Online monitoring system based on AI analysis

Abstract

The invention relates to the technical field of environment monitoring, aims to solve the problems that an existing environment monitoring system is insufficient in end-edge-cloud collaboration, difficult in multi-source data fusion, poor in decision-making one-sided interpretability, incapable of continuously and autonomously evolving a model, and difficult in consideration of data transmission redundancy and real-time performance, and provides an online monitoring system based on AI analysis. The system adopts an end-edge-cloud three-layer architecture, and is provided with an end-side lightweight trigger module, an edge-side progressive reasoning module, a cloud dual-engine collaborative decision-making module, a self-adaptive event-driven communication module and a closed-loop optimization feedback module. And a whole-process collaborative closed loop of monitoring, reasoning, decision making and optimization is formed. The real-time performance, the traceability accuracy and the adaptive capacity of environment monitoring are remarkably improved, and the method can be widely applied to scenes such as water quality monitoring in chemical industrial parks.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring technology, and in particular to an online monitoring system based on AI analysis. Background Technology

[0002] Existing technologies for AI-based online monitoring systems primarily focus on building a complete technical system around ecological and environmental monitoring scenarios. They generally employ machine learning algorithms to construct monitoring and analysis models, supporting real-time acquisition and analysis of environmental parameters and identification of abnormal states. Existing technologies often use support vector machine algorithms as their core, combining environmental monitoring data collected from multiple sources of sensors to complete model training and parameter optimization, enabling routine monitoring and early warning of exceedances of various environmental indicators. Simultaneously, existing technologies have built supporting multi-dimensional environmental data acquisition systems. These systems utilize front-end distributed sensor networks to complete real-time acquisition and standardized transmission of ecological and environmental parameters, combined with a back-end centralized computing platform to perform model inference calculations and output analysis results, forming a basic closed-loop monitoring and early warning process. These related technical solutions have been widely applied in various ecological and environmental monitoring scenarios, such as watershed management and industrial park supervision.

[0003] Regarding existing related technologies, the inventors believe that they often have the following shortcomings: existing technologies have poor adaptability to dynamic time-series data of the ecological environment, insufficient fitting accuracy for nonlinear and non-stationary ecological monitoring data, high model computing power requirements that are difficult to adapt to real-time monitoring scenarios at the edge, and the problem of model accuracy decay during long-term operation, resulting in insufficient refinement of ecological risk management. Summary of the Invention

[0004] The technical problem to be solved by this invention is that existing environmental monitoring systems suffer from insufficient end-edge-cloud collaboration and difficulty in fusion of multi-source data. Specifically, each level operates independently, data silos lead to delayed anomaly response, biased decision-making, and the model cannot continuously evolve. To address this, we propose an online monitoring system based on AI analysis.

[0005] To achieve the above objectives, this application adopts the following technical solution: an online monitoring system based on AI analysis, comprising a three-layer architecture consisting of edge sensing devices, edge computing nodes, and cloud servers, including: The edge-side lightweight triggering module is used to collect multi-source, multi-modal environmental monitoring data in real time and run a micro-capacity anomaly detection model to determine whether the current data deviates from the historical baseline. An upload command is triggered only when a deviation is detected, uploading the data packet to the edge computing node. The edge-side sensing device is various sensor terminals deployed at the monitoring site. The multi-source, multi-modal environmental monitoring data includes one or more of the following: sensor numerical sequences, image frame data, audio data, and text records. The data packet contains the original monitoring data, timestamp, geographic location identifier, and device identifier. The edge-side progressive inference module receives and fuses multi-source, multi-modal data from multiple edge devices, generates a fused feature vector, and inputs it into a dynamic deep neural network containing multiple cascaded layers and corresponding early exit branches for progressive computation. The computation terminates and outputs the environmental state identification result when the output confidence of any early exit branch meets a preset threshold. The feature vectors of high-confidence abnormal events and the preliminary inference results are reported to the cloud. The edge computing node is a regional server or gateway device deployed near the edge devices. The feature vector is a high-dimensional numerical representation extracted by the neural network. The confidence level is the normalized probability value of the neural network output layer. The cloud-based dual-engine collaborative decision-making module receives the abnormal event feature vectors reported by the edge-side progressive inference module and inputs them into the knowledge graph tracing engine and the digital twin inference engine for parallel processing. The knowledge graph tracing engine generates an explanation path and credibility score from the abnormal phenomenon to the potential pollution source based on a pre-built environmental causal knowledge graph. The digital twin inference engine simulates the spatiotemporal diffusion process of pollutants based on a lightweight environmental dynamics model and generates trend predictions. The tracing results and trend predictions are fused to generate an intelligent decision package containing disposal suggestions. The cloud server is a high-performance server cluster deployed in a data center. The explanation path is a causal chain from the abnormal indicator node to the potential pollution source node. The credibility score is a numerical score calculated based on path attributes. The intelligent decision package includes pollution source tracing results, pollution trend predictions, emergency disposal suggestions, and law enforcement evidence materials. The adaptive event-driven communication module is used to adjust the data transmission strategy in real time according to the confidence level of abnormal events. Confirmed events are transmitted in real time with priority, suspected events are stored in edge nodes and uploaded on demand, and summaries of regular events are uploaded periodically. It also enables local caching and resume transmission mechanisms when the network is interrupted. The closed-loop optimization feedback module is used to compare the actual monitoring data sequence after emergency response with the theoretical decay curve generated by the digital twin inference engine, calculate the deviation and generate quantitative evaluation information, and feed it back to the cloud model training platform to update the core model parameters of each module. The modules form a closed-loop collaborative process of monitoring, reasoning, decision-making, and optimization. The cloud-based dual-engine parallel processing integrates the tracing and inference results, which is different from the separate application of knowledge graphs and digital twins.

[0006] Preferably, the micro-capacity anomaly detection model running in the edge-side lightweight triggering module is an unsupervised learning model based on single-class classification, which is deployed locally on the edge-side microcontroller after structured pruning optimization. The unsupervised learning model based on single-class classification includes one of single-class support vector machine, isolated forest, and deep autoencoder. The structured pruning optimization is channel pruning or layer pruning of the neural network. The edge-side microcontroller is a low-power embedded processor. The micro-capacity anomaly detection model uses corresponding feature extraction sub-modules for different types of data. The feature extraction sub-modules include one or more of the following: a statistical feature extraction module for sensor numerical sequences, a lightweight CNN feature extraction module for image frame data, and a Mel-spectrum feature extraction module for audio data. The dynamic adaptive threshold used by the micro-capacity anomaly detection model is adjusted based on historical baseline statistical features and real-time data signal-to-noise ratio to determine the degree of deviation between the current data distribution and the historical baseline distribution. The historical baseline statistical features include the mean, variance, and percentiles of historical data. The real-time data signal-to-noise ratio is the ratio of signal power to noise power. The adjustment logic of the dynamic adaptive threshold is to dynamically adjust the threshold size according to the real-time data signal-to-noise ratio.

[0007] Preferably, the multi-source, multi-modal data fusion operation performed by the side-progressive inference module includes: using a dynamic time warping algorithm and geohashing encoding to complete the spatiotemporal alignment of multi-source data; the dynamic time warping algorithm is used to map time series data with different sampling frequencies to a unified time axis; the geohashing encoding is used to map location information with different precisions to a unified spatial grid; extracting time-frequency domain features, visual features, and semantic features from numerical, image, and text data respectively; the time-frequency domain features include the frequency band energy features of wavelet packet decomposition and the spectral features of short-time Fourier transform; the visual features are feature vectors output by the intermediate layer of a lightweight convolutional neural network; the semantic features are vector representations transformed by word embedding models or pre-trained language models; using an inter-modal multi-head cross-attention mechanism to adaptively weight and fuse each modality feature, the inter-modal multi-head cross-attention mechanism assigns corresponding attention weights to each modality feature, and the attention weights are dynamically adjusted based on the real-time signal-to-noise ratio and historical inference confidence of the corresponding modality data; the inter-modal multi-head cross-attention mechanism includes multiple attention heads, each attention head independently calculating the correlation between different modality features.

[0008] Preferably, the dynamic deep neural network comprises multiple stacked convolutional layers. Each convolutional layer employs a multi-scale dilated convolution parallel structure, including a regular convolution parallel branch and a dilated convolution parallel branch. The convolution kernels of the dilated convolution parallel branch are set with corresponding dilation rates, which increase with the depth of the layer group. The outputs of these branches are concatenated and input into the next layer group. The regular convolution branch is used to extract local detail features; the dilated convolution branch is used to extract global context features. The number of channels in each convolutional layer group increases progressively. Each layer group in the dynamic deep neural network is followed by an early exit branch consisting of a global average pooling layer and a fully connected classifier. This early exit branch employs an adaptive dropout regularization mechanism, which sets a corresponding neuron dropout rate. The dropout rate is dynamically adjusted based on the variance of the current layer group's output feature map. The dropout rate in the adaptive dropout regularization mechanism is positively correlated with the feature map variance. The adaptive dropout formula is: , in, The dropout rate for the current layer group. The variance of the output feature map for the current layer group. For reference variance, such as historical baseline variance or global average variance, For adjustment coefficients, , These are the upper and lower thresholds for the discard rate.

[0009] Furthermore, the side-progressive inference module also includes a confidence dynamic calculation unit, used to calculate a comprehensive confidence score based on the feature vector output by the current layer group, the formula of which is: , in, To calculate the overall confidence score, The original confidence level is denoted by , and the maximum probability value output by the fully connected classifier that exits the branch early is denoted by . The signal-to-noise ratio factor is the normalized value obtained by comparing the historical baseline signal-to-noise ratio with the current data signal-to-noise ratio. is the confidence level change rate, and is the absolute value of the difference between the built-in confidence levels of adjacent time windows. The information entropy of the output feature vector probability distribution is calculated based on the probability distribution of the current layer's output feature vector after softmax normalization. , , As an adjustable hyperparameter, the comprehensive confidence score is used to classify events into three levels: confirmed, suspected, and routine, and serves as the basis for adjusting the data transmission strategy.

[0010] Preferably, the knowledge graph tracing engine includes an environmental causal knowledge graph database, a graph reasoning adaptation unit, an interpretable path generation unit, a path confidence quantification unit, and a decision suggestion matching unit. The environmental causal knowledge graph database stores five types of entities: monitoring indicators, abnormal events, pollution sources, environmental factors, and treatment measures, as well as three types of association edges: causal, spatial, and regulatory. The association edges are configured with updatable confidence weights. The monitoring indicator entities include pollutant indicators and their standard thresholds. The abnormal event entities include events exceeding standards, illegal discharge events, and leakage events. The pollution source entities include industrial enterprises, sewage treatment plants, agricultural non-point sources, and mobile sources. The environmental factor entities include meteorological parameters and hydrological parameters. The treatment measure entities include emergency measures and remediation measures. The causal association edges represent the causal relationships between entities. The relationships are as follows: spatial proximity edges represent spatial adjacency relationships between entities; regulatory association edges represent regulatory constraint relationships between entities; the graph reasoning adaptation unit maps abnormal event information into graph query statements; the interpretable path generation unit uses a depth-first graph search algorithm to traverse multiple reasoning paths from the node of the exceeding indicator to the node of the potential pollution source, and performs pruning optimization based on spatiotemporal and causal relationships; the pruning optimization strategy is to remove nodes whose spatiotemporal deviation from the current abnormal event exceeds a preset range and to remove nodes that have no causal relationship with the exceeding indicator; the path confidence quantification unit calculates the confidence score of each path based on the path topology length, the confidence of intermediate nodes, and the weight of the associated edges; the decision suggestion matching unit matches disposal measures according to the source tracing results and outputs them to the digital twin inference engine as boundary conditions.

[0011] Preferably, the digital twin inference engine includes a four-dimensional environmental digital twin, a lightweight environmental dynamics proxy model, a trend prediction output unit, and a visualization rendering unit. The four-dimensional environmental digital twin is constructed based on geographic information, hydrological and meteorological data, and pollution source distribution data, and can be dynamically updated. The geographic information data includes digital elevation models, water system network data, and land use type data. The hydrological and meteorological data includes water level, flow velocity, flow rate, wind direction, wind speed, temperature, and rainfall. The pollution source distribution data includes fixed source coordinates, mobile source trajectories, and area source distribution. The dynamic update mechanism of the four-dimensional environmental digital twin is based on periodic updates of optimized parameters according to real-time monitoring data and cloud feedback. The lightweight environmental dynamics proxy model is deployed after model pruning and quantization, using knowledge graph source tracing results as core boundary conditions and real-time monitoring... The data serves as calibration parameters, enabling rapid simulation of the spatiotemporal diffusion process of pollutants on a digital twin. Model pruning involves removing neuronal connections or convolutional channels that contribute little to the output. Quantization converts model weights from high-bit-count to low-bit-count. Boundary conditions include the pollution source location, emission intensity, and emission duration output from the knowledge graph tracing results. Calibration parameters include real-time meteorological and hydrological monitoring data. The trend prediction output unit generates the pollution diffusion trajectory, impact range, and theoretical decay curve, and synchronizes them to the closed-loop optimization feedback module. The pollution diffusion trajectory is a spatiotemporal sequence of pollutant concentrations exceeding a preset threshold. The impact range is the geographical boundary of the pollutant-covered area. The theoretical decay curve is the curve showing the natural decay of pollutant concentration over time without human intervention after a pollution event. The visualization rendering unit presents the simulation results.

[0012] Preferably, the closed-loop optimization feedback module includes an actual data acquisition unit, a deviation calculation unit, a measure effectiveness evaluation unit, and a model feedback update unit; the actual data acquisition unit continuously collects monitoring data within a preset time period after emergency response; the preset time period is a time interval that can fully assess the recovery process of the pollution event; the deviation calculation unit compares the actual data with the theoretical decay curve generated by the digital twin simulation to calculate the pollution reduction rate deviation; the theoretical decay curve is the natural decay curve of pollutant concentration under no human intervention generated by the digital twin simulation engine; the pollution reduction rate is the percentage decrease in pollutant concentration after the pollution event occurs; the measure effectiveness evaluation unit is based on the data before the implementation of the response measures. The contribution of each measure is quantified by the subsequent concentration change, and an assessment report is generated. The method for quantifying the contribution of each measure is to use the ratio of the change in pollution concentration before and after the implementation of the treatment measure to the total pollution reduction as the contribution of the measure. The assessment report includes the pollution reduction rate, environmental restoration time, and contribution scores of each treatment measure. The model feedback update unit integrates the assessment report, monitoring data, and treatment records into labeled training samples and feeds them back to the cloud model training platform to update the dynamic deep neural network, knowledge graph association weights, and lightweight proxy model parameters. The labels in the labeled training samples include abnormal event type, pollution source category, treatment measure type, measure effectiveness score, and pollution reduction rate.

[0013] Furthermore, the method for constructing the lightweight environmental dynamics proxy model includes: using simulation results of a high-precision environmental dynamics numerical model covering a multi-dimensional parameter space as a training dataset; the high-precision environmental dynamics numerical model is a hydrodynamic-water quality coupling model or an atmospheric pollutant diffusion model; the multi-dimensional parameter space covers various combinations of working conditions such as pollution source intensity, meteorological conditions, and hydrological conditions; constructing a deep neural network regression model with pollution source parameters, meteorological and hydrological parameters, and spatiotemporal coordinates as inputs and pollutant concentration as output; the deep neural network regression model contains multiple hidden layers; training the deep neural network regression model using data augmentation and regularization strategies to learn an end-to-end nonlinear mapping from input to output; the data augmentation strategy involves adding noise or scaling the input parameters; the regularization strategy constrains model complexity to prevent overfitting; deploying the converged model as a lightweight proxy model, whose inference speed is at least two orders of magnitude faster than the original high-precision model, and whose inference accuracy meets a preset correlation requirement; the preset correlation requirement is that the correlation coefficient between the inference results and the actual monitoring data or the simulation results of the original high-precision model meets a threshold requirement.

[0014] Furthermore, the lightweight environmental dynamics proxy model supports an online incremental update mechanism: when the accumulated labeled training samples from the closed-loop optimization feedback module exceed a preset threshold, incremental training is triggered, and transfer learning is used to fine-tune only the parameters of the last two fully connected layers of the model, freezing the parameters of the remaining layers; in the transfer learning, all convolutional layer parameters except the last two fully connected layers are frozen; after the updated model is pre-warmed and verified to meet the accuracy requirements, it is switched to the online service model through a dual-model alternation mechanism; the pre-warming verification means that the updated model needs to be verified to meet the accuracy requirements through test samples; the dual-model alternation mechanism means that a backup model is used to maintain the inference service during the update process, and the updated model is switched to after the update is completed; the samples are pushed in real time by the closed-loop optimization feedback module, so that the model continuously adapts to environmental changes.

[0015] Specifically, this system also includes a hierarchical data storage and management module deployed on edge nodes and in the cloud to achieve hierarchical storage and management of data; the edge devices are equipped with a circular buffer for temporary storage of raw data; the edge nodes are equipped with a local database for storing suspected event data and periodic summary data; the cloud is equipped with a distributed data lake for storing full historical data, model parameters, knowledge graphs and training samples; the data storage adopts a multi-level caching strategy, with hot data stored on high-speed storage media and cold data archived to low-cost storage media.

[0016] Specifically, this system also includes a model training and version management module deployed in the cloud, used to realize model training updates and version management; the model training and version management module performs incremental training or full retraining on dynamic deep neural networks and lightweight environmental dynamics proxy models based on training samples pushed by the closed-loop optimization feedback module; the model training and version management module performs version management and performance evaluation on the trained models; the model training and version management module decides whether to issue model updates to edge nodes based on the evaluation results; the model training and version management module supports model canary release and rollback mechanisms.

[0017] Specifically, this system also includes a cloud-based user interaction and decision support module for implementing user interaction and decision support functions. This module provides a visual dashboard interface displaying real-time monitoring data, abnormal event distribution, pollution source tracing results, and diffusion simulation animations. It also supports a natural language query interface, allowing users to query environmental status, event details, and handling suggestions via voice or text input. Furthermore, the module generates various reports and supports multi-level access control.

[0018] Specifically, this system also includes a security and privacy protection module that runs through the three-layer architecture of end-edge-cloud to achieve data security and privacy protection; the security and privacy protection module encrypts data transmission between end devices and edge nodes, and between edge nodes and the cloud; the security and privacy protection module uses federated learning technology for sensitive data, uploading only model gradients instead of raw data, and adding differential privacy noise before uploading; the security and privacy protection module performs anonymization processing on stored historical data; and the security and privacy protection module records complete operation logs.

[0019] Specifically, this system also includes a clock synchronization and data consistency module deployed on each node of the edge-cloud architecture to achieve system clock synchronization and data consistency. The clock synchronization and data consistency module performs clock synchronization on each node to ensure that the time base of the entire system is consistent. The clock synchronization and data consistency module performs timestamp verification and deduplication on data that is resumed after a network outage. The clock synchronization and data consistency module adopts a distributed transaction mechanism for event data across nodes to ensure data consistency and integrity.

[0020] The technical effects and advantages of this invention are as follows: In this invention, a lightweight triggering module on the edge enables local anomaly screening, solving the problem of limited computing power and power consumption on the edge; a progressive inference module on the edge completes multi-source fusion and dynamic inference, resolving the contradiction between the difficulty of multi-source data fusion and insufficient edge computing power; a dual-engine collaborative decision-making module on the cloud runs knowledge graph tracing and digital twin inference in parallel, solving the problem of one-sided and poor interpretability of purely data-driven model decision-making; an adaptive event-driven communication module transmits data in a graded manner according to event levels, solving the problem of balancing data redundancy and real-time performance; and a closed-loop optimization feedback module compares actual data with inference results and updates the model, solving the problem of the system's inability to continuously evolve. The system as a whole achieves edge-cloud collaborative sensing, fusion inference, and closed-loop optimization, significantly improving the real-time performance, accuracy, and adaptability in the field of environmental monitoring. Attached Figure Description

[0021] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts: Figure 1 This is a flowchart illustrating the three-layer architecture and full-module collaboration process of the present invention. Figure 2 This is a flowchart illustrating the entire process of operation from edge-side exception triggering to closed-loop optimization in this invention. Figure 3 This is a diagram illustrating the adaptive operation mechanism under extreme operating conditions of the present invention; Figure 4This is a graph showing the iterative process of continuous evolution of the model in this invention and the trend of performance changes. Detailed Implementation

[0022] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0023] I. System Deployment Scenarios and Hardware Configuration like Figure 1-4 As shown, this embodiment selects a chemical industrial park and its downstream 10-kilometer watershed for system deployment verification. This scenario features a high pollution risk level, complex pollution source types, and a concentrated distribution of environmentally sensitive targets, allowing for comprehensive verification of the system's core functions such as multi-source data fusion, pollution source tracing analysis, and pollution trend prediction. The industrial park contains 32 chemical production enterprises, and a centralized drinking water source for a township is located 5 kilometers downstream. Five automatic water quality monitoring stations are deployed within the watershed, and eight automatic ambient air monitoring stations are set up at the park's boundary.

[0024] End-side device configuration: Fifteen water quality monitoring buoys were deployed at key sections of the river, comprehensively covering critical control points such as upstream inflow sections, industrial park sewage outlets, and downstream environmentally sensitive targets. Each buoy integrates a pH sensor, dissolved oxygen sensor, chemical oxygen demand (COD) sensor, ammonia nitrogen sensor, and water temperature sensor. The sampling frequency was differentiated based on the dynamic characteristics of the monitored parameters: pH, dissolved oxygen, and water temperature parameters exhibited relatively stable dynamic changes, with a sampling frequency of 5 minutes per sample; COD and ammonia nitrogen parameters showed rapid dynamic changes and were directly correlated with pollution events, with a sampling frequency of 15 minutes per sample. This sampling scheme ensured the monitoring density of key pollution indicators while avoiding excessive collection of invalid data.

[0025] Eight video surveillance cameras were installed around the company's sewage outlet, with a data acquisition frequency of 1 frame / second and an image resolution of 720p. The video data was used to help determine the sewage discharge conditions and could identify visual features such as abnormal water color and foam accumulation corresponding to abnormal sewage discharge behavior.

[0026] Two drones, equipped with multispectral cameras and gas sensors, are deployed for rapid watershed patrols during emergency response phases. As mobile monitoring nodes, the drones can compensate for the insufficient spatial coverage of fixed monitoring points.

[0027] All edge devices are equipped with an ARM Cortex-M4 microcontroller with a clock speed of 168MHz and 256KB of memory. This microcontroller has sufficient computing power to support the operation of lightweight algorithm models, while maintaining low power consumption, making it suitable for long-term outdoor deployment scenarios powered by solar energy.

[0028] Edge computing node configuration: The park management center has deployed two edge servers, each equipped with an Intel Xeon E-227 8G processor (an 8-core, 16-thread architecture), 32GB of RAM, and an NVIDIA Tesla T4 graphics processor with 16GB of video memory. This hardware configuration can support parallel processing of data streams from multiple edge devices, as well as real-time inference computations for dynamic deep neural networks.

[0029] A single edge server with equivalent configuration is deployed at the downstream water intake to ensure independent monitoring and emergency response capabilities for sensitive downstream areas.

[0030] Edge nodes and end-side devices interact with each other via 4G / 5G wireless communication networks, and are connected to the cloud platform via dedicated fiber optic lines. This hybrid networking mode can balance network coverage and data transmission stability.

[0031] Cloud server configuration: The cloud platform adopts a distributed cluster architecture, comprising 10 compute nodes. Each compute node is configured with two Intel Xeon Gold 6258R processors, 512GB of RAM, and four NVIDIA A100 graphics processors. The cloud platform needs to handle high-density computing tasks such as global algorithm model training, knowledge graph operation and maintenance management, and digital twin simulation, hence the adoption of a high-performance cluster architecture.

[0032] The storage system adopts a distributed data lake architecture with a total capacity of 1PB, which can support the storage of massive historical monitoring data and the use of algorithm model training.

[0033] II. Specific Implementation of the End-Side Lightweight Trigger Module 2.1 Derivation of Design Principles The system design phase faced a core technical challenge: limited computing resources and strict power consumption control requirements on the edge devices prevented the deployment of complex algorithm models for real-time data analysis; uploading all raw data to the cloud for processing would result in significant network bandwidth consumption and transmission latency, failing to meet the real-time early warning requirements for sudden pollution events. To address this challenge, technical analysis revealed that data in environmental monitoring scenarios falls within normal operating conditions for the vast majority of periods, with a very low percentage of abnormal events requiring identification. Therefore, this module employs a lightweight triggering mechanism. Only a small algorithm model is deployed on the edge device for initial screening of abnormal events. Normal operating data is cached locally or only a data summary is uploaded. The entire raw data is only uploaded when a suspected abnormal event is detected. This mechanism significantly reduces communication link overhead while ensuring no abnormal events are missed, enabling the design and development of a small-capacity anomaly detection model.

[0034] 2.2 Structure and Training of the Micro-volume Anomaly Detection Model In this embodiment, each end-side device is equipped with an anomaly detection model based on a deep autoencoder. This model can learn low-dimensional feature representations of normal operating data. When abnormal data is input, it will generate a significantly increased reconstruction error, which is suitable for the scenario requirements of unsupervised anomaly detection.

[0035] The specific structure of the model is as follows: The encoder part consists of a 3-layer fully connected network. The dimension of the input layer is consistent with the dimension of the input features, and the number of neurons decreases layer by layer, which can extract the core features of the data during the layer-by-layer feature compression process. The first layer has 64 neurons, the second layer has 32 neurons, and the third layer has 16 neurons. Each layer is followed by a ReLU activation function, which has the technical advantages of low computational complexity and mitigating the gradient vanishing problem.

[0036] The decoder is symmetrical to the encoder and consists of three fully connected layers. The number of neurons increases progressively with each layer: 16, 32, and 64, respectively. The output layer dimension matches the input layer dimension. The output layer uses a linear activation function to accommodate the continuous numerical output requirement of the reconstruction task, avoiding the numerical truncation problem caused by ReLU or sigmoid activation functions.

[0037] The total number of model parameters is calculated as follows: The input features cover 7 statistical features corresponding to 5 types of monitoring parameters, totaling 35 dimensions. The weight matrix from the input layer to the first layer has a dimension of 35×64, with approximately 2240 parameters after adding bias terms; the weight matrix from the first layer to the second layer has a dimension of 64×32, with approximately 2048 parameters; the weight matrix from the second layer to the third layer has a dimension of 32×16, with approximately 512 parameters; the encoder part has approximately 4800 parameters. The decoder is symmetrical to the encoder, with approximately 9600 parameters. After adding bias terms from each layer, the total number of model parameters is approximately 32,000. After structured pruning optimization, the model size can be controlled within 200KB, adapting to the deployment requirements of edge microcontrollers.

[0038] The model training process is as follows: A training set is constructed using historical normal operating condition data from the corresponding monitoring points of the edge device, with a time span of no less than 30 days. The training objective is to minimize the reconstruction error, and the mean squared error (MSE) is used as the loss function. The Adam optimizer is used during training, with a learning rate of 0.001, a batch size of 32, and 100 training epochs. After training, the model parameters are fixed and deployed to the edge device.

[0039] 2.3 Design of the Feature Extraction Submodule For different types of monitoring data, the model adopts differentiated feature extraction methods. The raw monitoring data has problems such as high dimensionality and large noise interference. Directly inputting it into the autoencoder will lead to a complex model structure and high convergence difficulty. Therefore, a feature extraction submodule is designed to convert the raw data into low-dimensional, high-information-density feature vectors.

[0040] Feature extraction of water quality sensor numerical sequences: The input window length is determined based on the persistence characteristics of pollution events and the response speed of different parameters, selecting the 12 most recent sampling points as the input window. For pH and dissolved oxygen parameters with a 5-minute sampling period, the 12 sampling points correspond to 1 hour of monitoring data; for COD and ammonia nitrogen parameters with a 15-minute sampling period, the 12 sampling points correspond to 3 hours of monitoring data. This window setting can retain sufficient historical feature information while avoiding introducing excessive analysis delay.

[0041] The extracted statistical features include mean, variance, rate of change, maximum value, minimum value, sum of absolute values ​​of first-order differences, and sum of absolute values ​​of second-order differences. These features comprehensively characterize key attributes of the data, such as central tendency, degree of fluctuation, and rate of change. The rate of change can capture sudden increases in parameters, and the sum of absolute values ​​of first-order differences reflects the overall fluctuation range of the data. The feature dimensions total 35 dimensions (5 types of parameters × 7 features). This dimensionality setting allows for control over the structural complexity of the autoencoder while preserving core feature information.

[0042] Feature extraction of video frames: Under normal operating conditions, the image acquisition frequency is 30 seconds / frame, and under abnormal triggering mode, the acquisition frequency is 1 frame / second. The low acquisition frequency is used under normal operating conditions to reduce device power consumption and storage usage; the high acquisition frequency is used in triggering mode to fully capture the detailed features of abnormal events.

[0043] The image preprocessing process is as follows: the image is scaled to 224×224 pixels to adapt to the input requirements of the MobileNetV3 network; the pixel values ​​are normalized to the range of 0 to 1 to improve the convergence stability of the model.

[0044] The feature extraction network employs the first three convolutional layers of the MobileNetV3-Small network, which boasts advantages such as low computational cost and few parameters, making it suitable for mobile deployment scenarios. The output feature map undergoes global average pooling to obtain a 128-dimensional feature vector. Global average pooling enables spatial information aggregation, outputting a fixed-length feature vector.

[0045] Feature extraction from audio data: The audio sampling rate was set to 16kHz, and the frame length was set to 3 seconds, corresponding to 48,000 sampling points. The 3-second frame length setting can fully cover the audio of abnormal device operation associated with the pollution event, avoiding the loss of feature information due to too short a frame length, or the increase in computational complexity due to too long a frame length.

[0046] The audio preprocessing process is as follows: the audio signal is framed and windowed, with a frame length of 25ms and a frame shift of 10ms. These parameters are common standard parameters for speech signal processing and can completely preserve the time-frequency domain feature information of the audio.

[0047] The feature extraction process is as follows: extract 40-dimensional Mel frequency cepstral coefficients, which are a common feature in the field of audio processing and can effectively characterize the spectral characteristics of audio; then further reduce the dimensionality through a 2-layer convolutional neural network (CNN) to output a 32-dimensional feature vector.

[0048] 2.4 Calculation Logic of Dynamic Adaptive Threshold The core of anomaly detection is threshold setting. Fixed thresholds cannot adapt to the dynamic changes in the statistical characteristics of environmental data with the seasons and diurnal rhythms, which can easily lead to a large number of false alarms or missed anomalies. Therefore, this module is designed with a dynamic adaptive threshold mechanism.

[0049] The edge model uses reconstruction error as the basis for anomaly detection for input samples. The model outputs reconstructed values. The mean square error of both is calculated as the reconstruction error. Under normal operating conditions, the reconstruction error is... It follows a specific statistical distribution; under abnormal operating conditions, the reconstruction error A significant increase will occur. The formula for calculating the dynamic adaptive threshold is as follows: , in, This is a dynamically adaptive threshold used to determine whether the current data is abnormal; The mean of the reconstruction error represents the average level of the model reconstruction error under normal conditions. It is calculated by averaging the reconstruction errors of all normal samples within the sliding time window. This is the sensitivity coefficient, used to control the tolerance of the threshold to normal fluctuations. In this example, it is set to 3.0, which corresponds to a 99.7% confidence interval under a normal distribution, and can balance the false alarm rate and the false negative rate. Let be the standard deviation of the reconstruction error, describing the range of fluctuation of the reconstruction error under normal conditions, and... Calculated within the same time window, reflecting the stability of normal data; The historical baseline signal-to-noise ratio represents the data quality during normal historical periods, and is related to... , Calculated within the same time window, it serves as a reference benchmark for signal-to-noise ratio adjustment; The current real-time signal-to-noise ratio (SNR) represents the data quality at the current moment. It is obtained through a local window SNR estimation algorithm and is used to dynamically adjust the threshold to cope with noise interference.

[0050] When real-time reconstruction error Exceeding the dynamic threshold If the event is detected as a suspected abnormal event, the full data upload process will be triggered.

[0051] The reasoning process for this formula is as follows: First, the mean of the reconstruction error... and standard deviation It describes the distribution range of normal data. A basic confidence interval is given. However, data quality affects reconstruction error: when the signal-to-noise ratio (SNR) is low, i.e., noise is high, the reconstruction error will increase even for normal data, so the threshold needs to be increased to avoid false alarms; conversely, when the SNR is high, the threshold can be appropriately lowered to improve sensitivity. The signal-to-noise ratio (SNR) is introduced. This is precisely to achieve such adaptive adjustment.

[0052] 2.5 Data Digest Generation and Local Caching For routine operating data that does not trigger anomalies, the edge device generates a data summary every hour. When there are no abnormal events, basic feature information should still be retained for long-term trend analysis and model iteration optimization; there is no need to upload the full original data.

[0053] The data summary includes the following: timestamp intervals, including the start and end times of the data; mean, variance, and rate of change of various monitoring parameters. These statistics can fully characterize the data distribution characteristics within the time period and meet the needs of long-term trend analysis; and data quality labels, including three categories: normal, slight fluctuation, and sensor anomaly. The quality labels are obtained through sensor self-testing algorithms, and the self-testing indicators include data missing rate, number of mutations, etc.

[0054] Summary data is stored in a local circular buffer with a capacity set to 24 hours of data. This 24-hour storage capacity covers the duration requirements of most network outage scenarios, while the relatively small data volume is well-suited to the storage capabilities of edge devices. Once the buffer is full, it automatically overwrites the oldest historical data.

[0055] When network communication is normal, data summaries are uploaded in batches every 4 hours. This upload frequency is a balanced design between device power consumption and data timeliness, avoiding frequent wake-ups of the communication module that increase power consumption, while ensuring the timeliness of cloud data. When network communication is interrupted, data is continuously stored in a local buffer. After communication is restored, data is automatically re-entered based on timestamps to avoid duplicate data uploads.

[0056] III. Specific Implementation of the Side-by-Side Progressive Reasoning Module 3.1 Design Principles Edge nodes receive uploaded data from multiple end-side devices and need to complete the fusion of multi-source heterogeneous data and real-time inference operations. The core technical challenges to be solved are: first, the heterogeneous fusion of data from different devices in terms of sampling frequency, spatial location, and data type; and second, the optimization of inference efficiency under the condition of limited computing resources of edge nodes.

[0057] To address the challenge of multi-source data fusion, this module adopts a technical approach of first aligning spatiotemporally and then fusing features: First, spatiotemporal alignment maps heterogeneous multi-source data to a unified time and space coordinate system; then, the core features of each modality are extracted; finally, an attention mechanism is used to achieve adaptive fusion of multimodal features. This hierarchical processing architecture eliminates data heterogeneity through the alignment stage, preserves the core information of each modality through the feature extraction stage, and automatically learns the weight proportions of different modalities through the attention fusion stage.

[0058] To address the issue of inference efficiency optimization, this module designs a dynamic deep neural network architecture, enabling rapid exit from simple samples and deep inference for complex samples. This architecture design is based on the data characteristics of environmental monitoring scenarios; the vast majority of monitoring data represents normal operating conditions and can be identified with high confidence using shallow networks, while only a few abnormal and complex samples require analysis through deep networks. The early exit mechanism significantly reduces the computational resource consumption for normal operating condition samples, thereby improving overall inference efficiency.

[0059] 3.2 Specific Implementation of Spatiotemporal Alignment of Multi-Source Data Time alignment: The sampling frequencies of different end-devices vary significantly. Water quality buoys sample every 5 to 15 minutes, video surveillance samples at 1 frame per second, and drones collect data on demand. To address the need for a unified timeline for multi-source data, this module employs a dynamic time warping algorithm.

[0060] The core principle of the dynamic time warping algorithm is to minimize the cumulative distance between two time series by nonlinearly stretching or compressing the time axis. In this scenario, a 1-minute time axis is used as the baseline. For each 1-minute time window, the average of all sampling points within the window is taken as the aligned value. The 1-minute time granularity can meet the timeliness requirements of emergency response and cover the fluctuation cycle of most environmental parameters.

[0061] If a sensor has no valid sampled data within a corresponding time window, linear interpolation is used to complete the data. For example, if the COD parameter sampling period is 15 minutes, the intermediate window with no data between two sampling points is filled by linear interpolation of the two sampling points. This method has low computational complexity and can achieve smooth estimation of intermediate values.

[0062] Spatial alignment: The GPS coordinate accuracy varies among different end devices, with some devices offering meter-level accuracy and others offering ten-meter-level accuracy. To address the spatial correlation requirements of multi-source data, this module employs a geo-hash encoding method.

[0063] The core principle of geohashing is to recursively encode two-dimensional latitude and longitude coordinates into strings, with coordinates having the same string prefix being spatially adjacent. In this embodiment, the grid resolution is set to 50 meters, corresponding to a geohashing code length of 6 bits. A 50-meter grid resolution can accurately capture the hourly diffusion details of pollutants in river channels, while avoiding a surge in computational load caused by an excessive number of grids.

[0064] Devices with the same geohash code are considered to be in the same spatial location. For pollution events that cross grids, a weighted average method that is inversely proportional to the distance is used to process the monitoring data of adjacent grids to ensure data continuity at grid boundaries.

[0065] 3.3 Specific Implementation of Multimodal Feature Extraction Feature extraction from numerical data: The input data includes 15 monitoring points, 5 types of water quality parameters, and monitoring data from the most recent 60 minutes. The original dimension is 4500. Directly inputting it into the model would result in excessive dimensionality and computational complexity. Therefore, dimensionality compression needs to be achieved through feature extraction.

[0066] Time-domain statistical feature extraction: For each monitoring parameter, the mean, variance, rate of change, and peak value are calculated within four time windows: 5 minutes, 15 minutes, 30 minutes, and 60 minutes. The multi-window setting is based on the following: the 5-minute window captures instantaneous changes in the parameter; the 15-minute window matches the COD parameter sampling period; the 30-minute window reflects short-term trends; and the 60-minute window reflects overall change patterns. The statistical features are: mean, used to characterize the average level of the parameter; variance, used to characterize the degree of fluctuation of the parameter; rate of change, used to characterize the overall speed of change of the parameter; and peak value, used to capture the maximum concentration value of the parameter.

[0067] The total dimensions of the time-domain statistical features are 4 windows × 5 types of parameters × 4 features, totaling 80 dimensions.

[0068] Frequency domain feature extraction: Wavelet packet decomposition is employed, a method that can simultaneously analyze both low-frequency and high-frequency components of the signal, adapting to the capture requirements of fluctuation patterns at different time scales. A three-level wavelet packet decomposition is used to obtain eight frequency bands, and the energy features of each band are extracted. The db4 wavelet is selected as the wavelet basis, as it possesses good time-frequency localization performance. The total dimension of the frequency domain features is 40 dimensions, consisting of 5 types of parameters × 8 frequency bands.

[0069] The total dimension of the numerical features is 120.

[0070] Image data feature extraction: The input data includes keyframes from 8 cameras over the past 5 minutes, totaling 300 frames per second. Directly processing all image frames would result in excessive computational overhead; therefore, single-frame image features are extracted first, followed by temporal pooling.

[0071] The feature extraction network uses the output of the 8th convolutional block of the MobileNetV3-Small network. Shallow layers can extract low-level features such as edges and textures, while deep layers can extract semantic features. The 8th convolutional block is located in the middle of the network, which can balance the extraction of detailed features and semantic features. A single frame image outputs a 128-dimensional feature vector.

[0072] Temporal pooling was performed on the feature sequence of 300 frames of images, and the maximum value, mean, and variance were calculated respectively. The maximum value captures the most significant anomalous features, the mean reflects the overall feature level, and the variance reflects the dynamic change of features. After pooling, 3×128 dimensions, totaling 384 dimensions of video features, were obtained.

[0073] Text data feature extraction: Input data includes text data such as enterprise emission reports, operation and maintenance logs, and alarm records. The ALBERT-tiny model is used to extract semantic features. This model has a small number of parameters, fast inference speed, and is suitable for edge node deployment scenarios. The model outputs a 128-dimensional feature vector. When there is no valid text data, this part of the feature vector is set to zero.

[0074] 3.4 Specific Implementation of Multi-Head Cross-Attention Fusion Across Modalities The core challenge of multimodal fusion lies in the dynamic adaptation of the importance of different modalities. Traditional methods such as simple concatenation or fixed-weighted averaging cannot adapt to the dynamic changes in the amount of information in each modality under different scenarios. For example, the effectiveness of video monitoring data decreases in nighttime environments, and the corresponding weights should be adjusted accordingly; when the sensor signal-to-noise ratio is low, the weights of the corresponding numerical modalities should also be adjusted accordingly. Based on this, this module adopts a multi-head cross-attention mechanism to achieve adaptive dynamic learning of the weights of each modality.

[0075] The fusion module is designed as follows: First, a fully connected layer maps the features of each modality to a unified dimension, set to 256 dimensions. The fully connected layer performs a linear transformation, ensuring that the dimensions of each modality feature remain consistent during subsequent attention calculations.

[0076] The multi-head attention layer uses eight attention heads, each with a 64-dimensional dimension. This eight-head configuration allows for the simultaneous capture of modal correlations across different feature subspaces while keeping computational costs within a reasonable range.

[0077] The attention calculation process is as follows: Linear mapping generates query, key, and value vectors: For the i-th target modality, a query vector is generated. ; Regarding the first Each source mode generates a key vector. Value vector In the formula: Number the target mode. Number the source mode; , The first The, the The original feature vector of a single modality after spatiotemporal alignment; , , These are the learnable linear transformation weight matrices corresponding to the query, key, and value, respectively, which are automatically optimized and converged during model training. This is the query vector used to initiate cross-modal feature matching; This is a key vector used to provide matching identifiers; This is a value vector used to provide the feature content to be fused.

[0078] Calculate the basic semantic attention weights: obtain the weights by scaling the dot product and normalizing. The modality pair of the first The basic semantic attention weights for each modality are calculated using the following formula: In the formula: These are the basic semantic attention weights, with values ​​ranging from [0,1], and the sum of the weights for all source modalities is 1. This is a normalized activation function used to convert relevance scores into valid weights; Key vector The transpose of the matrix; The dot product of the query vector and the key vector is used to quantify the semantic relevance between modalities; This is a scaling factor used to ensure numerical stability and proper gradient propagation in the model. The feature dimensions of the query and key vector.

[0079] Generating cross-modal fusion features: The final fusion feature is the weighted sum of all modal value vectors, calculated using the following formula: , in For the first The final cross-modal fusion feature vector of each modality, This represents the total number of modalities participating in the fusion.

[0080] The output layer reduces the dimensionality of the fused features to 128 dimensions, matching the input dimension of the first layer of the dynamic deep neural network.

[0081] The dynamic adjustment rules for attention weights are as follows: modalities with higher signal-to-noise ratios have better data quality and should be assigned higher weights; modalities with higher historical inference confidence have more stable recognition performance and should also be assigned higher weights. The weight calculation formula is: , in, For the first to participate in cross-modal fusion The modality number; For the first The adaptive quality correction weights for each modality, with values ​​ranging from [value range] to [value range] for all participating fusion modalities. The sum is 1; This is a normalized activation function used to convert the modal quality composite score into valid normalized weights; The signal-to-noise ratio (SNR) weighting coefficient is 0.6 in this embodiment, representing the relative importance of the real-time SNR in modal quality evaluation. The historical confidence weighting coefficient, in this embodiment, is set to 0.4, representing the relative importance of historical inference confidence in modal quality evaluation, and satisfying the following conditions: ; For the first The real-time signal-to-noise ratio estimate of each modality data, after min-max normalization, takes a value range of [0,1]. For the first The average confidence of the inference results of each modality over the past 24 hours, after min-max normalization, ranges from [0,1]. The final effective attention weight is the element-wise product of the basic semantic relevance weight and the modality quality correction weight. The value vectors of each modality are then weighted and summed based on the effective attention weight to obtain the final cross-modal fusion feature.

[0082] 3.5 Specific Structure of Dynamic Deep Neural Networks The design goal of dynamic deep neural networks is to minimize computational resource consumption while ensuring recognition accuracy. The inference logic is as follows: for simple samples, high-confidence recognition results can be achieved through shallow network computation, allowing the inference process to exit early; for complex samples, deep networks are used to complete the computation. This architecture is highly adaptable to environmental monitoring scenarios, with a high proportion of simple samples under normal operating conditions, significantly improving overall inference efficiency.

[0083] Multi-scale hollow convolutional layer design The reason for using a dilated convolutional structure is that traditional convolutional structures have a limited receptive field, requiring multiple layers to capture a wide range of contextual information. Dilated convolution, by setting gaps between kernel elements, can expand the receptive field of the convolutional kernel without increasing the number of parameters. This embodiment sets up regular convolution and dilated convolution in parallel, which can capture both local detailed features and global contextual information simultaneously.

[0084] Layer 1: The input is a 128-dimensional feature map with 64 channels, output after feature fusion, and contains two parallel branches. Branch A is a regular 3×3 convolution with a stride of 1, padding of 1, and 32 output channels, responsible for extracting local detail features; Branch B is a dilated 3×3 convolution with a dilation rate of 2, a stride of 1, padding of 2, and 32 output channels, responsible for extracting global context features. The outputs of the two branches are concatenated along the channel dimension to obtain a 64-channel feature map.

[0085] Layer 2: Input is a 64-channel feature map, output is 128 channels, and the dilation rate is set to 4. The design of increasing the number of channels is based on the fact that as the network depth increases, more feature channels are needed to express complex feature patterns.

[0086] Layer 3: Input is a 128-channel feature map, output is 256 channels, and the dilation rate is set to 6. The design principle of progressively increasing dilation rate is that shallow networks need to extract fine local features, so the dilation rate should not be too large; deep networks need to capture a wider range of contextual information, so the dilation rate needs to be increased.

[0087] Each layer is followed by a batch normalization layer and a ReLU activation function. Batch normalization can accelerate model training and improve convergence stability, while the ReLU activation function introduces nonlinear transformation capabilities.

[0088] Early exit branch design Each layer is followed by an early exit branch, with the following branch structure: a global average pooling layer that compresses the feature map into a fixed-length vector to reduce the number of parameters; a dropout layer to prevent overfitting, with an adaptive adjustment of the dropout rate; a fully connected layer with 128 neurons for further feature extraction; a ReLU activation function that introduces a non-linear transformation; a fully connected layer with an output dimension consistent with the number of categories, including five categories: normal operating conditions, COD exceeding standards, ammonia nitrogen exceeding standards, combined pollution exceeding standards, and equipment malfunction; and a Softmax output layer that outputs the probability distribution of each category.

[0089] The confidence thresholds are set as follows: Layer 1 threshold 0.98, Layer 2 threshold 0.95, and Layer 3 threshold 0.92. The reason for the progressively decreasing thresholds is that deeper networks extract richer feature information, resulting in generally higher recognition confidence; a lower threshold can be set to allow for early exit. The thresholds can be dynamically adjusted based on historical inference results, with a single adjustment increment of 0.01. If a certain layer frequently achieves early exit while maintaining the required recognition accuracy, the threshold can be appropriately lowered to further improve inference efficiency.

[0090] Adaptive Dropout drop rate calculation: Standard Dropout uses a fixed dropout rate, but the variance of feature maps varies across different layers and samples, making a fixed dropout rate unsuitable for the regularization requirements of different scenarios. High-variance feature maps contain richer feature information but are more prone to overfitting, requiring a higher dropout rate; low-variance feature maps have lower information density, and an excessively high dropout rate can easily lead to the loss of valuable information. Therefore, this module designs an adaptive Dropout mechanism based on feature map variance.

[0091] The specific calculation process is as follows: For the feature map output by each layer group, calculate the mean variance of the channel dimension, that is, take the average variance of each channel; Set reference variance The value is the global average variance of all samples in the training set. In this embodiment, the value is 0.15, which reflects the characteristic fluctuation level under normal operating conditions. The formula for calculating the discard rate is: , The coefficient 0.5 is an empirical value used to map the variance ratio to a reasonable dropout rate range; the dropout rate is limited to the range of 0.1 to 0.4 to avoid the loss of effective information due to an excessively high dropout rate, or the insufficient regularization effect due to an excessively low dropout rate. Example calculation: when When =0.3, the calculation yields =1.0, but due to upper limit restrictions, the actual value is 0.4; when When p = 0.05, the calculated value is p ≈ 0.167, and the actual value is 0.167.

[0092] 3.6 Example of Reasoning Process Taking a COD pollution incident as an example, the complete reasoning process is as follows: The data uploaded by the edge node receiving end is processed by spatiotemporal alignment and fused with multimodal features to generate a 128-dimensional fused feature vector; The fused feature vector is input into the dynamic deep neural network: After calculation by layer group 1, the branch exits early with a confidence score of 0.76, which is lower than the layer group 1 threshold of 0.98, and the forward propagation continues; after calculation by layer group 2, the branch exits early with a confidence score of 0.93, which is lower than the layer group 2 threshold of 0.95, and the forward propagation continues; after calculation by layer group 3, the branch exits early with a confidence score of 0.98, which is higher than the layer group 3 threshold of 0.92, and the inference calculation is terminated. The output identification result is a COD exceeding the standard event with a confidence level of 0.98. At the same time, the feature vector of layer group 3 is output for subsequent dynamic calculation of confidence level.

[0093] The complete inference process for this event took 43ms, meeting the real-time requirements. If the input sample is normal operating data, it is highly likely to exit early in layer 1 or layer 2, with even shorter inference time, fully demonstrating the efficiency advantages of dynamic networks.

[0094] IV. Specific Implementation of the Cloud-based Dual-Engine Collaborative Decision-Making Module 4.1 Derivation of Design Principles The core cloud-based functions are divided into two main dimensions: pollution event attribution analysis and pollution diffusion trend prediction. The former requires combining domain expertise to trace the source of pollution and identify the root cause of the pollution event; the latter requires simulating the diffusion process based on environmental dynamics and physical models to predict the subsequent development trend of the pollution event. Technical validation has shown that implementing these two functions using a dual-engine parallel collaborative architecture can produce significant synergistic effects: the source tracing results from the knowledge graph engine can provide boundary conditions such as pollution source location and emission intensity for the digital twin prediction engine; the simulation results from the digital twin prediction engine can, in turn, verify the rationality of the source tracing results. If the prediction results do not match the actual monitoring data, it can be determined that the source tracing results are biased, and a new source tracing analysis needs to be conducted. Based on this, this module designs a dual-engine parallel collaborative processing mechanism.

[0095] 4.2 Specific Implementation of the Knowledge Graph Source Tracing Engine Knowledge Graph Construction The core function of a knowledge graph is to structurally represent professional knowledge in the environmental domain, supporting machines in performing automated reasoning and analysis. The knowledge graph constructed in this embodiment includes the following entity types and relationships: The monitoring indicator entities cover 20 monitoring indicators, including COD, ammonia nitrogen, and pH. Each indicator node is associated with attributes such as standard thresholds and rules for judging exceedances. Relevant information is sourced from national environmental protection standards and expert knowledge in the field. The abnormal event entities cover 8 event types, including indicator exceedances, abnormal discharges, and pollutant leaks. Each event node is associated with a corresponding abnormal data pattern; for example, a sudden increase in COD concentration corresponds to an abnormal discharge event by an industrial enterprise. The pollution source entities cover 32 chemical enterprises within the park. Each enterprise node is associated with attributes such as discharge outlet coordinates, emission characteristics, and historical environmental violations. Relevant information is sourced from enterprise registration data and environmental enforcement records. The environmental factor entities cover dynamic environmental parameters such as wind speed, wind direction, water temperature, and river flow, which are dynamically updated through real-time monitoring data. The disposal measure entities cover 15 types of emergency measures, including production unit shutdown, production restriction and emission reduction, emergency monitoring, and personnel evacuation. Each measure node is associated with attributes such as applicable scenarios and expected disposal effects.

[0096] The design of the related edges is as follows: Causal edges represent the causal relationship between entities, such as a company's wastewater discharge causing water quality indicators to exceed standards, or water dilution leading to a decrease in pollutant concentration. The initial edge weight is set to 0.6 to 0.8, based on domain expert knowledge. Spatial proximity edges represent the spatial relationship between entities, such as a company located upstream of a river, or a wastewater outlet within 500 meters of a monitoring point. The edge weight is set to 0.4 to 0.6, determined based on spatial distance calculations. Regulatory related edges represent the regulatory constraints between entities, such as pollutant emission standard limits and environmental violation penalty clauses. The edge weight is set to 0.8 to 0.9, corresponding to the deterministic constraint attribute of regulatory clauses.

[0097] The knowledge graph is stored using the Neo4j graph database, which is deeply optimized for graph relationship queries and adapted to the needs of depth-first path search scenarios.

[0098] Explainable path generation process Taking a COD pollution incident as an example, the process of generating a source tracing path can be explained as follows: The graph reasoning adaptation unit converts pollution event information into graph database query statements, maps key information such as the event's exceeding indicators, monitoring points, and occurrence time into the starting node of the graph query, and retrieves pollution source entities with causal relationships by using the COD indicator node as the query starting point.

[0099] The path retrieval is performed using a depth-first graph search algorithm, which completes a full graph traversal along causal edges and spatially adjacent edges. The depth-first search algorithm can quickly locate the complete source path and facilitates retrieval pruning optimization.

[0100] Search pruning optimization: To improve search efficiency, nodes that are not related to the current event are removed, and nodes whose spatiotemporal coordinates deviate from the event by more than 500 meters are directly removed. This threshold is determined based on the hourly diffusion capacity of pollutants. Pollution source entities that have no causal relationship with the exceeding indicators are directly removed, such as enterprise nodes with no history of COD emissions.

[0101] Path confidence quantification: For multiple tracing paths retrieved, a confidence score is calculated for each path. The weighting rules are as follows: Topology length accounts for 20% of the weight; the longer the path, the higher the uncertainty. The weight is the normalized result of the inverse of the path length. The historical confidence of intermediate nodes accounts for 40% of the weight. The historical confidence of a node is determined based on the frequency and accuracy of the historical tracing events of that node. For example, if a company was identified as a pollution source in 2 out of 3 historical tracing events, the corresponding confidence is 0.67. The weight of associated edges accounts for 40% of the weight, which is the sum of the initial weight of the edge and the historical update value.

[0102] The credibility score of each path is obtained by weighted summation, and at least two paths with the highest scores are selected as the final tracing results.

[0103] 4.3 Specific Implementation of the Digital Twin Inference Engine Construction of a four-dimensional digital twin A digital twin is a digital mapping of a real watershed environment. It must contain complete geographical, hydrological, and meteorological information in order to accurately simulate the pollution diffusion process.

[0104] The basic geographic information uses 30-meter resolution digital elevation model (DEM) data, which is publicly available high-precision topographic data. River system vector data is overlaid simultaneously to determine water flow paths. Hydrological data uses measured data from river cross-sections, covering river width, water depth, and flow velocity parameters at multiple cross-sections. A water level-discharge relationship curve is established based on the measured data, allowing for the extrapolation of river flow based on real-time water level data. Meteorological data is integrated from real-time monitoring data from local meteorological stations, covering wind direction, wind speed, ambient temperature, and rainfall. These parameters directly affect the diffusion process of pollutants in water and the atmosphere. Grid division: The entire study area is divided into a 100m × 100m grid to suit the spatial scale of pollutant diffusion in the watershed; the grid in the core river area is densified to 20m × 20m to meet the high-precision extrapolation requirements of narrow river channels.

[0105] Lightweight environmental dynamics proxy model construction Traditional environmental dynamics models, such as the Environmental Fluid Dynamics Code (EFDC), offer high simulation accuracy but are computationally time-consuming and cannot meet the demands of real-time simulations. Therefore, this module constructs a lightweight proxy model that uses machine learning methods to fit the input-output mapping relationship of the original physical model.

[0106] The model building process is as follows: Training Dataset Construction: Training the surrogate model requires a large amount of sample data covering all operating conditions. However, the number of samples from real pollution events is extremely small, making it impossible to cover various combinations of operating conditions. Therefore, a high-precision environmental dynamics model is used to generate a simulation dataset offline. The input parameter space covers the full range of operating conditions that may occur in real-world scenarios, including pollution source intensity from 0.1 to 100 mg / s, river flow from 1 to 50 m³ / s, water temperature from 5 to 30℃, and wind speed from 0 to 10 m / s. A Latin hypercube sampling method is used to generate 100,000 sets of input parameter combinations. The corresponding output results are obtained through high-precision model simulation, thus constructing a complete training dataset.

[0107] Proxy model architecture design: A fully connected neural network architecture is adopted to adapt to the regression task requirements of numerical parameter input and pollutant concentration output. The network consists of 5 fully connected layers with 256, 512, 512, 256 and 1 neurons respectively. The input dimension is 16-dimensional, covering spatial coordinates, time, pollution source intensity, hydrological parameters, etc., and the output dimension is 1-dimensional, corresponding to the pollutant concentration value.

[0108] Model training settings: The Adam optimizer is used, the learning rate is set to 0.001, the training epochs are 100, and training is stopped early when the validation set loss drops to 0.002 to avoid model overfitting.

[0109] Model compression and optimization: The trained model needs to be pruned and quantized before it can be deployed in the real-time inference service. Structured pruning removes neurons that contribute little to the output, with a pruning ratio of 40%; INT8 quantization converts the model weights from 32-bit floating-point numbers to 8-bit integers, reducing the model size from MB to KB and the single-step inference time from seconds to milliseconds.

[0110] 4.4 Intelligent Decision Package Generation After the dual-engine parallel processing is completed, the source tracing results and the simulation results are merged to generate an intelligent decision package. The fusion logic is as follows: the source tracing results clarify the source and cause of the pollution incident, the simulation results predict the scope and development trend of the pollution incident, and the combination of the two can form a complete and implementable emergency response plan.

[0111] Taking a COD pollution incident as an example, the decision package generation logic is as follows: The source tracing result indicates that a certain enterprise is the most likely suspect for abnormal discharge, with a credibility of 71%, while agricultural non-point source pollution is the secondary suspect, with a credibility of 55%; the projection result indicates that the pollutants will affect downstream drinking water sources in 2.5 hours, with a peak concentration of 0.8 mg / L; the disposal recommendation generation logic is as follows: immediately conduct on-site law enforcement verification of the enterprise, as this enterprise is the primary suspected source of pollution and the source tracing credibility is high; initiate emergency intensified monitoring of drinking water sources, as the pollutants will affect the water sources in 2.5 hours, and it is necessary to monitor water quality changes in real time; issue water quality early warning information to downstream water supply plants, as although the peak concentration of pollutants has not exceeded the standard, it is necessary to prepare for risk prevention and control in advance; if the on-site verification confirms abnormal discharge behavior, initiate administrative penalty procedures in accordance with the law to form a complete closed-loop chain of evidence.

[0112] The intelligent decision package is simultaneously pushed to law enforcement officers' mobile terminals and stored in the event database for subsequent system closed-loop optimization.

[0113] V. Specific Implementation of the Adaptive Event-Driven Communication Module 5.1 Derivation of Design Principles Different levels of pollution events have varying requirements for the real-time performance, reliability, and bandwidth of data transmission: confirmed pollution events require real-time and reliable transmission, directly linked to emergency response; suspected abnormal events can have their transmission delayed, awaiting further analysis and confirmation; routine operating data only requires the upload of summary information, primarily for long-term trend analysis. Simultaneously, the wireless communication network status fluctuates dynamically, necessitating the design of an adaptive transmission mechanism to match these changes. Based on this, this module designs a three-level event-driven transmission strategy and a bandwidth adaptive adjustment mechanism.

[0114] 5.2 Three-level event-driven transmission strategy Confirm the event; the event confidence level is not less than 0.95. The transmission priority is the highest, prioritizing the use of dedicated 5G / fiber optic transmission channels. If these channels become unavailable, it automatically switches to a backup 4G / satellite communication channel. 5G communication offers the advantages of low latency and high bandwidth, while satellite communication serves as a backup in extreme scenarios. The transmitted content is a full data packet, including raw monitoring data, feature vectors, inference results, and evidence chain hash values. The evidence chain hash value ensures data integrity and supports subsequent event tracing and law enforcement applications. The retransmission mechanism automatically triggers data retransmission if no confirmation is received from the cloud within 5 seconds, with a maximum of 3 retransmissions. The 5-second timeout threshold is determined by considering both network round-trip latency and cloud processing time.

[0115] Suspected incidents, with a confidence level of not less than 0.80 and not less than 0.95: Data is stored in a local database on the edge node for 30 days, which meets the needs of the entire process of manual review and subsequent analysis. Upload is triggered by cloud access commands, cross-regional collaborative analysis requirements, and manual review commands, avoiding unnecessary data uploads that consume bandwidth resources.

[0116] For routine events, the event confidence level is below 0.80: Data summaries are uploaded in batches every hour, a frequency designed to balance device power consumption and data timeliness. Local storage on edge nodes lasts for 7 days, while cloud storage lasts for 1 year, after which data is automatically archived. The 1-year storage period meets the needs of long-term trend analysis while controlling cloud storage resource consumption.

[0117] 5.3 Bandwidth Adaptive Transmission Mechanism Edge nodes monitor network bandwidth every 5 seconds and calculate the moving average bandwidth. When the bandwidth falls below the 10Mbps threshold, an adaptive bandwidth adjustment mechanism is activated. The specific adjustment strategies are as follows: Prioritize uploading 128-dimensional key feature vectors to replace the full upload of raw data. The feature vectors already contain the core information required for cloud analysis, reducing the data volume from MB to KB. Image data is compressed, reducing JPEG image quality from 95% to 70% and resolution from 720p to 480p. The difference in human visual perception is small, and the data volume can be reduced by 80%. Video data is transmitted by frame extraction, reducing the sampling frequency from 1 frame / second to 1 frame / 5 seconds. This sampling frequency can still completely capture the key changing features of pollution events.

[0118] 5.4 Disconnection Resume Mechanism Both the endpoint and edge devices are equipped with 128GB solid-state drives for local data caching, which can store several months of summary data and several days of full raw data. When network communication is interrupted, the endpoint devices organize all data into files according to timestamps and store them locally. The edge nodes receive the endpoint data and store it synchronously, while marking it as pending upload.

[0119] After network communication is restored, the device automatically scans local cached data and completes data supplementation in chronological order. During the supplementation process, newly generated real-time data is uploaded first, while historical data awaiting supplementation is transmitted during periods of bandwidth idleness to ensure the timeliness of real-time monitoring data. The data conflict handling rule is as follows: if valid data with the same timestamp already exists in the cloud, the cloud data shall prevail and a conflict log shall be recorded to avoid duplicate data uploads and overwriting.

[0120] VI. Specific Implementation of the Closed-Loop Optimization Feedback Module 6.1 Derivation of Design Principles The natural environment is constantly changing, and the algorithm model needs to maintain stable analytical performance through continuous iterative optimization. This module quantifies the actual effectiveness of various emergency response measures by comparing the actual handling effects of pollution incidents with theoretical deductions. The effectiveness evaluation results are used as feedback signals to complete the iterative updates of the algorithm models of each module, constructing a closed-loop optimization system for the entire process of monitoring, decision-making, evaluation, and optimization, enabling the system to have long-term adaptive evolution capabilities.

[0121] 6.2 Actual Data Acquisition and Deviation Calculation After a pollution incident occurs, the system automatically tracks continuous monitoring data for 72 hours after the incident. The 72-hour duration can cover the complete handling and water quality restoration process of most pollution incidents.

[0122] The deviation calculation process is as follows: The theoretical pollutant decay curve derived from the digital twin model is extracted. This curve is generated based on the simulation of natural decay conditions without human intervention. The deviation between the actual monitoring curve and the theoretical curve can directly reflect the actual effect of artificial emergency response measures. The larger the deviation, the more significant the pollution reduction effect of the response measures. The overall effectiveness of the measures can be quantified by calculating the total deviation through integration.

[0123] The calculation indicators used are as follows: relative deviation, representing the percentage of concentration deviation at each time point; integral deviation. The trapezoidal integral method is used to perform approximate calculations, where, t represents the relative deviation at time t; the pollution reduction rate characterizes the overall decrease in pollutants throughout the entire event cycle; the deviation value quantifies the additional pollution reduction effect brought about by emergency response measures.

[0124] 6.3 Evaluation of the effectiveness of the measures The evaluation logic for the contribution of each treatment measure is as follows: assuming that the change in pollutant concentration before and after the implementation of the measure is mainly caused by the measure, the contribution of the measure is measured by the ratio of the amount of pollutant reduction during the implementation period to the total amount of reduction of the event.

[0125] The specific evaluation process is as follows: Record the execution timeline of each emergency response measure; The entire event cycle is divided into multiple time periods, each corresponding to the execution cycle or natural decay cycle of a specific response measure; Calculate the amount of pollutant reduction in each time period; Contribution of a single measure = Pollutant reduction amount during that period / Total pollutant reduction amount of the event.

[0126] The contribution score can be mapped to a range of 0 to 100 points. For example, a contribution score of 0.4 or higher corresponds to 90 points, and a contribution score of 0.3 to 0.4 corresponds to 80 points. These scores are used for label assignment during subsequent model training.

[0127] 6.4 Model Feedback Update The data from the entire incident handling process and the results of the performance evaluation are integrated into labeled training samples. The sample labels include: the type of abnormal event, such as abnormal industrial discharge, agricultural non-point source pollution, equipment failure, etc.; the type of pollution source, such as the corresponding enterprise, non-point source area, etc.; the type of handling measures, such as production plant shutdown, production restriction and emission reduction, emergency monitoring, etc.; the effectiveness score of the measures; and the pollution reduction rate.

[0128] Training samples are pushed to the cloud-based model training platform, triggering incremental updates to the algorithm models of each module. The specific update rules are as follows: Dynamic deep neural network update: Incremental training is used with a learning rate of 0.001 and 50 iterations. Incremental training only updates some parameters of the network, avoiding the huge computational overhead of full retraining. The smaller learning rate also avoids destroying the historical feature knowledge already learned by the model.

[0129] Knowledge graph association weight updates: The weights of association edges are adjusted based on the final verification results of the events. If a company is confirmed as a source of pollution, the weight of the corresponding causal association edge is increased; if a company is ruled out as a pollution suspect, the weight of the corresponding association edge is decreased. The weight update uses an exponential moving average method, with the following formula: , in, , For the new and old weights, The retention ratio representing historical weights, The value is the indication value for this event. Confirmed pollution is assigned a value of 1, and suspected pollution is assigned a value of 0.

[0130] Lightweight proxy model fine-tuning: New event samples are added to the training set, and transfer learning is used to fine-tune only the last two fully connected layers of the network. The shallow layers of the network learn the basic physical laws of pollutant diffusion and have general applicability; the deep layers of the network learn the input-output mapping relationship of a specific scenario and need to be fine-tuned with new samples to adapt to changes in the scenario.

[0131] VII. Overall System Performance The system was deployed and operated in the target chemical industrial park for 6 months. The statistical results of the core performance indicators are shown in the table below: Performance indicators Before system deployment After system deployment Performance improvement Time of discovery of abnormal events Average 4 hours Average 15 minutes Shortened by 93.8% Accuracy of pollution source tracing 65% 92% Increased by 27 percentage points Emergency response time Average 3 hours Average 35 minutes Shortened by 80.6% False alarm rate 12% 3.50% Reduced by 70.8% Average daily data transmission volume per point 1.2GB 360MB Reduce by 70% Monthly decay rate of model inference accuracy 3% per month 0.8% per month Reduced by 73.3% During the system's deployment and operation, it successfully issued early warnings and completed emergency responses to five pollution incidents, effectively mitigating three risks of drinking water source pollution. This significantly improved the intelligence and precision of environmental supervision in the park, fully verifying the advanced nature and engineering practicality of the technical solution of this invention.

[0132] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. An online monitoring system based on AI analysis, comprising a three-layer architecture consisting of edge sensing devices, edge computing nodes, and cloud servers, characterized in that, include: The edge-side lightweight triggering module is used to collect multi-source, multi-modal environmental monitoring data in real time and run a micro-capacity anomaly detection model to determine whether the current data deviates from the historical baseline. Only when a deviation is detected will an upload command be triggered to upload the data packet to the edge computing node. The edge progressive inference module is used to receive and fuse multi-source multimodal data from multiple edge devices, generate fused feature vectors, and input them into a dynamic deep neural network containing multiple cascaded layers and corresponding early exit branches for progressive calculation. When the output confidence of any early exit branch meets a preset threshold, the calculation is terminated and the environmental state recognition result is output. The feature vectors of high-confidence abnormal events and the preliminary inference results are reported to the cloud. The cloud-based dual-engine collaborative decision-making module receives the abnormal event feature vectors reported by the side-by-side progressive reasoning module and inputs them into the knowledge graph tracing engine and the digital twin inference engine for parallel processing. The knowledge graph tracing engine generates an explanation path and credibility score from the abnormal phenomenon to the potential pollution source based on a pre-built environmental causal knowledge graph. The digital twin inference engine simulates the spatiotemporal diffusion process of pollutants based on a lightweight environmental dynamics model and generates trend predictions. The tracing results and trend predictions are fused to generate an intelligent decision package containing disposal suggestions. The adaptive event-driven communication module is used to adjust the data transmission strategy in real time according to the confidence level of abnormal events. Confirmed events are transmitted in real time with priority, suspected events are stored in edge nodes and uploaded on demand, and summaries of regular events are uploaded periodically. It also enables local caching and resume transmission mechanisms when the network is interrupted. The closed-loop optimization feedback module is used to compare the actual monitoring data sequence after emergency response with the theoretical decay curve generated by the digital twin inference engine, calculate the deviation and generate quantitative evaluation information, and feed it back to the cloud model training platform to update the core model parameters of each module. The modules form a closed-loop collaborative process of monitoring, reasoning, decision-making, and optimization. The cloud-based dual-engine parallel processing integrates the tracing and inference results, which is different from the separate application of knowledge graphs and digital twins.

2. The online monitoring system based on AI analysis according to claim 1, characterized in that, The micro-capacity anomaly detection model running in the edge-side lightweight triggering module is an unsupervised learning model based on single-class classification, which is deployed locally on the edge-side microcontroller after structured pruning optimization; the micro-capacity anomaly detection model uses corresponding feature extraction sub-modules for different types of data; The micro-volume anomaly detection model employs a dynamic adaptive threshold based on historical baseline statistical characteristics and real-time data signal-to-noise ratio adjustments to determine the degree of deviation between the current data distribution and the historical baseline distribution.

3. The online monitoring system based on AI analysis according to claim 1, characterized in that, The multi-source, multi-modal data fusion operation performed by the side-progressive inference module includes: using dynamic time warping algorithm and geohashing encoding to complete the spatiotemporal alignment of multi-source data; extracting time-frequency domain features, visual features, and semantic features from numerical, image, and text data respectively; and using an inter-modal multi-head cross-attention mechanism to adaptively weight and fuse the features of each modality. The inter-modal multi-head cross-attention mechanism assigns corresponding attention weights to each modality feature, and the attention weights are dynamically adjusted based on the real-time signal-to-noise ratio and historical inference confidence of the corresponding modality data.

4. The online monitoring system based on AI analysis according to claim 1, characterized in that, The dynamic deep neural network comprises multiple stacked convolutional layers. Each convolutional layer employs a multi-scale dilated convolution parallel structure, including a regular convolution parallel branch and a dilated convolution parallel branch. The convolutional kernels of the dilated convolution parallel branch are configured with a corresponding dilation rate, which increases with the depth of the layer group. The outputs of these branches are concatenated and input into the next layer group. The number of channels in each convolutional layer group increases progressively. Each layer group in the dynamic deep neural network is followed by an early exit branch consisting of a global average pooling layer and a fully connected classifier. This early exit branch employs an adaptive dropout regularization mechanism, which sets a corresponding neuron dropout rate. This dropout rate is dynamically adjusted based on the variance of the current layer group's output feature map.

5. The online monitoring system based on AI analysis according to claim 4, characterized in that, The side-progressive inference module also includes a confidence dynamic calculation unit, used to calculate a comprehensive confidence score based on the feature vector output by the current layer group. The formula is as follows: ， in, To calculate the overall confidence score, The original confidence level. Signal-to-noise ratio factor The rate of change of confidence level. The information entropy of the output feature vector probability distribution. , , As an adjustable hyperparameter, the comprehensive confidence score is used to classify events into three levels: confirmed, suspected, and routine, and serves as the basis for adjusting the data transmission strategy.

6. The online monitoring system based on AI analysis according to claim 1, characterized in that, The knowledge graph tracing engine includes an environmental causal knowledge graph database, a graph reasoning adaptation unit, an interpretable path generation unit, a path confidence quantification unit, and a decision suggestion matching unit. The environmental causal knowledge graph database stores five types of entities: monitoring indicators, abnormal events, pollution sources, environmental factors, and treatment measures, as well as three types of association edges: causal, spatial, and regulatory. The association edges are configured with updatable confidence weights. The graph reasoning adaptation unit maps abnormal event information into graph query statements. The interpretable path generation unit uses a depth-first graph search algorithm to traverse multiple reasoning paths from the exceeding indicator node to the potential pollution source node, and performs pruning optimization based on spatiotemporal and causal relationships. The path confidence quantification unit calculates the confidence score of each path based on the path topology length, intermediate node confidence, and association edge weights. The decision suggestion matching unit matches treatment measures according to the tracing results and outputs them to the digital twin inference engine as boundary conditions.

7. The online monitoring system based on AI analysis according to claim 1, characterized in that, The digital twin simulation engine includes a four-dimensional environmental digital twin, a lightweight environmental dynamics proxy model, a trend prediction output unit, and a visualization rendering unit. The four-dimensional environmental digital twin is constructed based on geographic information, hydrological and meteorological data, and pollution source distribution data, and can be dynamically updated. The lightweight environmental dynamics proxy model is deployed after model pruning and quantization, using knowledge graph source tracing results as core boundary conditions and real-time monitoring data as calibration parameters to quickly simulate the spatiotemporal diffusion process of pollutants on the digital twin. The trend prediction output unit generates the pollution diffusion trajectory, impact range, and theoretical decay curve, and synchronizes them to the closed-loop optimization feedback module; the visualization rendering unit presents the simulation results.

8. The online monitoring system based on AI analysis according to claim 1, characterized in that, The closed-loop optimization feedback module includes an actual data acquisition unit, a deviation calculation unit, a measure effectiveness evaluation unit, and a model feedback update unit; the actual data acquisition unit continuously collects monitoring data within a preset time period after emergency response; The deviation calculation unit compares the actual data with the theoretical attenuation curve derived from the digital twin to calculate the pollution reduction rate deviation; the measure effectiveness evaluation unit quantifies the contribution of each measure based on the concentration changes before and after the implementation of the treatment measures and generates an evaluation report; the model feedback update unit integrates the evaluation report, monitoring data, and treatment records into labeled training samples and feeds them back to the cloud model training platform to update the dynamic deep neural network, knowledge graph association weights, and lightweight proxy model parameters.

9. The online monitoring system based on AI analysis according to claim 7, characterized in that, The method for constructing the lightweight environmental dynamics surrogate model includes: using simulation results of a high-precision environmental dynamics numerical model covering a multi-dimensional parameter space as a training dataset; constructing a deep neural network regression model with pollution source parameters, meteorological and hydrological parameters, and spatiotemporal coordinates as inputs and pollutant concentration as output; training the deep neural network regression model using data augmentation and regularization strategies to learn the end-to-end nonlinear mapping from the input to the output field; and deploying the converged model as a lightweight surrogate model, whose inference speed is improved by at least two orders of magnitude compared to the original high-precision model, and whose inference accuracy meets the preset correlation requirements.

10. The online monitoring system based on AI analysis according to claim 9, characterized in that, The lightweight environmental dynamics proxy model supports an online incremental update mechanism: when the accumulated labeled training samples fed back by the closed-loop optimization feedback module exceed a preset threshold, incremental training is triggered, and transfer learning is used to fine-tune only the parameters of the last two fully connected layers of the model, while freezing the parameters of the remaining layers. After the updated model is preheated and verified to meet the accuracy requirements, it is switched to the online service model through a dual-model alternation mechanism; the samples are pushed in real time by the closed-loop optimization feedback module, so that the model can continuously adapt to environmental changes.

Images