Electric power industry environmental protection supervision cloud edge algorithm warehouse device and cooperative processing method

Abstract

The invention provides a power industry environmental protection supervision cloud edge algorithm warehouse device and a cooperative processing method. The device comprises three modules and an encryption communication link: an algorithm warehouse cloud platform (a cloud core, integrated data annotation, model development training push and other functions), edge nodes (field deployment, video processing clear image screening, localized reasoning and real-time alarm), a block chain evidence storage platform (a distributed account book and full-process operation encryption evidence storage), and an encryption communication link. And secure transmission is realized through an SSL / TLS protocol and the like. FFmpeg decoding and perceptual hash de-duplication are adopted by edge nodes to reserve a clear image, and FPGA accelerates parallel reasoning, so that the processing efficiency is improved, and the reasoning delay is extremely low; the model and data security is guaranteed through full-process encryption; the uploading bandwidth is reduced through localized reasoning, manpower is replaced by automation, and the cost is reduced; the block chain evidence cannot be tampered, and the tracing is convenient.

Description

Technical Field

[0001] This invention relates to the field of environmental protection supervision in the power industry, specifically a cloud-edge algorithm warehouse device and collaborative processing method for environmental protection supervision in the power industry. Background Technology

[0002] Soil and water conservation is a crucial measure for protecting the ecological environment, an essential requirement for implementing the central government's major decisions on ecological civilization construction, and a necessity for the people's growing need for a better life. Currently, environmental supervision and management of power industry construction projects has reached a stage requiring innovation. Faced with a massive workload of inspections, the traditional approach involves collecting visual data on-site using high-resolution satellites, drones, cameras, or personnel, followed by manual identification. The shortcomings of this existing technology are as follows:

[0003] 1. High cost: The purchase and maintenance costs of satellite remote sensing and drone aerial photography equipment are high, and most power operations are located in the suburbs, requiring a large investment of manpower, and the long-term monitoring costs remain high;

[0004] 2. Low efficiency: Massive amounts of view data need to be processed manually one by one, resulting in long recognition cycles and making it impossible to detect environmental issues in a timely manner;

[0005] 3. Unstable accuracy: Manual identification is affected by subjective factors such as the experience and fatigue of the staff, and is prone to missed or misjudged problems such as small areas of exposed soil (e.g., <10㎡) and slight smoke emissions, resulting in large fluctuations in identification accuracy;

[0006] 4. Lack of closed-loop process: There is a lack of unified recording and traceability mechanisms for data collection, identification, and processing. If a supervisory dispute arises, it is impossible to quickly verify the operation process and it is difficult to determine responsibility.

[0007] With the continuous advancement of science and technology, remote, information-based, and intelligent environmental supervision of power transmission and transformation projects has become an inevitable trend. Summary of the Invention

[0008] In view of this, the present invention provides a cloud-edge algorithm warehouse device and collaborative processing method for environmental protection supervision in the power industry to solve the above-mentioned technical problems.

[0009] A cloud-edge algorithm warehouse device for environmental monitoring in the power industry, comprising:

[0010] The Algorithm Warehouse Cloud Platform module, as the core control center in the cloud, integrates a data annotation unit, a model development unit, a model training unit, a model optimization unit, a node management unit, and an image alarm processing unit. It is used to receive edge node registration requests, develop and train environmental monitoring models, push models to edge nodes, and process alarm information.

[0011] The edge node platform module, deployed at the monitoring site, includes a video processing unit, a model inference unit, and a communication and alarm unit. It is used to collect video streams and filter clear images, perform local inference based on cloud-pushed models, and trigger real-time alarms.

[0012] The blockchain evidence storage platform module is a distributed ledger system that includes an operation record collection unit, an encrypted evidence storage unit, and a traceability query unit. It is used to collect and encrypt the entire process operation data of the cloud platform and edge nodes.

[0013] An encrypted communication link, employing the SSL / TLS protocol and heartbeat mechanism, connects the algorithm warehouse cloud platform module, edge node platform module, and blockchain evidence storage platform module, enabling secure data transmission and status synchronization among the three.

[0014] Furthermore, the various units of the algorithm warehouse cloud platform module are connected sequentially via data interfaces;

[0015] The data annotation unit outputs annotation data to the model development unit, which then generates a model structure file and transmits it to the model training unit.

[0016] The model training unit trains the model using labeled data and outputs the initial model to the model tuning unit.

[0017] After optimizing the model parameters, the model tuning unit stores the qualified model in the model storage unit;

[0018] The edge node management unit retrieves the adapted model from the model storage unit and pushes it to the edge node platform module via an encrypted communication link.

[0019] The image alarm processing unit receives inference alarm images uploaded by edge nodes in real time through a message queue.

[0020] Furthermore, the data flow direction of each unit in the edge node platform module is as follows:

[0021] The video processing unit receives the video stream, decodes and filters it, and outputs optimized images with PSNR≥30dB and SSIM≥0.9 to the model inference unit.

[0022] The model inference unit loads the model pushed from the cloud, performs parallel inference on the optimized image, and outputs a recognition result with a confidence level ≥ 0.8 to the communication and alarm unit.

[0023] The communication and alarm unit encrypts the results that meet the alarm conditions and pushes them to the algorithm warehouse cloud platform module through the encrypted communication link.

[0024] Furthermore, the workflow of the blockchain evidence storage platform module is as follows:

[0025] The operation log acquisition unit collects real-time operation data for model uploading / downloading / training / deployment, edge node registration, and alarm handling.

[0026] The encrypted evidence storage unit encodes the operation data into JSON format, writes it into the blockchain using the BFT-SMaRT consensus algorithm, and encrypts the blocks using the SHA-256 algorithm.

[0027] The traceability query unit provides a retrieval interface based on operation ID and time range, outputs evidence records containing the entire operation chain, and automatically verifies the integrity of hash values.

[0028] Furthermore, the encrypted communication link includes:

[0029] When edge nodes register, permissions are verified using the RSA algorithm, and the cloud platform assigns a unique identifier and encryption key.

[0030] The model file and alarm information are transmitted using SSL / TLS protocol encryption. The edge node sends a heartbeat packet every 30 seconds, and the cloud platform triggers a reconnection mechanism if it does not receive a response.

[0031] A cloud-edge collaborative processing method for environmental monitoring algorithms in the power industry includes the following steps:

[0032] S1. System Initialization: Edge nodes send a registration request to the cloud platform via an encrypted communication link. After the cloud platform verifies the request, a secure connection is established.

[0033] S2. Model Development and Deployment: Develop and train environmental monitoring models on the cloud platform, push them to edge nodes, and verify their integrity;

[0034] S3. Localized Inference and Alarm: Edge nodes process video streams and infer based on models, and push alarms to the cloud platform after triggering them;

[0035] S4. Alarm Handling and Evidence Preservation: The cloud platform assigns personnel to handle alarms offline, and the entire process of blockchain-based synchronous evidence preservation is carried out.

[0036] S5. Traceability and Auditing: Verify the legitimacy of operation records through blockchain traceability query units.

[0037] Furthermore, step S2 includes:

[0038] S2.1 Data Labeling: Collect environmental monitoring images, label target areas through human-computer collaboration, and output labeled datasets;

[0039] S2.2 Model Training: The model is built based on the TensorFlow / PyTorch framework, and the parameters are optimized using distributed training to generate the initial model;

[0040] S2.3 Model Tuning: The hyperparameters, including the learning rate and batch size, are adjusted using the Bayesian optimization algorithm to make the model's F1 value ≥ 0.85 and then stored in the algorithm repository.

[0041] S2.4 Model Push: The node management unit selects suitable models based on the edge node hardware parameters, pushes them via HTTP / 2 protocol, and calculates MD5 hash values ​​for edge node verification.

[0042] Furthermore, step S3 includes:

[0043] S3.1 Video Processing: The video processing unit receives RTSP / HTTP video streams, decodes them into single-frame images using FFmpeg, and uses a perceptual hashing algorithm to remove duplicates, retaining the clearest images with PSNR≥30dB and SSIM≥0.9;

[0044] S3.2 Image Optimization: Gaussian filtering is applied to the selected images for noise reduction, and the images are normalized to 640×480 pixels to output optimized images;

[0045] S3.3 Parallel Inference: The model inference unit loads three models for earthwork exposure recognition, construction waste recognition, and smoke recognition, which run in parallel. The FPGA accelerates feature extraction and outputs the coordinates of the recognition box and the confidence score.

[0046] S3.4 Alarm Trigger: When the identification result meets the alarm conditions, the communication and alarm unit encrypts the alarm information and pushes it to the cloud platform.

[0047] Furthermore, step S4 includes:

[0048] S4.1 Alarm Reception: The cloud platform image alarm processing unit receives encrypted alarm information, decrypts it, and stores it in the MinIO distributed system;

[0049] S4.2 Personnel Allocation: Based on the alarm type and geographical location, automatically allocate qualified personnel for offline verification;

[0050] S4.3 Result Feedback: Offline personnel upload processing reports containing before-and-after comparison images of rectification, and the cloud platform generates closed-loop records by associating the original alarm information;

[0051] S4.4 Blockchain Evidence Storage: The operation record collection unit collects data such as model push and alarm processing in real time, encodes it into JSON format, and writes it into the blockchain through BFT-SMaRT consensus.

[0052] Furthermore, step S5 includes:

[0053] S5.1 Record Retrieval: Users can input an operation ID or time range through the traceability query unit, and the system will retrieve the corresponding operation record from the blockchain;

[0054] S5.2 Integrity Verification: Automatically calculate the SHA-256 hash value of the evidence record, and if it matches the value stored on the blockchain, it is determined that the record has not been tampered with;

[0055] S5.3 Link Display: Visualizes the entire operation process and supports exporting audit reports.

[0056] The present invention has the following beneficial effects:

[0057] 1. Improved data processing efficiency

[0058] This invention employs FFmpeg decoding (1 frame per second) and perceptual hashing for deduplication at its edge nodes, preserving clear images with PSNR ≥ 30dB and SSIM ≥ 0.9. It supports parallel inference of three models (FPGA accelerated), with inference time ≤ 200ms / frame. The daily video stream processing time of a single edge node is increased to more than 50 times that of traditional manual processing; inference latency is reduced to 1 / 1500 of the manual recognition response time (approximately 5-10 minutes), enabling real-time alarms.

[0059] 2. Model safety

[0060] The entire process of model push and alarm image push adopts SSL / TLS encrypted communication, SHA-256 hash encryption, and MD5 integrity verification, which can ensure model security and data security, and improve device security.

[0061] 3. Cost control

[0062] The environmental scene recognition model features localized inference at each node, covering the entire process from "data annotation - model development - localized inference - alarm handling - blockchain evidence storage," replacing manual identification and recording. This reduces view data upload bandwidth by over 90%, automates processes to replace 80% of manual workload, and lowers annual environmental supervision costs by 60%-70%.

[0063] 4. Traceability

[0064] The entire process of blockchain-based evidence storage supports retrieval based on operation ID / time range, and automatic verification of integrity using hash values; therefore, operation records cannot be tampered with, which is beneficial for liability determination and significantly reduces the time required for tracing. Attached Figure Description

[0065] Figure 1 This is a schematic diagram of the structure of the cloud-edge algorithm warehouse device for environmental protection supervision in the power industry according to the present invention.

[0066] Figure 2 This is a schematic diagram of the process for developing the environmental monitoring model of this invention.

[0067] Figure 3 This is a flowchart illustrating the cloud-edge collaborative processing method of the present invention. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] Figure 1 The diagram shown is a structural schematic of the cloud-edge algorithm warehouse device for environmental protection supervision in the power industry according to the present invention. The device includes: an algorithm warehouse cloud platform module, an edge node platform module, a blockchain evidence storage platform module, and an encrypted communication link. The hardware configuration, software functions, and processing flow of each module are as follows:

[0070] The algorithm warehouse cloud platform module consists of multiple cloud computing servers (including Intel Xeon CPUs and NVIDIA A100 GPU clusters), distributed storage devices, and 10Gbps Ethernet switches, deployed in a data center. Software configuration: It runs a Linux operating system, the TensorFlow / PyTorch deep learning framework, Kafka message queues, the MinIO distributed storage system, and a web management interface.

[0071] Figure 2 The diagram shows the entire process of "Environmental Protection Supervision Model Development" in the Algorithm Warehouse Cloud Platform module, which includes seven key steps: "Environmental Protection Supervision Image Library → Image Annotation → Project Creation → Model Development → Model Training → Model Optimization → Model Library Storage".

[0072] The algorithm warehouse cloud platform module is the core control center of the device, including:

[0073] Data Labeling Unit: Provides a professional set of data labeling tools and workflows, supporting labeling tasks for various data types (such as images, videos, and text). Employing a human-computer collaboration approach, combining automated labeling algorithms with manual review mechanisms, it ensures the accuracy and consistency of labeled data, providing high-quality datasets for subsequent model training.

[0074] Model Development Unit: This unit provides a model development environment based on mainstream deep learning frameworks (such as TensorFlow and PyTorch), supporting the design and implementation of various model architectures. It offers a rich library of algorithms and code templates, facilitating developers to quickly build customized machine learning and deep learning models.

[0075] Model training unit: Leveraging the powerful computing resources of the cloud platform, it enables efficient parallel training on large-scale data. It supports distributed training and multi-GPU acceleration, improving model training efficiency and performance through optimized training algorithms and parameter tuning strategies.

[0076] Model tuning unit: Utilizes automated tuning algorithms (such as random search, grid search, Bayesian optimization, etc.) to optimize the model's hyperparameters. Combined with model evaluation metrics (such as accuracy, recall, F1 score, etc.), it monitors model performance in real time and continuously adjusts the model structure and parameters to achieve the best prediction results.

[0077] Model storage unit: Provides a centralized model storage and management platform, supporting model version control, uploading, downloading, and sharing. Employing metadata management and indexing technologies, it facilitates developers in quickly finding and using the models they need.

[0078] Edge node management unit:

[0079] 1. Establish a database for edge nodes, recording basic information (such as device type, geographical location, hardware configuration, etc.) and operating status of each edge node. Enable remote monitoring, configuration, and management of edge nodes through a visual management interface.

[0080] 2. A high-efficiency file transfer protocol is used to push the trained model from the cloud platform to edge nodes. Incremental updates and version rollback mechanisms are supported to ensure that the model on the edge nodes is always up-to-date.

[0081] 3. Edge node heartbeat detection function: This function periodically sends heartbeat packets to edge nodes to monitor their online status and network connectivity in real time. When an edge node anomaly is detected, an alarm is issued promptly and appropriate measures are taken to ensure system stability and reliability.

[0082] The image alarm processing unit receives inference alarm images uploaded by edge nodes in real time via message queues (such as Kafka, RabbitMQ, etc.). It uses a distributed storage system (such as MinIO) to store and manage the images, facilitating subsequent querying and analysis.

[0083] The edge node platform module is deployed on edge devices at the monitoring site, including an edge computing card, a video capture card (supporting 8 video stream inputs), and a 4G / 5G communication module. Software configuration: It runs an Ubuntu embedded system, utilizing the FFmpeg video decoding library, the TensorFlow Lite inference framework, a lightweight message queue (MQTT), a video processing unit, a model inference unit, and a communication and alarm unit.

[0084] This module is a localized data processing terminal and includes the following units:

[0085] Video processing unit: Supports multiple video input formats (such as RTSP, HTTP, USB camera, etc.), and connects video streams to the edge node platform through the video acquisition module; calls the FFmpeg library to decode the video stream into single-frame images (1 frame extracted per second), calculates image fingerprints through perceptual hashing algorithm, and retains the clearest images with PSNR≥30dB and SSIM≥0.9 in the repeating sequence after deduplication; performs noise reduction (Gaussian filtering) and normalization (size adjustment to 640×480) processing on the selected images to enhance the model recognition effect.

[0086] Model Inference Unit: Monitors model push messages from the cloud platform, verifies push permissions using the RSA algorithm, calculates the MD5 hash value of the model file upon receipt, and stores it on the local SSD if it matches the value provided by the cloud; loads the model into the TensorFlow Lite framework, configures FPGA to accelerate feature extraction, and performs a health check during initialization (inference time ≤200ms indicates normal operation); supports parallel operation of 3 models (such as exposed earthwork recognition, construction waste recognition, and smoke recognition), performs target detection on optimized images, and outputs the coordinates of the recognition box and confidence score (≥0.8 triggers an alarm).

[0087] Communication and alarm unit: Sends a heartbeat packet (including CPU usage, GPU load, and network signal strength) to the cloud platform every 30 seconds and receives cloud responses to maintain the connection; when the model recognition result meets the alarm conditions (such as exposed earth area > 20㎡), it pushes the alarm time, location, and image (AES encrypted) to the cloud platform through an encrypted channel and waits for cloud confirmation.

[0088] The blockchain evidence storage platform module consists of 5 consensus node servers (4 CPUs, 16GB memory, 1TB SSD) and 1 gateway server (2 CPUs, 8GB memory, 200GB hard drive). This module is an operation record evidence storage terminal and includes the following units:

[0089] Operation Log Collection Unit: Log agents are embedded in key operation points such as model development and push nodes on the cloud platform, model deployment and alarm triggering on edge nodes, to collect operation information (time, operator ID, operation content, and result) in real time and send it asynchronously to the encrypted evidence storage unit via the MQTT protocol.

[0090] Encrypted Evidence Storage Unit: Converts the collected operation records into JSON format (e.g., {"Operation Type":"Model Push","Object":"Earthwork Recognition V1.2","Time":"2025-02-01 12:03:11"}); achieves consensus among 5 nodes using the BFT-SMaRT algorithm (≥3 nodes pass verification), packages the encoded information into blocks (including the hash value of the previous block), and writes it into the distributed ledger; uses the SHA-256 algorithm to encrypt the blocks, so any modification to the record by any node will cause a change in the hash value, which will be rejected by other nodes.

[0091] Traceability and Query Unit: Provides a query interface based on a blockchain explorer, supporting the retrieval of evidence records by conditions such as operation ID, time range, and operator, displaying the entire operation process (such as "model development → push → edge deployment → alarm handling"), and automatically verifying the integrity of the record's hash value.

[0092] The functions of the encrypted communication link include:

[0093] When edge nodes register, permissions are verified using the RSA algorithm, and the cloud platform assigns a unique identifier and encryption key.

[0094] The model file and alarm information are transmitted using SSL / TLS protocol encryption. The edge node sends a heartbeat packet every 30 seconds, and the cloud platform triggers a reconnection mechanism if it does not receive a response.

[0095] Figure 3 The diagram illustrates the 11 key steps of the collaborative processing method of this invention, presented in a step-by-step manner: "S1 Edge Node Registration → S2 Cloud Distribution Model → S3 Edge Node Receiving Model → S4 Edge Node Video Access → S5 Edge Node Deployment and Startup Model → S6 Video Data Processing → S7 Edge Node Model Inference → S8 Alarm Push → S9 Cloud Platform Alarm Reception → S10 Alarm Handling → S11 Report Generation." The specific process is as follows:

[0096] When the edge node service initializes and starts, it automatically triggers an operation to send a registration request to the cloud platform. This request contains detailed basic information about the edge node itself, such as device model, hardware configuration (including CPU model and number of cores, memory capacity, storage type and capacity, etc.), operating system version, and related configuration parameters, such as network communication protocols (TCP / IP, UDP, etc.), data transmission rate requirements, security encryption policies, etc.

[0097] Upon receiving the registration request, the cloud platform immediately initiates the verification process:

[0098] First, the information in the request is formatted according to preset rules to ensure the accuracy and integrity of the data;

[0099] Next, the edge node is compared with a pre-set whitelist or authentication database to verify its legitimacy.

[0100] If verification is successful, the cloud platform will assign a unique identifier to the edge node, in the format BYJD202502011203110000X, where the last five digits are sequentially incremented based on the number of registered edge nodes. Simultaneously, based on the edge node's type, purpose, and security level requirements, appropriate access permissions will be assigned, clearly defining the edge node's access scope and operational authority to cloud platform resources. Subsequently, an encrypted communication link based on the Secure Sockets Layer (SSL / TLS) protocol will be established to ensure the security and reliability of data transmission between the edge node and the cloud platform.

[0101] After completing registration and establishing a communication link, the edge node will periodically send heartbeat packets to the cloud platform at preset time intervals (such as every 30 seconds). The heartbeat packet contains real-time operating status information of the edge node, such as CPU utilization, memory usage, disk I / O status, etc., as well as network connection status information, such as network latency, packet loss rate, etc.

[0102] Upon receiving a heartbeat packet, the cloud platform immediately returns a heartbeat response. Edge nodes check the receipt and content of the heartbeat response to ensure real-time communication and synchronization with the cloud platform. If a heartbeat response is not received within a specified time, the corresponding exception handling mechanism will be triggered.

[0103] The entire process of the above collaborative processing method is recorded in the blockchain evidence storage platform. Later, a query interface based on a blockchain explorer will be provided to support searching evidence storage records by operation ID, time range, operator, and other conditions, and to display the entire collaborative operation process.

[0104] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A cloud-edge algorithm warehouse device for environmental monitoring in the power industry, characterized in that, include: The Algorithm Warehouse Cloud Platform module, as the core control center in the cloud, integrates a data annotation unit, a model development unit, a model training unit, a model optimization unit, a node management unit, and an image alarm processing unit. It is used to receive edge node registration requests, develop and train environmental monitoring models, push models to edge nodes, and process alarm information. The edge node platform module, deployed at the monitoring site, includes a video processing unit, a model inference unit, and a communication and alarm unit. It is used to collect video streams and filter clear images, perform local inference based on cloud-pushed models, and trigger real-time alarms. The blockchain evidence storage platform module is a distributed ledger system that includes an operation record collection unit, an encrypted evidence storage unit, and a traceability query unit. It is used to collect and encrypt the entire process operation data of the cloud platform and edge nodes. An encrypted communication link, employing the SSL / TLS protocol and heartbeat mechanism, connects the algorithm warehouse cloud platform module, edge node platform module, and blockchain evidence storage platform module, enabling secure data transmission and status synchronization among the three.

2. The apparatus according to claim 1, characterized in that, The various units of the algorithm warehouse cloud platform module are connected sequentially through data interfaces; The data annotation unit outputs annotation data to the model development unit, which then generates a model structure file and transmits it to the model training unit. The model training unit trains the model using labeled data and outputs the initial model to the model tuning unit. After optimizing the model parameters, the model tuning unit stores the qualified model in the model storage unit; The edge node management unit retrieves the adapted model from the model storage unit and pushes it to the edge node platform module via an encrypted communication link. The image alarm processing unit receives inference alarm images uploaded by edge nodes in real time via a message queue.

3. The apparatus according to claim 1, characterized in that, The data flow direction of each unit in the edge node platform module is as follows: The video processing unit receives the video stream, decodes and filters it, and outputs optimized images with PSNR≥30dB and SSIM≥0.9 to the model inference unit. The model inference unit loads the model pushed from the cloud, performs parallel inference on the optimized image, and outputs a recognition result with a confidence level ≥ 0.8 to the communication and alarm unit. The communication and alarm unit encrypts the results that meet the alarm conditions and pushes them to the algorithm warehouse cloud platform module through the encrypted communication link.

4. The apparatus according to claim 1, characterized in that, The workflow of the blockchain-based evidence storage platform module is as follows: The operation log acquisition unit collects real-time operation data for model uploading / downloading / training / deployment, edge node registration, and alarm handling. The encrypted evidence storage unit encodes the operation data into JSON format, writes it into the blockchain using the BFT-SMaRT consensus algorithm, and encrypts the blocks using the SHA-256 algorithm. The traceability query unit provides a retrieval interface based on operation ID and time range, outputs evidence records containing the entire operation chain, and automatically verifies the integrity of hash values.

5. The apparatus according to claim 1, characterized in that, The encrypted communication link includes: When edge nodes register, permissions are verified using the RSA algorithm, and the cloud platform assigns a unique identifier and encryption key. The model file and alarm information are transmitted using SSL / TLS protocol encryption. The edge node sends a heartbeat packet every 30 seconds, and the cloud platform triggers a reconnection mechanism if it does not receive a response.

6. A cloud-edge collaborative processing method for environmental monitoring algorithms in the power industry, characterized in that... Includes the following steps: S1. System Initialization: Edge nodes send a registration request to the cloud platform via an encrypted communication link. After the cloud platform verifies the request, a secure connection is established. S2. Model Development and Deployment: Develop and train environmental monitoring models on the cloud platform, push them to edge nodes, and verify their integrity; S3. Localized Inference and Alarm: Edge nodes process video streams and infer based on models, and push alarms to the cloud platform after triggering them; S4. Alarm Handling and Evidence Preservation: The cloud platform assigns personnel to handle alarms offline, and the entire process of blockchain-based synchronous evidence preservation is carried out. S5. Traceability and Auditing: Verify the legitimacy of operation records through blockchain traceability query units.

7. The method according to claim 6, characterized in that, Step S2 includes: S2.1 Data Labeling: Collect environmental monitoring images, label target areas through human-computer collaboration, and output labeled datasets; S2.2 Model Training: The model is built based on the TensorFlow / PyTorch framework, and the parameters are optimized using distributed training to generate the initial model; S2.3 Model Tuning: The hyperparameters, including the learning rate and batch size, are adjusted using the Bayesian optimization algorithm to make the model's F1 value ≥ 0.85 and then stored in the algorithm repository. S2.4 Model Push: The node management unit selects suitable models based on the hardware parameters of the edge nodes, pushes them via HTTP / 2 protocol, and calculates MD5 hash values ​​for edge nodes to verify.

8. The method according to claim 6, characterized in that, Step S3 includes: S3.1 Video Processing: The video processing unit receives RTSP / HTTP video streams, decodes them into single-frame images using FFmpeg, and uses a perceptual hashing algorithm to remove duplicates, retaining the clearest images with PSNR≥30dB and SSIM≥0.9; S3.2 Image Optimization: Gaussian filtering is applied to the selected images for noise reduction, and the images are normalized to 640×480 pixels to output optimized images; S3.3 Parallel Inference: The model inference unit loads three models for earthwork exposure recognition, construction waste recognition, and smoke recognition, which run in parallel. The FPGA accelerates feature extraction and outputs the coordinates of the recognition box and the confidence score. S3.4 Alarm Trigger: When the identification result meets the alarm conditions, the communication and alarm unit encrypts the alarm information and pushes it to the cloud platform.

9. The method according to claim 6, characterized in that, Step S4 includes: S4.1 Alarm Reception: The cloud platform image alarm processing unit receives encrypted alarm information, decrypts it, and stores it in the MinIO distributed system; S4.2 Personnel Allocation: Based on the alarm type and geographical location, automatically allocate qualified personnel for offline verification; S4.3 Result Feedback: Offline personnel upload processing reports containing before-and-after comparison images of rectification, and the cloud platform generates closed-loop records by associating the original alarm information; S4.4 Blockchain Evidence Storage: The operation record collection unit collects data such as model push and alarm processing in real time, encodes it into JSON format, and writes it into the blockchain through BFT-SMaRT consensus.

10. The method according to claim 6, characterized in that, Step S5 includes: S5.1 Record Retrieval: Users can input an operation ID or time range through the traceability query unit, and the system will retrieve the corresponding operation record from the blockchain; S5.2 Integrity Verification: Automatically calculate the SHA-256 hash value of the evidence record, and if it matches the value stored on the blockchain, it is determined that the record has not been tampered with; S5.3 Link Display: Visualizes the entire operation process and supports exporting audit reports.

Images