Self-evolution environment monitoring system and method based on digital twinning and causal perception

Abstract

The application provides a self-evolution environment monitoring system based on digital twinning and causal perception, which is suitable for high-precision, high-timeliness, interpretable and sustainable evolution monitoring and management of complex dynamic environments such as urban atmosphere, industrial parks and ecological sensitive areas. The system adopts a three-level collaborative architecture of "edge-fog-cloud", which is composed of the following three parts: pulse edge intelligent micro station (PEIM): deployed on the ground of the target area, responsible for event-driven multi-modal environment perception; light-weighted environment digital twinning (LEDT): deployed on the edge gateway or fog computing node, constructs a local virtual environment field and performs anchor-free self-calibration; causal perception center (CACH): deployed on the regional server or cloud, fuses multi-source data to construct a dynamic environment causal graph, and outputs pollution tracing results and intervention suggestions.

Description

Technical Field

[0001] This invention belongs to the field of environmental monitoring technology, and more specifically, relates to a self-evolving environmental monitoring system and method based on digital twins and causal perception. Background Technology

[0002] Environmental monitoring is a crucial technological foundation for continuously or periodically observing environmental elements such as air, water, and noise using sensors, instruments, and other equipment. This monitoring serves to understand changes in environmental quality, identify pollution risks, and support management decisions. With accelerating urbanization and industrialization, the demand for timely, high-precision, and wide-coverage environmental monitoring capabilities is becoming increasingly urgent.

[0003] Current mainstream environmental monitoring systems primarily rely on two approaches: one is to build a small number of high-precision standard monitoring stations. While this provides reliable data, it incurs high investment and maintenance costs, has limited spatial coverage, and struggles to capture localized or sudden pollution events. The other approach involves deploying a large number of low-cost IoT sensor nodes. Although this increases deployment density, it generally suffers from low sensing efficiency, poor data reliability, and weak analytical capabilities. Specifically, these nodes typically employ fixed-period sampling strategies, continuously collecting redundant data when no events occur, resulting in energy waste. Furthermore, they fail to trigger timely high-precision observations when sudden pollution events occur (such as industrial leaks or sudden increases in road dust), leading to the loss of crucial information. Simultaneously, low-cost sensors are susceptible to drift errors due to factors such as temperature, humidity, and aging, while existing calibration methods generally rely on high-precision reference equipment, making it difficult to achieve autonomous correction without anchor points. More critically, existing systems often rely on statistical correlation for pollution source tracing, failing to effectively distinguish between causal relationships and confounding interference. This results in inaccurate pollution source location and a lack of interpretability and scientific basis for proposed remediation measures. Summary of the Invention

[0004] This invention aims to provide a self-evolving environmental monitoring system and method based on digital twins and causal perception. By integrating pulsed edge intelligent perception, lightweight digital twins and causal graph reasoning, it achieves high-sensitivity capture of environmental mutation events, high-precision state mapping, explainable pollution source tracing, and continuous evolution driven by closed-loop feedback.

[0005] A self-evolving environmental monitoring system based on digital twins and causal perception includes:

[0006] Multiple pulse-type edge intelligent micro-stations are deployed in the target monitoring area and configured to trigger high-precision sampling when environmental abrupt events are detected, and generate and upload a multi-dimensional feature package containing time-series gradients, spectral features and cross-modal correlation indicators;

[0007] A lightweight environmental digital twin, corresponding one-to-one with each of the pulse-type edge intelligent micro-stations, is deployed on edge computing nodes or fog computing layers. It is configured to construct a virtual state mapping based on real-time sensing data from physical micro-stations and perform counterfactual inference and anchorless self-calibration based on spatiotemporal consistency constraints.

[0008] The causal perception center, configured on a regional server or cloud server, communicates with the pulsed edge intelligent micro-station and the environmental digital twin. It is configured to receive the multi-dimensional feature package, external context data and environmental prior knowledge, construct and dynamically update the environmental causal map, and output pollution source location results, environmental impact prediction and intervention suggestions based on the causal map.

[0009] Furthermore, the pulsed edge intelligent micro-station includes: a multimodal micro-sensor array integrating gas, acoustic, optical, and electromagnetic sensing units; an event-driven sampling engine configured to activate a high-precision sampling mode when any of the following conditions are met: (a) the gradient of a single pollutant concentration changes within a preset time window exceeds a threshold; (b) two or more heterogeneous sensing signals exhibit atypical coupling relationships; an ultra-low-power processor configured to maintain a dormant state under normal conditions and generate the multidimensional feature packet after an event is triggered; and a dual-mode communication module supporting LoRaWAN long-distance transmission and BLE Mesh short-range collaborative networking.

[0010] Furthermore, the environmental digital twin achieves sensor drift compensation through an anchorless self-calibration mechanism, which includes: (a) acquiring concurrent observation data, meteorological forecast information, and geographical topological relationships from neighboring microstations; (b) constructing a local environmental field smoothness model based on the data to determine whether the current microstation output violates the law of physical continuity; and (c) if drift is determined to exist, generating a correction offset based on counterfactual reasoning and dynamically adjusting the sensor output, wherein the counterfactual reasoning is calculated based on the state response differences under virtual perturbation.

[0011] Furthermore, the environmental causal graph constructed by the causal perception center includes: a set of nodes covering pollution sources, transmission media, and sensitive receptors; and a set of directed edges, where each directed edge represents a causal influence path from a source node to a receptor node, and its weights are learned by a cross-modal causal graph neural network. The cross-modal causal graph neural network integrates physical diffusion models and data-driven evidence. Its inputs include the multidimensional feature package, satellite remote sensing segments, traffic flow data, and social media sentiment keywords, and its outputs are the causal path confidence and the estimated value of the intervention effect.

[0012] Furthermore, the causal perception center is further configured to: (a) respond to a sudden pollution event by calling the cross-modal causal graph neural network to perform counterfactual simulation and identify the most likely pollution source; (b) perform strategy deduction based on the updated causal graph to generate an intervention plan, which includes at least one of traffic control, water spraying scheduling, or enterprise production restriction recommendations; and (c) feed back the implementation effect of the intervention plan to each environmental digital twin to update its state mapping model and calibration strategy, thereby driving the continuous evolution of the system's overall cognitive ability.

[0013] Furthermore, the system has self-diagnosis and self-maintenance capabilities, including: (a) continuously comparing the actual output of the physical micro-station with its virtual state prediction value in each environmental digital twin; (b) when the comparison deviation continues to exceed the preset tolerance and cannot be eliminated by the anchorless self-calibration, automatically marking the micro-station as a fault state and pushing a maintenance work order to the maintenance platform.

[0014] An environmental monitoring method includes the following steps: (a) implementing event-driven environmental perception through the pulsed edge intelligent micro-station, and generating and uploading a multi-dimensional feature package when an environmental mutation event is detected; (b) using the environmental digital twin based on spatiotemporal consistency constraints and neighborhood observation data to complete local state mapping, sensor drift correction, and missing data inference, wherein the correction does not rely on high-precision standard monitoring equipment; (c) fusing multi-source heterogeneous data in the causal perception center, constructing and incrementally updating the environmental causal graph through a cross-modal causal graph neural network; (d) outputting pollution source tracing results with causal path confidence and governance suggestions with quantitative intervention effects based on the causal graph; (e) feeding back the intervention implementation effect to the digital twin to optimize its state prediction and calibration model, thereby achieving self-evolution of the system's monitoring capabilities.

[0015] Furthermore, the anchorless self-calibration described in step (b) is based on the spatiotemporal continuity prior of the environmental field and the neighborhood consistency constraint. It detects abnormal outputs by constructing a local smoothness model and generates correction offsets using counterfactual reasoning.

[0016] Furthermore, the cross-modal causal graph neural network described in step (c) adopts an architecture that combines a structural causal model with a graph attention mechanism. It distinguishes between confounding factors and real causal effects through intervention simulation and backdoor path adjustment, and outputs an interpretable causal chain.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] (1) By using an event-triggered mechanism through a pulsed edge intelligent micro-station, it maintains an ultra-low power sleep state under normal conditions and activates high-precision sampling only when a sudden change in the environment is detected, which greatly reduces energy consumption and communication load, while ensuring high-fidelity capture of key events, effectively improving the response efficiency and sustainable operation capability of the monitoring system.

[0019] (2) By constructing a local environmental field smoothness model based on spatiotemporal consistency constraints and neighborhood observation data through an environmental digital twin, automatic identification and counterfactual reasoning correction of sensor drift can be achieved without relying on expensive standard monitoring equipment, which significantly improves data reliability and system stability under long-term deployment.

[0020] (3) The causal perception center integrates multi-dimensional feature packages, remote sensing, traffic, public opinion and other multi-source heterogeneous data through cross-modal causal graph neural network, and combines physical diffusion model and structural causal reasoning to accurately construct a dynamically updated environmental causal map. It can not only output high-confidence pollution source location results, but also quantify the expected effects of different intervention measures, providing an interpretable and verifiable basis for scientific decision-making. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the system architecture. Detailed Implementation

[0022] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of the invention.

[0023] Example:

[0024] like Figure 1 As shown, this invention provides a self-evolving environmental monitoring system based on digital twins and causal perception, suitable for high-precision, high-timeliness, interpretable, and sustainable evolutionary monitoring and governance of complex and dynamic environments such as urban atmospheres, industrial parks, and ecologically sensitive areas. The system adopts a three-level collaborative architecture of "edge-fog-cloud," consisting of the following three parts: Pulsed Edge Intelligent Micro-station (PEIM): deployed on the ground in the target area, responsible for event-driven multimodal environmental perception; Lightweight Environmental Digital Twin (LEDT): deployed on edge gateways or fog computing nodes, constructing a local virtual environmental field and performing anchor-free self-calibration; Causal Perception Hub (CACH): deployed on a regional server or in the cloud, fusing multi-source data to construct a dynamic environmental causal map, outputting pollution source tracing results and intervention suggestions.

[0025] The overall system operation forms a closed loop: PEIM detects mutation events → LEDT maps and calibrates the state → CACH infers causality and generates strategies → the intervention effect is fed back to LEDT → the system's cognitive ability continues to evolve.

[0026] In this scheme, each PEIM includes:

[0027] Multimodal micro-sensor array: integrating an electrochemical gas sensor (detecting NO2, SO2, O3, CO, VOCs), a MEMS acoustic microphone (sampling rate 8 kHz), and a micro light scattering PM sensor (outputting PM1 / PM2.5). 2.5 / PM 10 ), low-frequency electromagnetic probe;

[0028] Ultra-low power processor: The main control chip is ARM Cortex-M7 (with NPU acceleration unit), and the coprocessor is Cortex-M0+; Dual-mode communication module: Supports LoRaWAN (Class A, spread factor SF7–SF12 adaptive) and BLE 5.0 Mesh;

[0029] Power management unit: Supports lithium battery (5 Ah) or solar charging, with μA-level sleep current.

[0030] PEIM normally polls the sensor's coarse value at a frequency of 1 Hz, maintaining a sleep state (power consumption ≈ 20 μW). High-precision sampling mode is activated (lasting 10–30 seconds) when any of the following conditions are met:

[0031] Condition (a): Single-modal abrupt change detection: For the gas concentration signal x(t), calculate the gradient of the sliding window (Δt = 60 s):

[0032]

[0033] If |G x (t)∣>τ grad And the second rate of change A x If (t) > 0, then it is triggered. Where the threshold τ... grad Dynamically adjusted using Exponentially Weighted Moving Average (EWMA):

[0034]

[0035] Condition (b): Cross-modal coupling anomaly detection: For gas g(t) and acoustic energy s(t), calculate symbol mutual information: convert the signal into a symbol sequence (↑, ↓, →); statistically analyze the joint probability distribution and calculate the mutual information I(g;s)t; if I(g;s)t deviates from the historical mean by more than 2 standard deviations, it is determined to be atypical coupling.

[0036] Once triggered, all sensors enter a high sampling rate operating state (lasting 10–30 seconds), and the main processor performs feature extraction to generate a multidimensional feature packet (MFP) containing temporal gradients (such as first / second gradients, Hurst exponent), spectral features (such as wavelet packet energy entropy, MFCC coefficients), and cross-modal correlation indices (such as DTW distance, Granger causality p-value).

[0037] After Z-score normalization, the MFP is compressed to less than 2 KB in CBOR format and uploaded via a dual-mode communication module.

[0038] Furthermore, during uploads, BLE Mesh is prioritized for close-range collaborative networking to reduce latency and power consumption. If no nearby nodes are available, LoRaWAN is switched to achieve wide-area transmission. This design enables PEIM to maintain high sensitivity while achieving an average daily activation time of less than 60 seconds and a battery life of over 2 years, effectively overcoming the shortcomings of traditional continuous sampling systems, such as high power consumption and large data redundancy.

[0039] In this solution, the lightweight environmental digital twin (LEDT) is deployed at edge nodes (such as smart street light gateways), with resource limitations of 2-core ARM Cortex-A53 and 1 GB RAM. It constructs a pollutant concentration field C(x,t) based on MFP, meteorological API (wind speed, stability), and GIS topographic data uploaded by PEIM.

[0040] To balance accuracy and efficiency, LEDT uses sparse Gaussian process regression (SGPR) for modeling:

[0041]

[0042] Wherein, the mean function μ phys A simplified Gaussian plume model is introduced, including wind speed u and diffusion coefficient K; the covariance function is a spatial Matérn 3 / 2 kernel × a time exponential kernel; the induction point is the current microstation + the 5 nearest PEIM locations, M=6, and the computational complexity is controllable.

[0043] The anchorless self-calibration mechanism in this scheme does not rely on standard stations, and the process is as follows:

[0044] Data aggregation: Obtain nearby PEIM observations, meteorological data, and building occlusion factors within a 500 m radius (from OpenStreetMap ray projection);

[0045] Anomaly detection: Calculating Local Laplace Residuals (LLR):

[0046]

[0047] If the LLR is greater than 3 times the historical standard deviation, it is considered a drift.

[0048] Counterfactual correction: Solving for the solution that optimizes the smoothness of the entire field. :

[0049]

[0050] This correction problem is a quadratic optimization and can be solved analytically (<5 ms);

[0051] Output Update:

[0052]

[0053] Furthermore, when PEIM data is missing due to faults or communication interruptions, LEDT supports missing data extrapolation: for short-term missing data (<10 min), Kalman filtering is used, with the state transition matrix based on a first-order inertial model; for long-term missing data (>30 min), a neighborhood graph is constructed, with edge weights considering wind direction and advection, and pre-trained graph convolution (GCN) interpolation is used. This mechanism allows LEDT to achieve high-precision calibration (error <±8%) without relying on expensive standard monitoring stations, and maintains regional monitoring continuity when some nodes fail.

[0054] In this solution, the causal perception center is deployed on a regional server or cloud platform (such as a government cloud or environmental protection dedicated cloud), serving as the core of the system's cognition and decision-making. CACH first receives calibrated status, MFP, and uncertainty indicators from all LEDTs, and integrates multi-source external contextual data, including satellite remote sensing (such as NO2 column concentration from Sentinel-5P, time-aligned to ±15 minutes after aerosol correction), traffic flow data (GPS trajectories aggregated into road segment emission factors), social media sentiment (emotional intensity scores for keywords such as "pungent" and "dizzy" extracted through fine-tuning the BERT model), and enterprise prior knowledge (discharge permits, process types, and historical violation records).

[0055] All spatiotemporal data are unified to a 100m x 100m, 1-minute granularity grid through GeoHash encoding and Unix timestamps.

[0056] Based on this, CACH runs a cross-modal causal graph neural network (CM-CGNN) to construct and dynamically update an environmental causal graph. The graph nodes cover pollution sources (such as chemical plants and traffic routes), transmission media (such as wind field units), and sensitive receptors (such as schools and hospitals). The initial edge structure is constructed based on HYSPLIT wind direction trajectories or Gaussian plume direction cones, and the edge weights are initialized as diffusion attenuation factors.

[0057] CM-CGNN uses graph attention mechanism (GAT) to aggregate neighbor information, and the attention coefficients fuse cross-modal evidence (such as time-lag mutual information and spatial co-occurrence frequency). It also introduces a dual machine learning (DML) framework to eliminate confounding bias: first, XGBoost is used to estimate the impact of confounding variables (such as common upstream weather) on the source and receptor, and then the residuals are used to fit the true causal effect, thereby effectively distinguishing between statistical correlation and causal relationship.

[0058] The causal graph adopts an incremental update strategy. Sudden pollution events trigger local subgraph reconstruction, and full graph fine-tuning is performed every morning. Only causal paths with a Bootstrap resampling confidence greater than 0.7 are retained.

[0059] When an abnormal event occurs, CACH performs counterfactual simulations: it applies a do operation (e.g., do(source=0)) to candidate pollution sources, calculates the expected concentration if the source is shut down, and generates a source tracing probability ranking accordingly; at the same time, the system enumerates feasible intervention strategies (e.g., traffic restrictions, water spraying scheduling, and enterprise production restrictions), predicts the environmental improvement effect and social cost of each strategy, recommends the solution that meets the improvement threshold and has the lowest cost, and outputs it in the form of structured instructions.

[0060] After the intervention is implemented, PEIM / LEDT uploads the actual environmental changes, and CACH compares the predicted and measured effects. If the error exceeds 20%, the CM-CGNN parameters are fine-tuned online, and successful cases are stored in the causal strategy library to drive the continuous evolution of the system's cognitive ability.

[0061] This system also possesses comprehensive self-diagnosis and self-maintenance capabilities. Each LEDT continuously compares its virtual state prediction value with the PEIM calibration output value. If the deviation exceeds 20% for three consecutive times and cannot be eliminated through anchorless self-calibration, the corresponding PEIM is automatically marked as faulty, and a maintenance work order containing the fault type (such as sensor failure, communication interruption), suggested replacement parts, and temporary data alternatives (interpolated from neighboring LEDTs) is pushed to the maintenance platform, significantly reducing the cost of manual inspection.

[0062] Example 2: System Operation Example

[0063] At 23:00 one night, PEIM detected that VOCs increased from 50 ppb to 500 ppb within 5 minutes, and at the same time, the acoustic energy suddenly increased, triggering high-precision sampling;

[0064] MFP is uploaded to the fog node, LEDT constructs a VOCs field, an LLR anomaly is found at this point, and counterfactual calibration is initiated;

[0065] CACH integrates MFP, satellite patch data, and enterprise logs, and CM-CGNN locates the pollution source as "illegal operation of a paint factory" (92% confidence level).

[0066] Intervention measures: "Order a 6-hour production halt, expected to reduce VOCs by 80%";

[0067] At 05:00 the next day, PEIM data showed that VOCs dropped to 60 ppb → CACH verified the effect and fine-tuned the weight of the factory path in CM-CGNN;

[0068] In summary, this invention achieves ultra-low power consumption and high sensitivity event-driven perception through pulse-type edge intelligent micro-stations, completes anchor-free self-calibration and state mapping through lightweight environmental digital twins, and constructs interpretable and quantifiable dynamic causal graphs and generates intervention strategies through a causal perception center. These three elements work together to form a closed-loop system with self-diagnosis, self-calibration, and self-evolution capabilities. Field tests show that this system can reduce edge power consumption by more than 85%, improve pollution source tracing accuracy to over 85%, control calibration errors within ±8%, and eliminate the need for densely deployed standard monitoring stations, significantly reducing deployment and maintenance costs.

[0069] The embodiments of the present invention are given for illustrative and descriptive purposes only, and are not intended to be exhaustive or to limit the invention to the forms disclosed. Many modifications and variations will be apparent to those skilled in the art. The embodiments were chosen and described in order to better illustrate the principles and practical application of the invention, and to enable those skilled in the art to understand the invention and to design various embodiments with various modifications suitable for a particular purpose.

Claims

1. A self-evolving environmental monitoring system based on digital twins and causal perception, characterized in that, include: Multiple pulse-type edge intelligent micro-stations are deployed in the target monitoring area and configured to trigger high-precision sampling when environmental abrupt events are detected, and generate and upload a multi-dimensional feature package containing time-series gradients, spectral features and cross-modal correlation indicators; A lightweight environmental digital twin, corresponding one-to-one with each of the pulse-type edge intelligent micro-stations, is deployed on edge computing nodes or fog computing layers. It is configured to construct a virtual state mapping based on real-time sensing data from physical micro-stations and perform counterfactual inference and anchorless self-calibration based on spatiotemporal consistency constraints. The causal perception center is configured on a regional server or cloud server, and is communicatively connected to the pulse-type edge intelligent micro-station and the environmental digital twin. It is configured to receive the multi-dimensional feature package, external context data and environmental prior knowledge, construct and dynamically update the environmental causal map, and output pollution source location results, environmental impact prediction and intervention suggestions based on the causal map. The environmental digital twin achieves sensor drift compensation through an anchorless self-calibration mechanism, which includes: (a) Obtain concurrent observation data, meteorological forecast information, and geographic topological relationships from neighboring microstations; (b) Construct a local environmental field smoothness model based on the data to determine whether the current microstation output violates the law of physical continuity; (c) If a drift is determined to exist, a correction offset is generated based on counterfactual reasoning, and the sensor output is dynamically adjusted, wherein the counterfactual reasoning is calculated based on the difference in state response under virtual perturbation; The environmental causal map constructed by the causal perception center includes: The node set encompasses pollution sources, transmission vectors, and sensitive receptors; A set of directed edges, each representing a causal influence path from a source node to a recipient node, with weights learned by a cross-modal causal graph neural network. This cross-modal causal graph neural network integrates a physical diffusion model with data-driven evidence. Its inputs include the multidimensional feature package, satellite remote sensing segments, traffic flow data, and social media sentiment keywords, and its outputs are the causal path confidence and intervention effect estimates.

2. The self-evolving environmental monitoring system based on digital twin and causal perception according to claim 1, characterized in that, The pulse-type edge intelligent micro-station includes: A multimodal micro-sensor array that integrates gas, acoustic, optical, and electromagnetic sensing units; The event-driven sampling engine is configured to activate the high-precision sampling mode when any of the following conditions are met: (a) the gradient of the single pollutant concentration changes within a preset time window exceeds a threshold; (b) two or more heterogeneous sensor signals exhibit atypical coupling relationships. An ultra-low power processor is configured to maintain a sleep state under normal conditions and generate the multi-dimensional feature package after an event is triggered. The dual-mode communication module supports LoRaWAN long-distance transmission and BLE Mesh short-range collaborative networking.

3. The self-evolving environmental monitoring system based on digital twins and causal perception according to claim 1, characterized in that, The causal perception center is further configured as follows: (a) In response to a sudden pollution event, invoke the cross-modal causal graph neural network to perform counterfactual simulations to identify the most likely source of pollution; (b) Based on the updated causal graph, perform strategy deduction to generate intervention plans, which include at least one of traffic control, sprinkler scheduling, or enterprise production restriction recommendations; (c) Feedback the implementation effect of the intervention scheme to each environmental digital twin to update its state mapping model and calibration strategy, thereby driving the continuous evolution of the system's overall cognitive ability.

4. The self-evolving environmental monitoring system based on digital twin and causal perception according to claim 1, characterized in that, The system possesses self-diagnosis and self-maintenance capabilities, including: (a) The digital twins of each environment continuously compare the actual output of the physical micro-station with its virtual state prediction; (b) When the comparison deviation continues to exceed the preset tolerance and cannot be eliminated by the anchorless self-calibration, the micro-station is automatically marked as faulty and a maintenance work order is pushed to the operation and maintenance platform.

5. An environmental monitoring method for the system as described in any one of claims 1 to 4, characterized in that, Includes the following steps: (a) Event-driven environmental perception is achieved through the pulse-type edge intelligent micro-station, and a multi-dimensional feature package is generated and uploaded when a sudden environmental event is detected; (b) Using the environmental digital twin based on spatiotemporal consistency constraints and neighborhood observation data, local state mapping, sensor drift correction and missing data inference are completed, wherein the correction does not depend on high-precision standard monitoring equipment; (c) Multi-source heterogeneous data are fused in the causal perception center, and an environmental causal graph is constructed and incrementally updated through a cross-modal causal graph neural network; (d) Based on the causal map, output pollution source tracing results with causal path confidence and governance recommendations with quantitative intervention effects; (e) Feedback the effects of the intervention to the digital twin to optimize its state prediction and calibration model, thereby enabling the self-evolution of the system's monitoring capabilities.

6. The environmental monitoring method according to claim 5, characterized in that, The anchorless self-calibration described in step (b) is based on the spatiotemporal continuity prior of the environmental field and the neighborhood consistency constraint. It detects abnormal outputs by constructing a local smoothness model and generates correction offsets using counterfactual reasoning.

7. The environmental monitoring method according to claim 5, characterized in that, The cross-modal causal graph neural network described in step (c) adopts an architecture that combines a structural causal model with a graph attention mechanism. It distinguishes between confounding factors and real causal effects through intervention simulation and backdoor path adjustment, and outputs an interpretable causal chain.

Images