built for campus scenarios, integrating data collection, intelligent verification, and carbon asset operation. Its system architecture, from bottom to top, includes: The data acquisition layer is used to connect to and standardize access to various low-carbon behavior-related data sources on campus, including intelligent waste sorting and recycling equipment, library management system, mobile health application step counting interface and image acquisition terminal. Among them, the image acquisition terminal is deployed in canteen, teaching building and dormitory, and is specifically used to capture on-site images of low-carbon behaviors. The data layer is used for unified cleaning, structured governance, persistent storage, and in-depth analysis of various low-carbon behavior-related data sources; The support layer is used to integrate IoT platforms to enable device connectivity and status monitoring, as well as to integrate AI platforms to support the deployment and inference optimization of pre-trained AI recognition models; The application layer includes a personal carbon ledger module, a low-carbon behavior verification module, and a low-carbon classroom knowledge base. The personal carbon ledger module is used to record the accumulation and usage of carbon credits in the user's carbon account. The low-carbon behavior verification module is used to realize the low-carbon behavior certification request and response and the multimodal verification data flow interaction. The low-carbon classroom knowledge base is used to provide personalized green and low-carbon education content recommendations and learning outcome certification. The acquisition layer and data layer use the MQTT (Message Queuing Telemetry Transport) message transmission protocol based on TCP / IP. The data layer and support layer connect to a relational database through the JDBC (Java Database Connectivity) interface, and use the SQL (Structured Query Language) standard for structured data querying and writing. The support layer and application layer use the GraphQL interface query language, which allows the application layer to flexibly define data query fields according to the needs of different modules, reducing redundant data transmission. All interfaces between layers use the OAuth 2.0 authorization framework to control interface access permissions, verify the legitimacy of requests through Access Tokens, and encrypt transmitted data using TLS / SSL (Transport Layer Security / Secure Sockets Layer) to ensure the confidentiality and integrity of data during transmission.

[0030] The working principle of the above technical solution is as follows: The acquisition layer is mainly responsible for connecting to various data sources related to low-carbon behaviors on campus, and unifying the data from different devices and systems; for example, intelligent waste sorting equipment will upload the time, frequency, and type of waste sorted by users; the library borrowing system will extract electronic borrowing records to identify green reading behaviors; the mobile health APP will synchronize users' walking and cycling step data through an interface; image acquisition terminals will be deployed in key areas such as canteens, teaching buildings, and dormitories to capture on-site images of low-carbon behaviors such as the "Clean Plate Campaign," energy-saving electricity use, and green commuting. All raw data will undergo unified protocol conversion and format processing, becoming structured messages transmitted to the data layer to ensure data integrity and semantic consistency during subsequent processing; After receiving various data from the collection layer, the first step of the data layer is to clean the data, removing duplicate, missing, or abnormal data. For example, garbage sorting events are organized into a unified format that includes device ID, timestamp, category, and weight; book borrowing data is transformed into fields such as user ID, book title, borrowing duration, and whether the book has been renewed; step count data is summarized into daily low-carbon travel mileage; and image data is also labeled with shooting location, time, and scene. The cleaned and organized data is stored in a distributed database and data lake, supporting simultaneous reading and writing by a large number of users and allowing for the tracking of historical data. In addition, the data layer has a dedicated analysis engine that deeply mines this behavioral data, builds a low-carbon behavior profile for each user, identifies behavioral patterns and trends, and provides data support for subsequent behavior verification and content recommendation. The support layer primarily relies on an IoT platform to manage front-end data acquisition devices, including device registration, real-time monitoring of operational status, remote parameter configuration, and fault alarms, ensuring uninterrupted data acquisition and stable system operation. Simultaneously, the support layer integrates an artificial intelligence platform, deploying a pre-trained deep learning image recognition model specifically designed to analyze images transmitted from the image acquisition terminals. This model automatically identifies low-carbon behaviors such as empty plates, bringing one's own cutlery, and shared meals, and provides confidence levels for the identification. This AI model has been specifically optimized for campus scenarios, enabling accurate identification even in situations with complex lighting, obstructions, or multiple targets appearing simultaneously. Furthermore, inference combines edge computing and cloud collaboration, improving response speed and reducing bandwidth consumption, ensuring verification results are fed back to users in real time. The application layer has three core functional modules, all implemented based on the data and technical capabilities provided by the layers below. The Personal Carbon Ledger module creates a unique carbon account for each user, converting various low-carbon behaviors into carbon credits according to preset accounting rules. It records the entire process of credit accumulation, redemption, and clearing, and also visualizes individual emission reduction contributions and ranking changes, giving users a sense of participation and accomplishment. The Low-Carbon Behavior Verification module receives users' carbon credit claim requests, retrieves device logs, system records, and on-site images for the corresponding time period, and compares them with AI recognition results and business data to determine if the behavior actually occurred, preventing false declarations. After verification, an immutable digital certificate is generated and stored on the blockchain, ensuring transparency and credibility. The Low-Carbon Learning Knowledge Base pushes customized green courses based on user behavior profiles. For example, users who frequently borrow environmental books will be recommended lectures on the forefront of carbon neutrality; users who frequently participate in the "Clean Plate Campaign" will be recommended educational materials on food waste. After completing online quizzes and tasks, users can receive learning achievement certification, which in turn rewards them with carbon credits.

[0031] The beneficial effects of the above technical solution are as follows: by adopting the solution provided in this embodiment, the acquisition layer completes the data mapping of low-carbon behavior in the physical world to the digital space, the data layer realizes the preparation of data assetization, the support layer provides system-level service capabilities through IoT and intelligent connectivity technologies, and the application layer ultimately transforms the technical capabilities into interactive, operable, and incentivizing user value, thus building a fully automated carbon inclusive mechanism from behavior capture to value realization, which can promote the large-scale and sustainable development of green lifestyles on campus.

[0032] In one embodiment, the image acquisition terminal is configured with a lightweight model. The lightweight model is based on the cloud-based full recognition model deployed on a cloud server and is obtained after processing by one or more of the following methods: knowledge distillation, pruning, or quantization. It is deployed using a containerized encapsulation and edge collaborative scheduling architecture and supports incremental updates. Specifically, knowledge distillation uses the cloud-based full recognition model as the teacher model to guide the lightweight student model to learn its output layer soft label distribution and intermediate layer feature representation. Pruning evaluates the importance of redundant convolutional kernels or channels in the cloud-based full recognition model and filters and removes low-contribution weights based on preset indicators, including L1 norm, gradient sensitivity, or Taylor expansion. Quantization converts the floating-point parameters in the cloud-based full recognition model into low-bit integers.

[0033] The working principle of the above technical solution is as follows: The image acquisition terminal has a built-in lightweight model, which is transformed from the full recognition model deployed in the cloud through compression techniques such as knowledge distillation, pruning and quantization, realizing the transformation from a complex and computationally demanding model to a low-latency, low-power and small-size model suitable for edge devices. In the knowledge distillation process, the cloud-based full recognition model serves as the teacher model, using the soft label probability distribution of its output layer (i.e., the non-normalized prediction confidence of each category) and the feature maps of the intermediate hidden layers as learning objectives. The student model (i.e., the lightweight model) obtains a generalization ability close to that of the teacher model by minimizing the differences between itself and the teacher model on these soft objectives (e.g., using the KL divergence loss function), thereby inheriting its semantic expressive ability while maintaining a smaller structure. The pruning operation performs structured or unstructured sparsity processing on the convolutional layer parameters in the full-scale cloud recognition model. First, the importance score of each convolutional kernel or channel is calculated. Commonly used preset indicators include L1 norm (measures the sum of the absolute values ​​of the weights, reflecting the activation strength), gradient sensitivity (assessing the degree of influence of parameter changes on the loss function), and Taylor expansion approximation (estimates the increase in loss after removal based on the product of the first-order gradient and the weights). Based on the scores, low-contribution channels or filters below the threshold are removed, thereby reducing the number of model parameters and computational cost, and improving inference efficiency. The quantization process converts the model weights and activation values, which were originally stored and operated in 32-bit floating-point numbers (FP32), into 8-bit or even lower-bit integer values ​​(such as INT8). By establishing a linear mapping relationship between the floating-point range and the integer range, memory usage is significantly reduced and the inference process is accelerated while ensuring controllable precision loss. It is also compatible with edge hardware that supports low-precision computing (such as NPU and DSP), further improving energy efficiency. The lightweight model adopts containerized encapsulation (such as Docker) to achieve environment isolation, unified dependencies, and consistent cross-platform deployment. Combined with an edge collaborative scheduling architecture (such as Kubernetes + KubeEdge), it can uniformly manage model versions, monitor running status, and dynamically allocate computing resources across multiple image acquisition terminals. It also supports an incremental update mechanism, pushing only the differences in the model for remote upgrades, reducing bandwidth consumption and terminal downtime, and ensuring continuous and efficient system operation.

[0034] The beneficial effects of the above technical solution are as follows: by adopting the solution provided in this embodiment, the large model capability of the cloud is effectively migrated to the edge, and the actual deployment requirements such as real-time performance and resource constraints are met while ensuring recognition accuracy.

[0035] In one embodiment, the low-carbon behaviors to be certified include at least one of the following: the "Clean Plate Campaign," proper sorting and disposal of recyclables, electricity conservation, water conservation, low-carbon walking, school bus travel, and second-hand goods trading.

[0036] The working principle of the above technical solution is as follows: For several common low-carbon behaviors on campus, the system will authenticate and record them in different ways, as follows: The system analyzes images of tableware uploaded by users after meals using image recognition technology to determine whether there is food residue on the plate. Combining timestamps and geographical location information, it verifies whether the behavior occurred during the mealtime and at the restaurant. Using a food residue detection algorithm trained with a deep learning model (such as a convolutional neural network CNN), it quantifies the proportion of leftover food. When the proportion of leftover food is lower than a preset threshold, it is determined to be a valid clean plate behavior and certified. When users dispose of recyclables, they need to take pictures of the trash can during or after disposal. The system uses object detection algorithms (such as YOLO or Faster R-CNN) to identify the category of items in the image, determine whether they belong to recyclables (such as paper, plastic, metal, etc.), and further analyze whether they have been put into the correct sorting container. At the same time, it combines the sensor data of the smart trash can (such as weight changes, opening action) for cross-verification to ensure that the behavior actually occurs, prevent fake images, and improve the accuracy of authentication for correct sorting and disposal of recyclables. The system connects to the data interface of smart home meters or power monitoring equipment to collect the electricity load curve of homes or offices in real time. By comparing the historical electricity consumption benchmark (such as the average electricity consumption of the past 7 days) with the current electricity consumption, if the current electricity consumption is significantly reduced (for example, a decrease of more than 20%) and continues for a certain period of time, it is judged as an electricity-saving behavior. At the same time, abnormal power outages are excluded, and the rationality is verified in combination with the user's work and rest patterns to avoid misjudgment. Relying on smart water meters or water-using devices with flow monitoring functions (such as smart shower heads and water-saving faucets), the system records users' daily water usage data. By establishing a personal water usage behavior model, it identifies typical water usage scenarios (such as washing, laundry, and flushing toilets). When the water usage in a certain instance is significantly lower than the historical average for that scenario (such as a reduction of more than 30% in water usage for washing), and is within a reasonable range, it is considered an effective water-saving behavior. The system also sets a minimum duration to prevent false rapid switching from interfering with the judgment. Using the phone's built-in GPS and accelerometer sensors, the system collects the user's movement trajectory and gait characteristics. Trajectory analysis confirms walking as the mode of transportation, and map services are used to determine the path length. Simultaneously, a machine learning model distinguishes walking from other low-speed movements (such as slow-moving vehicles). Only when the distance traveled exceeds a set threshold (e.g., 500 meters), the speed is within the walking range (1-6 km / h), and there are no signs of transportation assistance, is it considered a low-carbon walking activity and included in the carbon emission reduction contribution. By comparing the spatial relationship between the user's pick-up and drop-off points and the school and residence through location data, and combining it with the schedule information of the school bus operation system (such as departure time and route ID) for matching and verification, users need to keep their location moving along the predetermined route over time during the school bus operation, and the stop points must conform to the station settings. At the same time, boarding confirmation is achieved by using Bluetooth beacons or NFC card check-in to enhance the reliability of authentication. Only when all data chains are complete and consistent will the trip be recorded as a low-carbon school bus travel behavior.

[0037] The system requires users to upload clear images or videos of the items to be transferred when initiating a transaction, and to fill in basic item information (such as name, category, brand, years of use, original price, etc.). The system uses image recognition technology combined with natural language processing algorithms to perform multimodal analysis on the uploaded content. First, it uses a convolutional neural network (CNN) to extract item features from the image, determining its authenticity and condition (such as its age and whether it is damaged), and then verifies the consistency with the information filled in by the user. At the same time, the system calls historical consumption data interfaces (such as e-commerce platform purchase records and electronic invoice information) or blockchain evidence data to verify that the item is indeed held by the user and has tradable attributes, preventing fictitious items or duplicate authentication. During the transaction, both the buyer and seller must complete the peer-to-peer transfer operation through real-name authenticated accounts. The system records a complete transaction log, including posting time, communication records, payment vouchers (if any), logistics information, or offline handover confirmation codes (such as QR code signing, face-to-face delivery location records within geofences), ensuring that the transaction actually occurred. To prevent fraudulent transactions or fake circular transactions, the system introduces a time window restriction mechanism (the interval between re-transactions of the same item must not be less than 6 months) and relationship graph analysis to detect whether there is high-frequency mutual recognition, credit abnormalities, or collusion risk between buyers and sellers. After the transaction is completed, the system estimates the carbon emission reduction brought about by its reuse based on the item type, original value, and depreciation rate (such as the carbon emission reduction from avoiding the production of new products). After weighted calculation, it is incorporated into the user's low-carbon points system, ultimately achieving a credible, quantifiable, and tamper-proof full-process authentication of second-hand goods transactions.

[0038] The beneficial effects of the above technical solution are as follows: by adopting the solution provided in this embodiment, it is possible to achieve accurate and automated authentication and recording of a variety of typical low-carbon behaviors on campus, effectively avoiding the problems of strong subjectivity, low efficiency and easy omissions in traditional manual recording methods.

[0039] In one embodiment, in step S2, the AI ​​recognition model adopts a multi-task learning architecture to simultaneously complete the category recognition of the low-carbon behavior to be certified, the detection of the number of participating entities, the estimation of duration, and the annotation of environmental interference factors. The environmental interference factors include changes in illumination, occlusion degree, motion blur, and semantic similarity interference. An attention mechanism is introduced to adaptively weight and focus on key visual regions. Key visual regions are visual feature regions highly correlated with the category of the low-carbon behavior to be certified. The multi-task learning architecture uses an improved ResNet-34 as the backbone feature extraction network. Its output feature map is input to four branches: the classification head, the detection head, the temporal evaluation head, and the interference factor evaluation head. The classification head (Head_cls) consists of a global average pooling layer (GAP) and two fully connected layers (FC), outputting the probability distribution of the behavior category using a cross-entropy loss function. The detection head (Head_det) uses a lightweight YOLO algorithm. The Head structure outputs the target bounding box coordinates and confidence score, using the CIoU loss function. The temporal evaluation head (Head_seq) is connected to a lightweight temporal convolutional module (TCN) to process the feature sequence of consecutive frames and output an action integrity score (0-1), using mean squared error loss. The interference factor evaluation head (Head_env) is similar in structure to the classification head, but outputs a four-dimensional vector corresponding to the severity scores of illumination, occlusion, blur, and semantic interference, using Smooth L1 loss. During the training phase, the total loss function is the weighted sum of the losses of each task, expressed by the formula... The calculation yielded that, Represents the total loss function. Represents the cross-entropy loss function. Represents the CIoU loss function. Represents the mean squared error loss function. Represents the Smooth L1 loss function, weights , , , It is not fixed, but dynamically adjusted using an uncertainty-weighted strategy, allowing the model to automatically learn the uncertainty of each task. The specific calculation formula is as follows: , The observable noise parameters for this task obtained by the model learning; The output of the AI ​​recognition model includes multiple branch output heads, which respectively generate the category probability distribution, the object detection bounding box sequence, the temporal action integrity score, and the confidence interval. The multi-task learning architecture shares backbone network parameters during the AI ​​recognition model training phase and combines gradient balancing strategies to balance the convergence speed differences among sub-tasks.

[0040] The working principle of the above technical solution is as follows: The AI ​​recognition model adopts a multi-task learning architecture, which can simultaneously identify the category of low-carbon behavior, detect the number of participants, estimate the duration of behavior, and also label environmental interference factors such as changes in lighting, degree of occlusion, motion blur, and semantic similarity interference. It does not require separate training of multiple models, which reduces redundant calculations, lowers deployment costs, and can also improve the overall recognition effect by utilizing the correlation between tasks. Environmental interference factors are external conditions that affect image recognition. For example, strong sunlight in the cafeteria at noon, someone blocking the subject of the action, or the action being too fast, causing the image to be blurry. Some actions may look similar to low-carbon behaviors and be easily confused. These factors will be used as additional information input into the model. During training, the data affected by interference will be weighted and processed. During inference, the results will also be corrected according to the interference situation, so that the model can stably recognize images in complex campus scenes. The model also incorporates an attention mechanism that can automatically determine which regions in an image are important. For example, when identifying garbage classification, it will focus on the area of ​​hand operations; when identifying cycling, it will focus on the area of ​​human joint movement. By focusing on analyzing these key regions and suppressing irrelevant background noise, the accuracy of classification and detection will be higher. The model's output has multiple branches, each responsible for different outputs: the first branch outputs the probability distribution of the behavior category, reflecting the model's confidence in judging the behavior type; the second branch outputs the object detection bounding box sequence, which can mark the position of the participating subjects in each frame, facilitating the counting of the number of people and spatial distribution; the third branch outputs the action integrity score (0-1 points), assessing whether the observed behavior segment was fully executed; and the fourth branch outputs the confidence interval, reflecting the confidence range of the model's prediction results, facilitating subsequent manual review or system judgment to determine whether supplementary verification is needed. The backbone network of the model uses an improved ResNet-34. During training, all sub-tasks share the parameters of this backbone network, eliminating the need for repeated feature extraction and saving resources. However, different sub-tasks converge at different speeds. For example, recognizing behavior categories is faster than estimating duration. Therefore, a gradient balancing strategy is used to dynamically adjust the loss weights of each task to prevent one task from dominating parameter updates, allowing all tasks to optimize collaboratively and converge stably.

[0041] The beneficial effects of the above technical solution are as follows: By adopting the solution provided in this embodiment, multi-dimensional information synchronous acquisition for low-carbon behavior recognition is achieved through a multi-task learning architecture. It can not only accurately identify behavior categories, but also simultaneously complete the statistics of the number of participants, the evaluation of behavior duration, and the labeling of environmental interference factors, providing comprehensive data support for the quantitative analysis of low-carbon behavior on campus.

[0042] In one embodiment, after step S3, the method further includes: organizing the standard carbon emission reduction data that meets preset conditions into a carbon asset declaration material package for review by the carbon inclusion platform; the preset conditions include and simultaneously meet the following: the cumulative carbon emission reduction value of a single verified low-carbon behavior record reaches the minimum carbon asset declaration unit threshold, the behavior of the verified low-carbon behavior record occurs within the carbon asset project cycle, and the integrity of the IoT sensor data associated with the low-carbon behavior in the verified low-carbon behavior record meets the carbon asset regulatory requirements.

[0043] The working principle of the above technical solution is as follows: After the system completes the verification of low-carbon behavior records (step S3), it will not immediately generate carbon asset declaration materials. Instead, it must first meet three preset conditions. Only when the three conditions are met at the same time will the declaration material package for the carbon inclusive platform to review be automatically generated. This can ensure the quality of the declared carbon assets and meet regulatory requirements. The first condition is that carbon emission reductions must meet the standard. For a single verified low-carbon behavior record, the cumulative carbon emission reductions must reach the minimum reporting unit threshold. This threshold is set in advance by the platform according to the project type, industry standards or policy requirements. This is mainly to avoid frequent reporting of scattered, low-value emission reductions, improve management efficiency and market liquidity. The system will continuously accumulate the verified emission reduction data of the same user or the same device. Only when the threshold is reached will it meet the requirements. The second condition is that the timing of the behavior must be compliant. Low-carbon behaviors must occur within the cycle of the carbon asset project. Each carbon asset project has a clear start and end time. Only behaviors that occur within this time period and are timestamped will be recognized. The system will compare the timestamp of each behavior record with the project cycle and exclude behavior data that occurs beyond the deadline or ahead of schedule to avoid invalid or duplicate entries. The third condition is that the sensor data must be complete. The completeness of IoT sensor data related to low-carbon behavior must meet the requirements of carbon asset regulation. IoT devices will collect raw behavioral data in real time. The system will use algorithms to detect data continuity, calculate missing rate, and identify outliers. For example, it is required that the data missing rate be less than 5%, key fields cannot have empty values, and time series cannot have breakpoints. Only when the data completeness meets the standard can the behavior be proven to be real and credible and have the evidence to support assetization. Once all three conditions are met, the system will automatically integrate the relevant data and package it into a standardized carbon asset declaration package according to the format specified by the platform. It will also include digital signatures or blockchain evidence, which will facilitate automated or manual review by the auditing agency, and ultimately transform the scattered low-carbon behavior records into compliant carbon assets.

[0044] The beneficial effects of the above technical solution are as follows: The solution provided in this embodiment can effectively improve the standardization and feasibility of carbon assetization of low-carbon behavior on campus; on the one hand, by setting multiple strict preset conditions, the waste of management costs caused by fragmented emission reductions is avoided, and non-compliant data is intercepted from entering the carbon asset circulation process at the source; on the other hand, the system automatically generates standardized application material packages and attaches digital signatures or blockchain evidence, realizing the full-process digitalization and traceability from low-carbon behavior recording to carbon asset application, which not only reduces the error risk that may be caused by manual operation, but also provides a solid data foundation and trust guarantee for subsequent carbon asset transactions, audits and other links.

[0045] In one embodiment, in step S5 Logical binding is achieved through smart contracts deployed on a blockchain network built and maintained by the Carbon Inclusive platform; When a carbon asset project package is approved by the regulatory node of the carbon benefit platform and a unique serial number is issued to it, a smart contract pre-deployed in the blockchain network is automatically triggered to associate the serial number with the block hash or globally unique identifier of all source behavior evidence blocks that constitute the carbon asset project package. This creates a traceable and tamper-proof mapping relationship between the serial number and each source behavior evidence block. The source behavior evidence block is the evidence block corresponding to the low-carbon behavior of each user that constitutes the carbon asset project package.

[0046] The working principle of the above technical solution is as follows: The blockchain network is built and maintained by the carbon benefit platform, which can provide a decentralized, tamper-proof, and traceable data storage environment to ensure the security and trustworthiness of carbon asset-related data. Smart contracts are automated programs deployed on the blockchain. They will be executed automatically when preset conditions are met without human intervention, ensuring transparent and consistent operation. Once a carbon asset project package passes the review of the platform's regulatory node, the node will issue it a unique serial number. This serial number serves as the project package's exclusive identity on the blockchain. Successful serial number issuance automatically triggers a pre-deployed smart contract on the blockchain, initiating the subsequent binding logic. The smart contract will read the source behavior evidence block information corresponding to all user low-carbon behaviors contained in this carbon asset project package. These blocks have been completed on the chain before. Each block has complete evidence data and timestamp, as well as a unique block hash or globally unique identifier (such as UUID). This identifier can accurately point to a specific block and will not be repeated or forged. The smart contract will structurally associate the project package's serial number with the hashes or identifiers of all associated source behavior evidence blocks, forming a mapping table of serial numbers to multiple source behavior identifiers. This mapping relationship will be permanently stored on the blockchain, and subsequent queries, verifications, and audits can all be completed through on-chain data, ensuring full traceability. Moreover, because the blockchain is immutable, this mapping relationship cannot be changed or deleted once it is on the chain, effectively preventing data forgery, reuse, or malicious tampering. External systems or regulators can also use public interfaces to query all source behavior records bound to a specific serial number, verifying the true source and compliance of carbon assets, thus making the entire carbon asset process more credible.

[0047] The beneficial effects of the above technical solution are as follows: by adopting the solution provided in this embodiment, the entire logical binding process is automated, standardized and verifiable, which can significantly improve the efficiency of carbon asset management, reduce the risk of human operation, and support the assetization and circulation of large-scale low-carbon behaviors.

[0048] In one embodiment, in step S3, the verified low-carbon behavior records are quantified into standard carbon emission reduction data according to a preset carbon inclusiveness methodology model, introducing uncertainty quantification, specifically including: Using the comprehensive confidence level corresponding to the verified low-carbon behavior records as prior information, and combined with standard carbon emission reduction data, the measurement error range of each input parameter is calculated as a characterization of observation uncertainty, and a Bayesian inference framework is constructed. Within this Bayesian inference framework, the posterior probability distribution of standard carbon emission reduction data is modeled using Markov chain Monte Carlo or variational inference methods to obtain the uncertainty distribution; Based on the uncertainty distribution, the expected value of the standard carbon emission reduction data is extracted as the central value of carbon emission reduction, and the uncertainty interval is determined by the coverage interval under a pre-set confidence level. The uncertainty interval is used as a component of the evidence information in step S4, and an error propagation model based on covariance propagation is used to mathematically synthesize the uncertainty intervals of the components of the carbon asset project package in step S5. The joint uncertainty distribution of the total carbon emission reduction of the carbon asset project package is calculated, and the aggregated total emission reduction center value and its corresponding total uncertainty interval are derived from it.

[0049] The working principle of the above technical solution is as follows: In step S3, the system will convert the verified low-carbon behavior records into standard carbon emission reduction data according to the preset carbon inclusive methodology model. This process is not just a simple numerical mapping, but also considers data quality and measurement reliability. Each low-carbon behavior record will be accompanied by a comprehensive confidence level. This comprehensive confidence level is obtained by the fusion of multi-source information in the early stage, which can reflect the credibility of the record and will be used as prior information in subsequent calculations. Based on this overall confidence level, and combined with the measurement error range of various input parameters used in the calculation of carbon emission reduction, a Bayesian statistical model (i.e., a Bayesian inference framework) is established so that the assessment of carbon emission reduction can reflect the uncertainty of the data. Within this Bayesian framework, either Markov Chain Monte Carlo (MCMC) or Variational Inference (VI) is used to model the posterior probability distribution of standard carbon emission reductions. MCMC approximates the target distribution by constructing Markov chains, making it suitable for complex nonlinear models. VI, on the other hand, optimizes the difference between the approximate distribution and the true posterior, improving computational efficiency while maintaining accuracy. Both methods can obtain the complete probability distribution of standard carbon emission reductions, laying the foundation for subsequent extraction of uncertainty information. Based on the obtained probability distribution, the expected value is taken as the center value of the standard carbon emission reduction. This value is the optimal estimate in a statistical sense. At the same time, based on the preset confidence level (such as 95%), the corresponding coverage interval is calculated to form an uncertainty interval. This interval can intuitively reflect the reliable range of the emission reduction estimate and is an important indicator for measuring the quality of carbon credit. It will be included in the evidence storage information in step S4 to make the data more transparent and facilitate subsequent audit traceability. In step S5, multiple certified emission reduction units are integrated into a carbon asset project package. Each emission reduction unit has its own central value and independent uncertainty range for its standard carbon emission reduction. The system adopts an error propagation model based on covariance propagation, considers the possible correlation between different emission reduction sources, and mathematically synthesizes the variance and covariance matrices of each unit to calculate the joint probability distribution of the total carbon emission reduction of the project package. This yields the central value of the aggregated total emission reduction and the corresponding total uncertainty range, realizing the transmission and integration of uncertainty from individual units to the whole, ensuring that the summary results are scientific and rigorous, neither overestimating nor underestimating the actual emission reduction effect.

[0050] The beneficial effects of the above technical solution are as follows: By introducing an uncertainty quantification mechanism, the scientific validity and credibility of carbon emission reduction data can be improved; by integrating the comprehensive confidence level as prior information into the Bayesian inference framework, the calculation of carbon emission reduction is no longer a single deterministic value, but a probability distribution that reflects the inherent uncertainty of the data; by using an error propagation model based on covariance propagation to mathematically synthesize the uncertainty intervals of each component, the joint uncertainty distribution of the total carbon emission reduction of the carbon asset project package and its corresponding total central value and total uncertainty interval can be accurately calculated, ensuring that the aggregated total emission reduction data, under the premise of scientific rigor, neither overestimates nor underestimates the actual emission reduction effect, providing a high-quality and reliable data foundation for carbon asset accounting, trading, and subsequent policy formulation.

[0051] In one embodiment, such as Figure 3 As shown, it also includes enhanced verification processing for abnormal low-carbon behavior authentication requests initiated by the carbon benefit platform. The specific process is as follows: Real-time monitoring of the frequency of low-carbon behavior authentication requests initiated by the same user, the same device, or the same geographical area within a unit time window, wherein the unit time window is a continuous 60-minute time interval; If the frequency exceeds the frequency threshold dynamically calculated based on historical baselines, and the confidence interval generated by the output of the AI ​​recognition model exceeds the preset threshold interval, the low-carbon behavior certification request will be marked as a potential anomaly, and enhanced verification will be initiated. Only after the enhanced verification is passed will the low-carbon behavior certification request be marked as normal and allowed to proceed to the next steps.

[0052] The working principle of the above technical solution is as follows: the system will monitor low-carbon behavior authentication requests in real time to prevent malicious requests and fake authentication; specifically, it will focus on the frequency of authentication requests initiated by the same user, the same device or the same geographical area within a unit time window; the unit time window is a continuous 60-minute time interval. The system will build a dynamic behavior baseline based on historical data (such as the average number of daily requests per user, request characteristics at different times, and activity levels in different regions). This baseline will be continuously updated to allow the frequency threshold to adapt to different times and different users. If the request frequency of a user, device, or region is detected to exceed this dynamically calculated threshold, a preliminary risk assessment will be triggered. Next, we will combine the results of the AI ​​recognition model to make further judgments, analyze whether the behavior type of the current authentication request is consistent, whether the geographical location is reasonable, whether the time series is logical, and whether the metadata is complete. If the confidence interval output by the model exceeds the preset security range (for example, the confidence level is lower than 0.6 or higher than 0.95, both of which may be abnormal), then the double abnormality condition is met. If both abnormal conditions are met, the system will mark the authentication request as potentially abnormal and initiate an enhanced verification process. Only after the enhanced verification is passed will the request be marked as normal and proceed to subsequent steps such as points calculation, on-chain notarization, or reward distribution.

[0053] The beneficial effects of the above technical solution are as follows: By adopting the solution provided in this embodiment, it is possible to accurately identify and intercept high-frequency abnormal low-carbon behavior authentication requests through a dual verification mechanism of real-time frequency monitoring and AI recognition credibility, effectively preventing risky behaviors such as malicious carbon credit farming and false authentication.

[0054] In one embodiment, enhanced verification includes at least one of the following verification mechanisms: When the acquisition angle of the on-site images of the low-carbon behavior to be certified is insufficient to fully represent the low-carbon behavior to be certified, an interactive instruction is generated to request the user to provide supplementary video stream data of the on-site low-carbon behavior to be certified from different spatial perspectives. When there are time discontinuities or data sparsity issues in the environmental perception information in IoT sensor data, a data retrieval request is sent to the IoT sensor to obtain the relevant sensor data sequence covering a longer time window for time sequence continuity verification. When the location information in the campus management system data is not well matched or there is spatial overlap interference from multiple entities, a location cross-verification based on geofencing and GNSS / Bluetooth beacon fusion positioning is initiated.

[0055] The working principle of the above technical solution is as follows: Enhanced verification will adopt corresponding verification mechanisms based on different abnormal situations to ensure the authenticity of the authentication request. Specifically, there are three situations: To address insufficient image acquisition angles, if the system detects that the angle of the on-site image uploaded by the user is not good and key behavioral features are not fully captured (e.g., only partial actions are captured, or there is severe occlusion), it will automatically send an interactive command to allow the user to record and upload a continuous video stream from different angles (e.g., front, side, top) using their mobile phone camera. The system will use a 3D behavior understanding model to determine whether the viewpoint is complete. If the current image or video is insufficient to clearly identify the behavior, data must be supplemented to improve the accuracy of subsequent recognition and verification. To address issues with IoT sensor data, when IoT sensors are running for extended periods, data may become intermittent due to network outages or device hibernation, or data may become sparse due to low reporting frequency. In such cases, the system will proactively send high-priority data retrieval requests to the sensors to obtain sensor data over a longer time span (e.g., 30 minutes before and after). These extended time-series data will then be interpolated, trend-fitted, and anomaly detected. The system will also analyze whether changes in environmental parameters conform to physical laws (e.g., gradual changes in temperature and humidity, and a steady decrease in energy consumption), thereby eliminating forged or spliced ​​false data. To address insufficient location information matching, if the location information provided by the campus management system deviates from the expected behavioral trajectory by more than 5 meters, or if the location trajectories of multiple users overlap significantly during the same period, it may lead to identity confusion and incorrect attribution of behavior. In such cases, the system will initiate multi-source positioning fusion verification, combining outdoor GNSS positioning and indoor sub-meter-level positioning via Bluetooth beacons. Kalman filtering or particle filtering algorithms will then be used to integrate these positioning signals, reconstruct the user's actual movement path, and then accurately match it with preset low-carbon behavior areas such as garbage sorting points and bicycle parking areas to ensure that the geographical location of the behavior is true and unique.

[0056] The beneficial effects of the above technical solution are as follows: by adopting the solution provided in this embodiment, the reliability and accuracy of campus low-carbon behavior certification can be effectively improved through a multi-dimensional and multi-level verification mechanism.

[0057] In specific application implementation, taking a student named Xiao Li's participation in campus waste sorting activities in a carbon credit project at a certain university as an example, the specific implementation steps are as follows: At the school's waste sorting and disposal point, Xiao Li correctly sorted and disposed of the recyclables (plastic bottles, waste paper, etc.) he collected; the smart trash can at the disposal point recorded information such as the disposal time, weight, and the real-time filling status of various types of waste in the trash can; Xiao Li used the Campus Carbon Benefits APP to take photos of the garbage and the disposal point before and after disposal. The photos included time watermarks and geographical location information. The school's access control system and surveillance cameras recorded the time Xiao Li entered and left the garbage sorting and disposal area, which matched the time of the event. The location information showed that he was accurately located at the garbage sorting and disposal point. After preliminary verification of these data, the system generates a source behavior evidence block for Xiao Li's garbage sorting behavior and puts it on the blockchain. This block generates a unique block hash H2=hash(block_data) and writes it into the blockchain network to form an immutable original behavior certificate. Based on a pre-defined carbon inclusion methodology model, the system transforms Xiao Li's waste sorting behavior records into standard carbon emission reduction data. The overall confidence level of this behavior is 0.9, which serves as prior information. A Bayesian inference framework is constructed by combining the measurement error ranges of various input parameters relied upon in the standard carbon emission reduction calculation process. The Markov Chain Monte Carlo (MCMC) method is used to model the posterior probability distribution of the standard carbon emission reduction. By constructing a Markov chain that converges to the target distribution, sampling approximation is achieved, resulting in the complete probability distribution of the standard carbon emission reduction. For example, the measurement error range of the recyclable weight in the waste sorting behavior is ±0.1 kg. According to the nonlinear transformation function between recyclable weight and carbon emission reduction in the methodology model, the fluctuation range of the error propagated to the carbon emission reduction calculation result is [-0.02 kg CO2e, +0.03 kg CO2e]. The image recognition accuracy for plastic bottles and waste paper was 0.95 and 0.92, respectively. The joint recognition error was converted into an uncertainty factor for carbon emission reduction through a confusion matrix. The standard deviation of the error distribution in this dimension was 0.015 kgCO2e, obtained by Monte Carlo simulation. The validity verification error of auxiliary data such as timestamps and location information was quantified as a systematic bias of 0.005 kgCO2e. The above prior information and the uncertainty distribution of each parameter were input into a Bayesian inference framework. The Markov Chain Monte Carlo (MCMC) method was used for 10,000 iterations of sampling. Finally, the expected standard carbon emission reduction value corresponding to this waste sorting behavior was 0.5 kgCO2e, with a 95% confidence interval of [0.46 kgCO2e, 0.54 kgCO2e]. This result was used as the basis for subsequent carbon asset accounting. Based on the obtained posterior distribution, its expected value is taken as the central value of the standard carbon emission reduction, assumed to be 0.5 kg CO2e; according to the 95% confidence level, the coverage interval defined by the corresponding quantile is calculated to form the uncertainty interval [0.46 kg CO2e, 0.54 kg CO2e], which is included in the evidence information in step S4. The school integrates the certified carbon reduction units of multiple students into a carbon asset project package. Assuming that in addition to Xiao Li's garbage sorting behavior, there are also Xiao Zhang's green travel (walking to and from get out of class) and Xiao Wang's electricity saving (turning off the appliances in the unattended classroom), the standard carbon reduction of each behavior has its own central value and uncertainty range. The system adopts an error propagation model based on covariance propagation, considering the possible correlations between different emission reduction sources (e.g., there may be no direct correlation between walking and saving electricity, but there may be a correlation between waste sorting and energy consumption in the waste treatment process). It mathematically synthesizes the variance and covariance matrices of each input variable, calculates the joint probability distribution of the total carbon emission reduction of the carbon asset project package, and then derives the central value of the aggregated total emission reduction, which is assumed to be 1 kgCO2e, and the corresponding total uncertainty interval is assumed to be [0.89 kgCO2e, 1.11 kgCO2e], realizing the transmission and integration of uncertainty from individual to overall. When a school's carbon asset project package (containing the low-carbon behaviors of multiple students) is submitted to the regulatory node for review, the regulatory node conducts a rigorous review of all information in the project package, including the authenticity of the students' low-carbon behaviors, the accuracy of the data, and the rationality of the emission reduction calculation. After the review is approved, the regulatory node issues a unique serial number SN20251114001 to the project package as the unique identity of the project package on the blockchain. The generation and issuance of serial numbers triggers the execution mechanism of the smart contract. The smart contract reads the source behavior evidence block information corresponding to all the students' low-carbon behaviors that constitute the carbon asset project package, such as the source behavior evidence block hash H2 for Xiao Li's garbage sorting behavior, the source behavior evidence block hash H3 for Xiao Zhang's green travel behavior, and the source behavior evidence block hash H4 for Xiao Wang's electricity saving behavior, etc. The smart contract structurally associates the serial number SN20251114001 of the carbon asset project package with the block hashes of all associated source behavior evidence blocks, forming a mapping table. This mapping is written into the blockchain and permanently stored. Any subsequent query, verification or audit of this relationship can be completed through on-chain data, ensuring that the entire process is traceable. For abnormal low-carbon behavior certification requests, the school's carbon credit platform monitors in real time the frequency of low-carbon behavior certification requests initiated by the same student, the same device, or the same geographical area within a unit time window. For example, it was found that student Xiao Zhao initiated 5 garbage sorting behavior certification requests within 1 hour, while the frequency threshold dynamically calculated based on historical baselines is 3 times / hour. At the same time, the AI ​​recognition model performs multi-dimensional semantic and contextual analysis on Xiao Zhao's certification request content, including consistency of behavior type, reasonableness of geographical location, logicality of time series, and integrity of metadata. If the confidence interval of the model output exceeds the preset threshold interval (confidence level below 0.6), the system marks Xiao Zhao's low-carbon behavior certification request as a potential anomaly and initiates enhanced verification. If the system detects that the angle of the garbage sorting photos uploaded by Xiao Zhao is insufficient to fully represent his behavior (e.g., only part of the garbage is captured, and key sorting actions are obscured), the system automatically generates an interactive command, prompting Xiao Zhao to record and upload a continuous video stream from different spatial positions (e.g., front, side, top, etc.) using his mobile device's camera to fully demonstrate the garbage sorting process. If the IoT sensor data at the disposal point has time discontinuity (e.g., some data is missing due to network interruption) or data sparsity (reporting frequency is too low), the system sends a high-priority data retrieval request to the relevant sensor nodes to obtain the original sensor data sequence covering a longer time range (e.g., 30 minutes before and after). By interpolating, fitting trends, and detecting outliers on the expanded time series, the system performs time series continuity analysis. The system determines whether changes in environmental parameters conform to physical laws. When the matching degree between Xiao Zhao's location information provided by the campus management system and the expected behavioral trajectory is lower than a set threshold (e.g., the location deviation exceeds 5 meters), or multiple students exhibit highly overlapping spatial trajectories within the same time period, the system initiates multi-source positioning fusion verification under geofence constraints. Combining macroscopic outdoor positioning provided by GNSS and Bluetooth beacons, it achieves sub-meter level precise indoor positioning. Kalman filtering or particle filtering algorithms are used to perform spatiotemporal alignment and weighted fusion of multi-source signals to reconstruct Xiao Zhao's actual movement path and accurately match it with the preset low-carbon behavior occurrence area (e.g., garbage sorting point). Only after the enhanced verification passes is Xiao Zhao's low-carbon behavior certification request marked as normal and allowed to proceed to subsequent steps, such as points calculation, on-chain notarization, or reward distribution. It should be noted that the above embodiments demonstrate the process of the system handling the allocation of emission reductions for collective behavior in a specific scenario. The data arbitration and allocation mechanism protected by this invention is not limited to this. When the multi-source evidence chain points to a single behavior being contributed by multiple participating entities, or when a single entity's behavior involves multiple divisible quantifiable units, the system can fairly and transparently divide the final verified carbon emission reductions according to preset, traceable rules (such as average allocation, weighted allocation, contribution proof, etc.), and store the division logic and results on the blockchain together. This ensures the scalability and fairness of the solution of this invention when encouraging collective low-carbon behavior.

[0058] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for identifying and recording low-carbon behaviors on campus based on AI image recognition, characterized in that, Includes the following steps: S1: In response to a user's request for low-carbon behavior certification initiated through the carbon benefit platform, obtain multimodal verification data, including on-site images of the low-carbon behavior to be certified; S2: Use a pre-trained AI recognition model to identify multimodal verification data, and combine it with at least one of the IoT sensor data and campus management system data associated with the low-carbon behavior to be certified for cross-validation to generate verified low-carbon behavior records. S3: Based on the preset carbon inclusiveness methodology model, the verified low-carbon behavior records are quantified into standard carbon emission reduction data, and corresponding carbon credits are issued to the user's carbon account. The carbon inclusive methodology model is constructed based on a pre-set low-carbon behavior emission reduction factor database and / or a behavior-driven carbon emission reduction quantitative assessment model trained on historical data; the emission reduction factor database includes standard parameters for unit emission reduction of different low-carbon behavior categories; S4: Upload the target information of standard carbon emission reduction data and on-site images of low-carbon behaviors to be certified as evidence to the blockchain network to generate an immutable evidence block; the target information includes a verifiable digital digest, a searchable storage identifier, a timestamp, and a user anonymity identifier; S5: On the blockchain network, the identifier of the evidence storage block is logically bound to one or more carbon asset project packages that meet the carbon inclusive certification emission reduction declaration requirements through smart contracts; among them, the carbon asset project package is formed by the carbon inclusive platform operator by aggregating the records of similar low-carbon behaviors of multiple users and the standard carbon emission reduction data generated therefrom and then submitting them in a unified manner. S6: Provides a traceable query interface for low-carbon behavior records associated with a user's carbon account. This traceable query interface is used to display the evidence block information corresponding to the low-carbon behavior record through the blockchain network.

2. The campus low-carbon behavior recognition and recording method based on AI image recognition according to claim 1, characterized in that, The method is implemented through a carbon equity platform, which is a comprehensive technology platform built for campus scenarios, integrating data collection, intelligent verification, and carbon asset management. Its system architecture, from bottom to top, includes: The data acquisition layer is used to connect to and standardize access to various low-carbon behavior-related data sources on campus, including intelligent waste sorting and recycling equipment, library management system, mobile health application step counting interface and image acquisition terminal. Among them, the image acquisition terminal is deployed in canteen, teaching building and dormitory, and is specifically used to capture on-site images of low-carbon behaviors. The data layer is used for unified cleaning, structured governance, persistent storage, and in-depth analysis of various low-carbon behavior-related data sources; The support layer is used to integrate IoT platforms to enable device connectivity and status monitoring, as well as to integrate AI platforms to support the deployment and inference optimization of pre-trained AI recognition models; The application layer includes a personal carbon ledger module, a low-carbon behavior verification module, and a low-carbon learning knowledge base. The personal carbon ledger module records the accumulation and usage of carbon credits in the user's carbon account. The low-carbon behavior verification module is used to realize the low-carbon behavior certification request and response and the multimodal verification data flow interaction. The low-carbon learning knowledge base is used to provide personalized green and low-carbon education content recommendations and learning outcome certification.

3. The campus low-carbon behavior recognition and recording method based on AI image recognition according to claim 2, characterized in that, The image acquisition terminal is equipped with a lightweight model, which is based on the full cloud recognition model deployed on a cloud server. It is obtained after processing by one or more of the following methods: knowledge distillation, pruning, or quantization. It is deployed using a containerized encapsulation and edge collaborative scheduling architecture, and supports incremental updates. Among them, knowledge distillation uses the full cloud recognition model as the teacher model to guide the lightweight student model to learn its output layer soft label distribution and intermediate layer feature representation. Pruning evaluates the importance of redundant convolution kernels or channels in the full cloud recognition model and filters and removes low-contribution weights based on preset indicators, including L1 norm, gradient sensitivity, or Taylor expansion. Quantization converts the floating-point parameters in the full cloud recognition model into low-bit integers.

4. The campus low-carbon behavior recognition and recording method based on AI image recognition according to claim 1, characterized in that, Low-carbon behaviors to be certified include at least one of the following: Clean Plate Campaign, proper sorting and disposal of recyclables, saving electricity, saving water, low-carbon walking, school bus travel, and second-hand goods trading.

5. The campus low-carbon behavior recognition and recording method based on AI image recognition according to claim 1, characterized in that, In step S2, the AI ​​recognition model adopts a multi-task learning architecture to simultaneously complete the category recognition of the low-carbon behavior to be certified, the detection of the number of participating entities, the estimation of duration, and the labeling of environmental interference factors. Among them, environmental interference factors include changes in illumination, degree of occlusion, motion blur, and semantic similarity interference. An attention mechanism is introduced to adaptively weight and focus on key visual regions. Key visual regions are visual feature regions that are highly correlated with the category of low-carbon behavior to be certified. The output of the AI ​​recognition model includes multiple branch output heads, which respectively generate the category probability distribution, the object detection bounding box sequence, the temporal action integrity score, and the confidence interval. The multi-task learning architecture shares backbone network parameters during the AI ​​recognition model training phase and combines gradient balancing strategies to balance the convergence speed differences among sub-tasks.

6. The campus low-carbon behavior recognition and recording method based on AI image recognition according to claim 1, characterized in that, Following step S3, the process further includes: organizing the standard carbon emission reduction data that meets the preset conditions into a carbon asset declaration material package for review by the carbon inclusion platform; the preset conditions include and simultaneously meet the following: the cumulative carbon emission reduction value of a single verified low-carbon behavior record reaches the minimum carbon asset declaration unit threshold, the behavior of the verified low-carbon behavior record occurs within the carbon asset project cycle, and the integrity of the IoT sensor data associated with the low-carbon behavior in the verified low-carbon behavior record meets the carbon asset regulatory requirements.

7. The campus low-carbon behavior recognition and recording method based on AI image recognition according to claim 1, characterized in that, In step S5, Logical binding is achieved through smart contracts deployed on a blockchain network built and maintained by the Carbon Inclusive platform; When a carbon asset project package is approved by the regulatory node of the carbon benefit platform and a unique serial number is issued to it, a smart contract pre-deployed in the blockchain network is automatically triggered to associate the serial number with the block hash or globally unique identifier of all source behavior evidence blocks that constitute the carbon asset project package. This creates a traceable and tamper-proof mapping relationship between the serial number and each source behavior evidence block. The source behavior evidence blocks are evidence blocks corresponding to the low-carbon behaviors of each user that constitutes the carbon asset project package.

8. The campus low-carbon behavior recognition and recording method based on AI image recognition according to claim 1, characterized in that, In step S3, based on the preset carbon inclusiveness methodology model, the verified low-carbon behavior records are quantified into standard carbon emission reduction data, introducing uncertainty quantification, specifically including: Using the comprehensive confidence level corresponding to the verified low-carbon behavior records as prior information, and combined with standard carbon emission reduction data, the measurement error range of each input parameter is calculated as a characterization of observation uncertainty, and a Bayesian inference framework is constructed. Within this Bayesian inference framework, the posterior probability distribution of standard carbon emission reduction data is modeled using Markov chain Monte Carlo or variational inference methods to obtain the uncertainty distribution; Based on the uncertainty distribution, the expected value of the standard carbon emission reduction data is extracted as the central value of carbon emission reduction, and the uncertainty interval is determined by the coverage interval under a pre-set confidence level. The uncertainty interval is used as a component of the evidence information in step S4, and an error propagation model based on covariance propagation is used to mathematically synthesize the uncertainty intervals of the components of the carbon asset project package in step S5. The joint uncertainty distribution of the total carbon emission reduction of the carbon asset project package is calculated, and the aggregated total emission reduction center value and its corresponding total uncertainty interval are derived from it.

9. The campus low-carbon behavior recognition and recording method based on AI image recognition according to claim 1, characterized in that, This also includes enhanced verification processing for abnormal low-carbon behavior authentication requests initiated by the carbon benefit platform. The specific process is as follows: Real-time monitoring of the frequency of low-carbon behavior authentication requests initiated by the same user, the same device, or the same geographical area within a unit time window, wherein the unit time window is a continuous 60-minute time interval; If the frequency exceeds the frequency threshold dynamically calculated based on historical baselines, and the confidence interval generated by the output of the AI ​​recognition model exceeds the preset threshold interval, the low-carbon behavior certification request will be marked as a potential anomaly, and enhanced verification will be initiated. Only after the enhanced verification is passed will the low-carbon behavior certification request be marked as normal and allowed to proceed to the next steps.

10. The campus low-carbon behavior recognition and recording method based on AI image recognition according to claim 9, characterized in that, Enhanced verification includes at least one of the following verification mechanisms: When the acquisition angle of the on-site images of the low-carbon behavior to be certified is insufficient to fully represent the low-carbon behavior to be certified, an interactive instruction is generated to request the user to provide supplementary video stream data of the on-site low-carbon behavior to be certified from different spatial perspectives. When there are time discontinuities or data sparsity issues in the environmental perception information in IoT sensor data, a data retrieval request is sent to the IoT sensor to obtain the relevant sensor data sequence covering a longer time window for time sequence continuity verification. When the location information in the campus management system data is not well matched or there is spatial overlap interference from multiple entities, a location cross-verification based on geofencing and GNSS / Bluetooth beacon fusion positioning is initiated.

Images