Edge cloud intelligent agent cooperative networking system for regional individual environmental exposure monitoring

Abstract

The application discloses a regional individual environmental exposure monitoring edge cloud intelligent agent cooperative networking system, relates to the cross field of environmental monitoring and edge intelligence technology, and comprises the following: an edge intelligent agent deployed on each individual exposure monitoring device and performing an autonomous AI agent; a central intelligent agent deployed on a cloud server cluster; and a communication protocol layer adopting a dual-channel communication architecture separating uplink HTTPS batch transmission from downlink MQTT long connection. The application simultaneously meets multiple requirements of large-scale device management and control, high-precision data quality control, low-threshold intelligent operation and maintenance, safe version iteration and upgrade, and multi-role collaborative management.

Description

Technical Field

[0001] This invention relates to the intersection of environmental monitoring and edge intelligence technology, and more specifically, to an edge-cloud intelligent agent collaborative networking system for monitoring individual environmental exposure in a region. Background Technology

[0002] Individual environmental exposure monitoring is a cutting-edge research direction in the field of environmental health. It can collect real individual exposure data of air pollutants such as PM2.5, O3, NO2, and CO through portable sensors, providing core data support for environmental epidemiological research and precision health management. Currently, a single monitoring device, relying on sensing and edge computing technologies, can achieve basic local data collection, quality control, and anomaly diagnosis, but it can only be adapted to small-scale monitoring scenarios. In large-scale device network monitoring scenarios, existing technologies have significant shortcomings: Traditional IoT platforms adopt a centralized architecture of "dumb terminals + intelligent cloud," where edge devices lack autonomy and cannot perform self-control when the network is abnormal, easily leading to data distortion and loss. Furthermore, the computing power and bandwidth costs of centrally processing massive amounts of raw data in the cloud are high, and the professional threshold for device operation and maintenance is high. Existing multi-agent systems are mostly applied to virtual simulation scenarios, lacking dedicated collaborative capabilities for environmental monitoring. They cannot achieve spatiotemporal consistency verification of monitoring data, cross-device domain data quality control, or natural language intelligent operation and maintenance, making them unsuitable for the large-scale collaborative operation needs of sensor networks. Simultaneously, distributed AI technologies such as federated learning, which focus on privacy-preserving training through model parameter transmission, cannot meet the core requirements of complete data transmission and analysis in this field, and are therefore unsuitable for this monitoring scenario. In summary, the industry currently lacks a monitoring and management solution for edge and cloud-based intelligent agent collaborative networking, failing to simultaneously address the diverse needs of large-scale device control, high-precision data quality control, low-threshold intelligent operation and maintenance, secure version iteration, and multi-role collaborative management. Summary of the Invention

[0003] To address the aforementioned issues, the present invention aims to provide an edge-cloud intelligent agent collaborative networking system for monitoring individual environmental exposure in a region. This system is designed to simultaneously meet multiple requirements, including large-scale equipment management and control, high-precision data quality control, low-threshold intelligent operation and maintenance, secure version iteration and upgrades, and multi-role collaborative management.

[0004] To achieve the above technical objectives, this application provides an edge-cloud intelligent agent collaborative networking system for regional individual environmental exposure monitoring, including: Edge agents, deployed on each individual exposure monitoring device and acting as autonomous AI agents, communicate with the central agent via a wireless network; The central intelligent agent, deployed on a cloud server cluster, executes the central AI agent to coordinate and schedule all online edge intelligent agents by setting up a device status knowledge graph, a spatiotemporal consistency verification engine, a conversational management engine, a task orchestration and canary release engine, and a user permission management module. The communication protocol layer is used to employ a dual-channel communication architecture that separates uplink HTTPS bulk transmission from downlink MQTT long connections, supplemented by a periodic heartbeat mechanism to sense and monitor the real-time status of the device.

[0005] Preferably, the edge agent also includes a device fingerprint module, which serves as a unique identifier for the central agent to identify and manage the monitoring device. The fingerprint generation logic is loaded in the form of a skill package and is uniformly managed and upgraded by the central agent.

[0006] Preferably, the edge agent further includes a local data preservation module for storing the raw sensor data of the edge agent using a partitioned storage strategy to ensure that the raw sensor data is not lost under any network conditions.

[0007] Preferably, the edge agent further includes a communication management module for communicating with the central agent, including: Heartbeat transmission is used to send a device status summary to the center at a specified frequency. Data upload is used to upload sensor data in batches via HTTPS. Command reception is used to receive commands issued by the central intelligent agent via an MQTT long connection; Network outage detection and recovery is used to monitor network status. When the network is out of service, it automatically switches to local security mode and uploads backlogged data according to priority after recovery.

[0008] Preferably, the edge agent further includes an intelligent adaptive policy uploading module, used to automatically switch working modes based on network status and the presence of abnormal events, wherein the working modes include four types: Normal mode: All data within the cycle is uploaded in batches every 5 minutes, and the confirmed queue data is cleared after successful upload; Alarm mode: Upload data one by one in real time, set abnormal data as high priority, and switch back to normal mode after the abnormality is eliminated; Offline mode: Data is stored in a local buffer, and abnormal data is additionally stored in a priority queue. Network connection is continuously retried according to an exponential backoff strategy. Recovery mode: Prioritize uploading abnormal high-priority data, then upload ordinary backlog data in reverse order, adaptive bandwidth transmission, and upload all data.

[0009] Preferably, the central intelligent agent is used to complete the full-link modeling through a device status knowledge graph, which consists of three layers of entities and four types of relationships: using a geographical region as the spatial carrier to mount all edge intelligent devices; constructing the adjacency topology between devices based on GPS coordinates; binding all alarm events generated by the devices; and then performing anomaly pattern similarity association on all alarms, ultimately forming a complete edge intelligent device association knowledge network covering geographical space, device entities, operational alarms, and anomaly associations.

[0010] Preferably, the central intelligent agent is used to construct a spatiotemporal consistency verification engine by utilizing a knowledge graph of the spatiotemporal distribution of pollutants and a cross-device spatiotemporal consistency verification algorithm. The cross-device spatiotemporal consistency verification algorithm first identifies single-point data deviations through spatial neighborhood statistics, and then distinguishes between equipment failures and real pollution phenomena through multi-layer business verification, while taking into account data quality control and environmental event capture, so as to verify the uploaded sensor data records.

[0011] Preferably, the central intelligent agent is used to generate a conversational management engine by invoking the ToolUse mechanism through a large language model and tools.

[0012] Preferably, the central intelligent agent is used to execute configuration change tasks and skill package canary release tasks through a task orchestration and canary release engine. When executing the skill package canary release task, the canary verification, canary expansion and full push operations are performed in sequence to execute the skill package canary release task. If the data integrity rate of upgraded devices decreases by more than 10%, or the frequency of abnormal alarms on upgraded devices increases by more than 200%, or any upgraded device reports a skill loading failure log / alarm, a global automatic hard rollback operation will be triggered. Meanwhile, skipping stages is prohibited if the required monitoring duration is not completed at each stage; after automatic rollback execution, root cause analysis and version repair must be completed before the canary rollout process can be restarted; batch one-click rollback is no longer supported after snapshots expire, only manual downgrade of a single machine is supported.

[0013] Preferably, the edge agent adopts a GPS+NTP dual-source time synchronization strategy: Master clock source: UTC time provided by the GPS module; Backup clock source: 4G network NTP synchronization; Synchronization cycle: Calibrate every 300 seconds; Each data record carries a timestamp and clock source identifier, which are used by the central intelligent agent to evaluate time quality.

[0014] The present invention discloses the following technical effects: This invention enables efficient networking and collaboration between edge intelligent agents and cloud-based central intelligent agents, possessing significant technical advantages and beneficial effects.

[0015] This invention, while ensuring the complete transmission, retention, and analysis of original monitoring data, enables unified registration, status monitoring, and full lifecycle management of large-scale edge monitoring devices, effectively supporting large-scale, large-scale network operations for individual environmental exposure monitoring.

[0016] By incorporating knowledge from atmospheric and environmental sciences, this invention can perform cross-device data quality verification and spatiotemporal consistency correction, significantly improving the overall monitoring data accuracy and reliability of large-scale sensor networks.

[0017] This invention supports natural language-driven device management and intelligent operation and maintenance, which lowers the technical threshold for system operation and maintenance, and allows routine device management operations to be completed without the need for professional technicians.

[0018] This invention enables secure gray-scale release, batch upgrades, and anomaly rollback of monitoring skill packages, ensuring the security and stability of system iteration updates. It also features a multi-role hierarchical permission management mechanism, supporting multi-user collaborative monitoring management and scientific research analysis. This invention comprehensively solves the problem of insufficient adaptability of traditional IoT platforms, existing multi-agent architectures, and federated learning distributed technologies in individual environment exposure monitoring scenarios. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the system architecture described in this invention.

[0021] Figure 2 This is the spatiotemporal consistency verification process described in this invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0023] like Figure 1 As shown, the present invention provides an edge-cloud intelligent agent collaborative networking system for monitoring individual environmental exposure in a region, which consists of an edge intelligent agent layer, a central intelligent agent layer and a communication protocol layer.

[0024] In one implementation, the edge agent layer is an autonomous AI agent deployed on each individual exposure monitoring device. It has capabilities such as sensor data acquisition, local quality control, sensor health self-diagnosis, and local data preservation. It communicates with the central agent via a 4G wireless network. Each edge agent has a unique device fingerprint and supports hot-swappable skill packages.

[0025] For example, each individual exposure monitoring device deploys an edge agent, and in network mode, the edge agent also has the following components: (1) Device fingerprint module: Each device generates a unique device fingerprint, which is generated by combining and hashing the following elements: DeviceFingerprint=SHA-256( hardware_serial+ / / Main control board hardware serial number sensor_config_hash+ / / Summary of current sensor configuration firmware_version + / / Firmware version number first_boot_timestamp / / First boot timestamp ) For example, a device fingerprint is generated upon initial registration and serves as a unique identifier for the central agent to identify and manage the device. The fingerprint generation logic is loaded as a skill package (referencing the P2 skill management system) and is uniformly managed and upgraded by the central agent.

[0026] (2) Local data preservation module: The edge agent adopts a partitioned storage strategy to ensure that the original sensor data is not lost under any network conditions, as shown in Table 1.

[0027] Table 1

[0028] Taking a 128GB memory card as an example, the data buffer is approximately 96GB. Assuming each sensor record is approximately 2KB and the sampling interval is 5 minutes: the daily data volume = 24h × 60min ÷ 5min × 2KB = 576KB / day; 96GB ÷ 576KB / day ≈ 174,000 days ≈ far exceeding 180 days (including redundancy).

[0029] For example, the above is the minimum data volume for a single channel. In reality, it includes additional information such as 12 sensor channels, GPS coordinates, quality control marks, and device status. Each record is about 8-12KB. Calculated at 10KB: Daily data volume = 288 records × 10KB = 2.88MB / day; 96GB ÷ 2.88MB / day ≈ 34,000 days >> 180 days.

[0030] For example, even at the highest sampling frequency (1-minute interval): daily data volume = 1440 records × 10KB = 14.4MB / day; 96GB ÷ 14.4MB / day ≈ 6,800 days >> 180 days, so 128GB of storage can fully meet the data preservation requirements for 180 days.

[0031] (3) Communication management module, which is responsible for communication with the central intelligent agent, including: heartbeat sending: sending device status summary to the center every 60 seconds (via MQTT connection); data upload: uploading sensor data in batches via HTTPS; instruction reception: receiving instructions issued by the central intelligent agent via MQTT long connection; network outage detection and recovery: monitoring network status, automatically switching to local preservation mode when the network is out of service, and uploading backlogged data according to priority after recovery.

[0032] (4) Intelligent adaptive strategy upload module, including an upload strategy that is dynamically adjusted by the edge agent based on network status and data characteristics: State machine definition: NORMAL mode (normal network, no abnormal events): Batch upload every 5 minutes; each upload includes all data records within the last 5 minutes; after the upload is completed, the confirmed data in the priority queue is released.

[0033] ALERT mode (anomaly detected): Switches to real-time upload (each piece of data is uploaded immediately); data related to the anomaly is marked as high priority; after the anomaly ends, it reverts to NORMAL mode.

[0034] OFFLINE mode (network disconnection): All data is written to the local data buffer; abnormal event data is simultaneously written to the priority upload queue; network reconnection is continuously attempted (exponential backoff: 1s, 2s, 4s, ... maximum 300s).

[0035] RECOVERY mode (network recovery): First priority: Upload abnormal event data in the priority queue; Second priority: Upload regular data (in reverse chronological order, latest first); Bandwidth adaptive: Dynamically adjust the upload rate based on 4G signal strength; Switch back to NORMAL mode after all backlogged data has been uploaded.

[0036] In one implementation, the intelligent adaptive strategy upload module automatically switches its operating mode based on network status and the presence of abnormal events. The operating modes include four types: Normal mode: All data within the specified period is uploaded in batches every 5 minutes; the confirmed queue data is cleared after successful upload; Alarm mode: Data is uploaded one by one in real time; abnormal data is prioritized; the system switches back to normal mode after the anomaly is resolved; Offline mode: Data is stored in a local buffer; abnormal data is additionally stored in a priority queue; network retry is continuously performed according to an exponential backoff strategy; Recovery mode: Abnormal high-priority data is uploaded first, then ordinary backlog data is uploaded in reverse order, with adaptive bandwidth transmission, uploading all data.

[0037] In one implementation, the central agent layer is deployed on a cloud server cluster as a central AI agent, acting as the "general coordinator" of the entire agent network.

[0038] In one embodiment, the central intelligent agent layer has the following core subsystems, including: a device state knowledge graph, a spatiotemporal consistency verification engine, a conversational management engine, a task orchestration and canary release engine, and a user permission management module.

[0039] For example, a device state knowledge graph is a structured knowledge representation used to maintain information such as the real-time state, geographic location, sensor configuration, and health status of all online edge agents; For example, the central agent maintains a continuously updated device state knowledge graph, with the following storage structure: Knowledge graph node types: DeviceNode: One node corresponds to each edge agent; Attributes: device_id, fingerprint, region, gps, online_status, battery,storage,last_heartbeat,sensors_config, health_status,installed_skills,registration_time; RegionNode: Geographic region node; Attributes: region_id, name, bounds_polygon, device_count; AlertNode: Alert event node; Attributes: alert_id, device_id, alert_type, severity, timestamp; Knowledge graph relationship types: BELONGS_TO:DeviceNode->RegionNode; NEIGHBORS: DeviceNode <-> DeviceNode (spatial distance < threshold R); HAS_ALERT: DeviceNode->AlertNode; SIMILAR_PATTERN: AlertNode <-> AlertNode (same type of exception mode).

[0040] For example, this image recognition map is constructed for three core entities: edge intelligent devices, geographical regions, and device alarms. It includes three types of node entities and four types of association relationships, fully depicting the spatial affiliation, spatial adjacency, alarm association, and abnormal pattern similarity logic of edge intelligent devices. The specific entities and association rules are explained as follows: For example, the graph node (entity) definition is as follows: 1. DeviceNode: Each independent edge intelligent device in the graph is mapped to a separate device node, used to store all basic information, operating status, configuration and business capability data of the device. Built-in attribute fields are as follows: unique identifier device_id, device hardware fingerprint, region identifier, device GPS location (gps), online status (online_status), remaining battery power (battery), storage space (storage), last heartbeat reporting time (last_heartbeat), sensor configuration parameters (sensors_config), device overall health status (health_status), installed business skills set (installed_skills), and device registration time (registration_time).

[0041] 2. RegionNode: Geographic region node: Constructs regional entity nodes based on administrative / planned geographical scope. It is used to uniformly manage the spatial boundaries of the region and the statistical information of the devices under its jurisdiction. The attributes include: region_id (unique region ID), region name (name), region boundary polygon coordinates (bounds_polygon), and the total number of devices in the current region (device_count).

[0042] 3. AlertNode Device Alarm Event Node: Each abnormal alarm triggered by a device generates a separate alarm entity node, recording complete alarm event information. Attributes include: unique alarm ID (alert_id), device number to which the alarm belongs (device_id), alarm category type (alert_type), alarm severity level (severity), and alarm timestamp.

[0043] For example, the graph relationship (entity association) is defined as follows: 1. BELONGS_TO Attribution Relationship: Directed unidirectional association, from DeviceNode to the corresponding RegionNode, semantically meaning "this edge smart device belongs to a certain geographical region", realizing the binding mapping between the device and the geographic space.

[0044] 2. NEIGHBORS Spatial Adjacency Relationship: Undirected bidirectional mutual association, establishing a connection between two DeviceNode devices; the determination rule is that if the GPS spatial straight-line distance between two devices is less than a preset distance threshold R, it means that the two edge agents are spatially adjacent devices.

[0045] 3. HAS_ALERT generates alarm relationships: a directed unidirectional association, where the DeviceNode points to the AlertNode generated by itself, with the semantics "this edge device generated this alarm event", binding the device with all its own alarm records.

[0046] 4. SIMILAR_PATTERN Abnormal Pattern Similarity Relationship: Undirected bidirectional mutual correlation, connecting two AlertNode alarm nodes; the judgment rule is that two alarms belong to the same type of abnormal operation mode, which is used to explore the correlation characteristics of abnormal alarms of the same source and type under multiple devices and multiple time periods.

[0047] In one implementation, the entire knowledge graph completes full-link modeling through three layers of entities and four types of relationships: using geographical regions as spatial carriers to mount all edge intelligent devices; constructing the adjacency topology between devices based on GPS coordinates; binding all alarm events generated by the devices; and then performing anomaly pattern similarity association on all alarms, ultimately forming a complete edge intelligent device association knowledge network covering geographical space, device entities, operational alarms, and anomaly associations.

[0048] For example, the spatiotemporal consistency verification engine is used to perform quality verification on cross-device data based on a knowledge graph of the spatiotemporal distribution characteristics of pollutants / aerosols, and can continuously self-optimize through the accumulation of literature knowledge and operational data.

[0049] For example, the spatiotemporal consistency verification process is as follows: Figure 2 As shown. The engine contains two components that work together: Component A: Spatial-Temporal Distribution Knowledge Graph of Pollutants: This knowledge graph encodes the spatiotemporal distribution characteristics of atmospheric pollutants and aerosols under different conditions. The initialization of the knowledge graph is based on empirical knowledge from atmospheric science literature, such as: PM2.5 has a spatial correlation scale of approximately 2-5 km under stable weather conditions (literature reference); NO2 has a spatial correlation scale of approximately 0.5-1 km due to the locality of traffic sources (literature reference); O3, as a secondary pollutant, has high spatial homogeneity at the regional scale, with a correlation scale reaching 10-20 km (literature reference).

[0050] Self-evolution mechanism: During operation, the spatiotemporal consistency verification engine continuously counts the spatial correlation index of the actual observation data and updates the knowledge graph weights using an exponentially weighted moving average: W_new=α×W_observed+(1-α)×W_prior, where: W_new is the updated weight; W_observed is the actual value obtained from the observation data of the past 30 days; W_prior is the current value in the knowledge graph; α is the learning rate (default 0.1, which can be adjusted by the administrator).

[0051] Component B: Cross-device spatiotemporal consistency verification algorithm: Perform the following verification process for each sensor data record uploaded to the center: Algorithm: Cross-device spatiotemporal consistency verification; Input: Observation vector X_i(t) of device i at time t.

[0052] Step 1: Constructing a dynamic spatial neighborhood: For each pollutant channel p: R_p=KG.getSpatialCorrelation(p,currentMetCondition,currentSeason); N_p(i) = {j | distance(i,j) < R_p and j is online}; Step 2: Calculate the local consistency score: If |N_p(i)|>=3: median_p=median({x_p(j,t)|j∈N_p(i)}); MAD_p=median({|x_p(j,t)-median_p||j∈N_p(i)}); S_p(i,t)=|x_p(i,t)-median_p| / max(MAD_p,ε); Step 3: Anomaly detection and classification: For channels where S_p(i,t) > θ (θ defaults to 3.0), perform real event exclusion checks: (a) time pattern check; (b) multi-channel correlation check; (c) knowledge graph cross-validation.

[0053] Overall judgment results: DEVICE_ANOMALY (device malfunction) → push maintenance alarm; ENVIRONMENTAL_EVENT (real-world event) → data is valid and marked as high value; UNCERTAIN (uncertain) → marked as requiring manual review.

[0054] For example, a conversational management engine is used to support managers in performing operations such as device querying, anomaly diagnosis, and task assignment through natural language, based on a large language model and tool invocation mechanism.

[0055] For example, this engine is implemented based on a large language model and a tool use mechanism. Workflow example: The administrator inputs: "How is the equipment operating in that area?" LLM generation tool call: →tool:query_devices(region="a region",fields=["status","alerts"]) Query results: { "region":"a certain region", "total_devices":856, "online":841, "offline":15, "alerts":[ {"device":"TJ-BH-023","type":"SO2_drift","severity":"WARNING"}, {"device":"TJ-NH-107","type":"offline_72h","severity":"CRITICAL"} ] } Natural language response: "There are a total of 856 devices in a certain region, and 841 of them are online (online rate of 98.2%)."

[0056] Two alarms require attention: 1. Zero drift is observed in the SO2 channel of TJ-BH-023 (Binhai New Area); calibration is recommended. 2. TJ-NH-107 (Nankai District) has been offline for over 72 hours; on-site investigation is recommended. The available toolset is shown in Table 2.

[0057] Table 2

[0058] For example, for a task orchestration and canary release engine, the canary release strategy used to support configuration changes and skill package upgrades includes three stages: canary verification, canary expansion, and full push.

[0059] For example, the task orchestration and canary release engine supports the orchestration and deployment of two types of tasks: Type A Configuration Change Task: Task Orchestration Process: 1. The administrator issues the intention: "Increase the sampling frequency of equipment in Binhai New Area to 1 minute for 3 days"; 2. The central agent generates an execution plan (including impact assessment); 3. The plan is submitted to the administrator for confirmation; 4. The configuration command is sent to the target device via MQTT; 5. Each edge agent executes the command immediately upon receiving it and replies with ACK; 6. The central agent monitors the execution status.

[0060] Type B skill pack gray-scale release tasks.

[0061] Three-stage canary release process: Phase 1 Canary Validation: Select 3-5 representative devices and monitor them continuously for 24 hours; Pass: No significant degradation in any indicator (change < 5%); Fail: Degradation of any indicator > 10% → Automatic rollback and termination of release.

[0062] Phase 2 Gray-scale expansion: Expand to 10% of the target devices (evenly distributed by area), continuously monitor for 48 hours, and use the same judgment criteria as Phase 1.

[0063] Phase 3 Full Push: Push to the remaining devices in batches (20% per batch, 1 hour apart); after the full push is completed, the previous version snapshot will be retained for 7 days, and a full rollback can be performed with one click within 7 days.

[0064] Automatic rollback trigger conditions (triggering a global automatic hard rollback operation): The data integrity rate of the upgraded device decreases by more than 10%; the frequency of abnormal alarms on the upgraded device increases by more than 200%; any upgraded device reports a skill loading failure.

[0065] For example, skipping a stage is prohibited if the specified monitoring duration is not completed in each stage; after automatic rollback execution, root cause analysis and version repair must be completed before the canary rollout process can be restarted; after the snapshot expires, batch one-click rollback is no longer supported, and only manual downgrade of a single machine is supported.

[0066] For example, the user permission management module is used to support hierarchical permission control for three roles: operation and maintenance administrator, researcher, and project manager.

[0067] For example, the present invention adopts a role-based access control (RBAC) model, defining three standard roles as shown in Table 3.

[0068] Table 3

[0069] For example, the edge agent and the central agent adopt a dual-channel communication architecture of uplink HTTPS + downlink MQTT.

[0070] In one implementation, the communication protocol layer adopts a dual-channel communication architecture that separates uplink (data upload) HTTPS bulk transmission from downlink (command issuance) MQTT long connection, and is supplemented by a periodic heartbeat mechanism to realize real-time device status perception.

[0071] For example, the uplink channel (HTTPS) is used for batch uploading of sensor data. The reasons for choosing HTTPS instead of MQTT for uploading data are: HTTPS has higher throughput when transmitting data in batches (reducing the protocol overhead of MQTT publishing one message at a time); HTTPS natively supports resuming interrupted uploads (Content-Range header); and the HTTPS protocol stack of the edge device's 4G module is more mature and stable.

[0072] Upload data packet format (JSON example): { "device_id":"BJ-CY-045", "fingerprint":"a3f8...7d2e", "batch_id":"20260413-103000-001", "upload_mode":"NORMAL", "records":[ { "timestamp":"2026-04-13T10:25:00+08:00", "channels":{ "SO2":{"value":12.3,"unit":"ppb","qc_flag":"VALID"}, "CO":{"value":0.8,"unit":"ppm","qc_flag":"VALID"}, "NO2":{"value":28.5,"unit":"ppb","qc_flag":"VALID"}, "O3":{"value":45.2,"unit":"ppb","qc_flag":"VALID"}, "PM25":{"value":35.0,"unit":"ug / m3","qc_flag":"VALID"}, ... }, "gps":{"lat":39.9042,"lon":116.4074,"alt":52.0}, "device_status":{"battery_pct":78,"storage_used_pct":23} } ], "edge_diagnostics":{ "sensor_health_summary":"ALL_NORMAL" } } For example, the MQTT (Message Queuing) channel is used by the central agent to send commands to the edge agents and receive heartbeats. MQTT topic hierarchy: iems / ├──heartbeat / {device_id} / / Heartbeat reporting (edge ​​→ center) ├──command / {device_id} / / Command issuance (center → edge) ├──diagnostic / {device_id} / req / / diagnostic request (center → edge) ├──diagnostic / {device_id} / res / / diagnostic results (edge ​​→ center) ├──skill / {device_id} / / Skill pack management command (center → edge) └──broadcast / region / {region} / / Regional broadcast (center → all edges within the region) For example, the time synchronization mechanism is as follows: all edge agents adopt a GPS+NTP dual-source time synchronization strategy: the main clock source is UTC time provided by the GPS module (with second-level accuracy to meet environmental monitoring requirements); the backup clock source is NTP synchronization via the 4G network (enabled when the GPS signal is weak); the synchronization cycle is calibrated every 300 seconds; each data record carries a timestamp and clock source identifier ("GPS" or "NTP") for the central agent to evaluate the time quality.

[0073] In one implementation, device authentication employs dual authentication using two-way TLS (mTLS) and device fingerprint: At the communication layer, mTLS ensures encrypted transmission and mutual identity verification; at the application layer, each data upload carries the device fingerprint, and the central agent verifies the binding relationship between the fingerprint and the device_id. Device loss handling: Administrators can trigger device locking. Upon receiving the locking command, the edge agent stops data collection and uploading, retaining only GPS location reporting to assist in device retrieval.

[0074] In one implementation, a communication mechanism between edge agents is designed: for security reasons, direct point-to-point communication is not performed between edge agents. All cross-device interactions are relayed and approved through a central agent to prevent malicious devices from attacking other devices through lateral movement.

[0075] In summary, the autonomous edge agent networking architecture of this invention differs from the traditional "dumb terminal + intelligent cloud" architecture of IoT platforms. In this invention, each edge node is itself an AI agent with reasoning, diagnostic, and memory capabilities. After networking, the edge agents maintain local autonomy while accepting unified coordination from the central agent. In the event of a network outage, the edge agents can independently complete data collection, quality control, and anomaly diagnosis, and securely store the data in local storage (supporting at least 180 days of data buffering). After the network is restored, the backlogged data is transmitted back to the center according to a priority strategy.

[0076] This invention maintains a knowledge graph of the spatiotemporal distribution characteristics of pollutants / aerosols. This graph encodes the spatial uniformity characteristics of different pollutants under different meteorological conditions, geographical environments, and time periods (e.g., the spatial correlation scale of PM2.5 under stable weather conditions is approximately 2-5 km, while that of NO2, due to its more localized source, is approximately 0.5-1 km). This knowledge graph is initialized with literature knowledge and continuously learns and corrects itself from actual observation data as the system operates, achieving self-evolution. Based on this knowledge graph, the spatiotemporal consistency verification engine can distinguish between abnormal readings caused by equipment failures and localized high values ​​caused by real environmental events, avoiding the misclassification of scientifically valuable extreme observations as faulty data.

[0077] This invention enables semantic-level collaborative communication between intelligent agents: the interaction between the central intelligent agent and edge intelligent agents is not limited to data transmission and command issuance, but also supports semantic-level collaborative reasoning. For example, the central intelligent agent can initiate a "diagnosis request" to a specific edge intelligent agent, and the edge intelligent agent returns a structured diagnostic report after running a local diagnostic algorithm; the central intelligent agent integrates the diagnostic results of multiple edge intelligent agents to perform cross-device correlation analysis. This "request-reasoning-reporting" collaborative model between intelligent agents differs from the unidirectional data flow of "collection-upload-processing" in traditional IoT platforms.

[0078] This invention proposes a skill package canary release and security upgrade mechanism: It proposes a skill package canary release strategy for tens of thousands of devices, including a three-stage release process (canary verification → canary expansion → full push) and an automatic rollback mechanism. The central intelligent agent automatically compares data quality indicators before and after the upgrade at each stage. Once a quality degradation is detected, an automatic rollback is triggered to ensure the security of large-scale upgrades.

[0079] This invention designs a unified management interface driven by natural language: administrators do not need to learn a professional device management interface; they can complete all management operations, such as device status query, anomaly diagnosis, task assignment, and data analysis, by conversing with the central intelligent agent through natural language. The central intelligent agent translates the natural language intent into specific tool call sequences (querying the device knowledge graph, initiating diagnosis to edge intelligent agents, generating analysis reports, etc.) and returns the results in natural language form.

[0080] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0081] In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0082] 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 cloud-edge intelligent agent collaborative networking system for monitoring individual environmental exposure in a region, characterized in that, include: Edge agents, deployed on each individual exposure monitoring device and acting as autonomous AI agents, communicate with the central agent via a wireless network; The central intelligent agent, deployed on a cloud server cluster, executes the central AI agent to coordinate and schedule all online edge intelligent agents by setting up a device status knowledge graph, a spatiotemporal consistency verification engine, a conversational management engine, a task orchestration and canary release engine, and a user permission management module. The communication protocol layer is used to employ a dual-channel communication architecture that separates uplink HTTPS bulk transmission from downlink MQTT long connections, supplemented by a periodic heartbeat mechanism to sense and monitor the real-time status of the device.

2. The edge-cloud intelligent agent collaborative networking system for monitoring individual environmental exposure in a region according to claim 1, characterized in that: The edge agent also includes a device fingerprint module, which serves as a unique identifier for the central agent to identify and manage the monitoring device. The fingerprint generation logic is loaded in the form of a skill package and is uniformly managed and upgraded by the central agent.

3. The edge-cloud intelligent agent collaborative networking system for monitoring regional individual environmental exposure according to claim 1, characterized in that: The edge agent also includes a local data preservation module, which uses a partitioned storage strategy to store the raw sensor data of the edge agent to ensure that the raw sensor data is not lost under any network conditions.

4. The edge-cloud intelligent agent collaborative networking system for regional individual environmental exposure monitoring according to claim 1, characterized in that: The edge agent further includes a communication management module for communicating with the central agent, including: Heartbeat transmission is used to send a device status summary to the center at a specified frequency. Data upload is used to upload sensor data in batches via HTTPS. Command reception is used to receive commands issued by the central intelligent agent via an MQTT long connection; Network outage detection and recovery is used to monitor network status. When the network is out of service, it automatically switches to local security mode and uploads backlogged data according to priority after recovery.

5. The edge-cloud intelligent agent collaborative networking system for monitoring individual environmental exposure in a region according to claim 1, characterized in that: The edge agent also includes an intelligent adaptive policy uploading module, used to automatically switch working modes based on network status and the presence of abnormal events. The working modes include four types: Normal mode: All data within the cycle is uploaded in batches every 5 minutes, and the confirmed queue data is cleared after successful upload; Alarm mode: Upload data one by one in real time, set abnormal data as high priority, and switch back to normal mode after the abnormality is eliminated; Offline mode: Data is stored in a local buffer, and abnormal data is additionally stored in a priority queue. Network connection is continuously retried according to an exponential backoff strategy. Recovery mode: Prioritize uploading abnormal high-priority data, then upload ordinary backlog data in reverse order, adaptive bandwidth transmission, and upload all data.

6. The edge-cloud intelligent agent collaborative networking system for monitoring regional individual environmental exposure according to claim 1, characterized in that: The central intelligent agent is used to complete the full-link modeling through the device state knowledge graph, which consists of three layers of entities and four types of relationships: using a geographical region as the spatial carrier to mount all edge intelligent devices; constructing the adjacency topology between devices based on GPS coordinates; binding all alarm events generated by the devices; and then performing anomaly pattern similarity association on all alarms, ultimately forming a complete edge intelligent device association knowledge network covering geographical space, device entities, operational alarms, and anomaly associations.

7. The edge-cloud intelligent agent collaborative networking system for monitoring individual environmental exposure in a region according to claim 1, characterized in that: The central intelligent agent is used to construct the spatiotemporal consistency verification engine by utilizing a knowledge graph of the spatiotemporal distribution of pollutants and a cross-device spatiotemporal consistency verification algorithm. The cross-device spatiotemporal consistency verification algorithm first identifies single-point data deviations through spatial neighborhood statistics, then distinguishes between equipment failures and real pollution phenomena through multi-layer business verification, and takes into account both data quality control and environmental event capture to verify the uploaded sensor data records.

8. The edge-cloud intelligent agent collaborative networking system for monitoring regional individual environmental exposure according to claim 1, characterized in that: The central intelligent agent is used to generate the conversational management engine by calling the ToolUse mechanism through a large language model and tools.

9. The edge-cloud intelligent agent collaborative networking system for monitoring regional individual environmental exposure according to claim 1, characterized in that: The central intelligent agent is used to execute configuration change tasks and skill package canary release tasks through the task orchestration and canary release engine. When executing the skill package canary release task, the canary verification, canary expansion and full push operations are performed in sequence to execute the skill package canary release task. If the data integrity rate of upgraded devices decreases by more than 10%, or the frequency of abnormal alarms on upgraded devices increases by more than 200%, or any upgraded device reports a skill loading failure log / alarm, a global automatic hard rollback operation will be triggered. Meanwhile, skipping stages is prohibited if the required monitoring duration is not completed at each stage; after automatic rollback execution, root cause analysis and version repair must be completed before the canary rollout process can be restarted; batch one-click rollback is no longer supported after snapshots expire, only manual downgrade of a single machine is supported.

10. The edge-cloud intelligent agent collaborative networking system for monitoring individual environmental exposure in a region according to claim 1, characterized in that: The edge agent employs a GPS+NTP dual-source time synchronization strategy: Master clock source: UTC time provided by the GPS module; Backup clock source: 4G network NTP synchronization; Synchronization cycle: Calibrate every 300 seconds; Each data record carries a timestamp and clock source identifier, which are used by the central agent to evaluate time quality.

Images