CN-122021950-A - Online monitoring system based on AI analysis

CN122021950ACN 122021950 ACN122021950 ACN 122021950ACN-122021950-A

Abstract

The invention relates to the technical field of environmental monitoring, aims to solve the problems that an existing environmental monitoring system is insufficient in end-side-cloud cooperation, multi-source data fusion is difficult, decision one-sided interpretability is poor, a model cannot continuously and autonomously evolve, and data transmission redundancy and instantaneity are difficult to consider, and provides an online monitoring system based on AI analysis. The system adopts an end-side-Yun Sanceng architecture, is provided with an end-side light triggering module, a side progressive reasoning module, a cloud dual-engine collaborative decision module, a self-adaptive event driven communication module and a closed-loop optimization feedback module, and forms a full-flow collaborative closed loop through knowledge graph tracing and digital twin deduction dual-engine parallel decision, and forms monitoring, reasoning, decision and optimization. The method and the system remarkably improve the instantaneity, the traceability accuracy and the self-adaption capability of environment monitoring, and can be widely applied to the scenes such as water quality monitoring in chemical industry parks.

Inventors

FENG YUMING
WANG QICANG
ZHANG TING
Xiang Yingbo
WANG JUAN

Assignees

江苏方洋环境监测有限公司

Dates

Publication Date: 20260512
Application Date: 20260410

Claims (10)

1. The utility model provides an on-line monitoring system based on AI analysis, includes three-layer framework that terminal side perception device, edge computing node and high in the clouds server constitute, its characterized in that includes: The terminal side light triggering module is used for collecting multi-source multi-mode environment monitoring data in real time, running the micro-capacity anomaly detection model to judge whether the current data deviate from a historical base line, triggering an uploading instruction only when the deviation is detected, and uploading the data packet to an edge computing node; The side progressive reasoning module is used for receiving and fusing multi-source multi-mode data from a plurality of terminal side devices, generating a fused feature vector, inputting the fused feature vector into a dynamic deep neural network comprising a plurality of cascade layer groups and corresponding early exit branches to perform progressive calculation, terminating calculation and outputting an environment state recognition result when the output confidence level of any early exit branch meets a preset threshold value, and reporting the feature vector of a high-confidence abnormal event and the initial reasoning result to the cloud; The cloud dual-engine collaborative decision-making module is used for receiving the abnormal event feature vectors reported by the side progressive inference module, respectively inputting the abnormal event feature vectors to the knowledge graph tracing engine and the digital twin deduction engine for parallel processing, wherein the knowledge graph tracing engine generates an interpretation path and a credibility score from an abnormal phenomenon to a potential pollution source based on a pre-constructed environment causal knowledge graph, and the digital twin deduction engine simulates a space-time diffusion process of the pollutant and generates trend prediction based on a lightweight environment dynamics model, and fuses a tracing result and the trend prediction to generate an intelligent decision package containing treatment suggestions; The self-adaptive event-driven communication module is used for adjusting a data transmission strategy in real time according to the confidence level of an abnormal event, transmitting the confirmed event preferentially in real time, storing the suspected event in an edge node and uploading the suspected event as required, uploading a summary periodically for a conventional event, and starting a local caching and continuous transmission mechanism when the network is interrupted; The closed-loop optimization feedback module is used for comparing the actual monitoring data sequence after emergency treatment with a theoretical attenuation curve generated by the digital twin deduction engine, calculating deviation and generating quantitative evaluation information, and feeding the quantitative evaluation information back to the cloud model training platform to update the core model parameters of each module; All the modules form a full-flow collaborative closed loop of monitoring, reasoning, decision-making and optimizing, and the cloud dual-engine processes and fuses the tracing and deduction results in parallel, and is different from a mode of independently applying a knowledge graph and a digital twin.
2. The online monitoring system based on AI analysis according to claim 1, wherein the micro-capacity abnormality detection model operated in the end-side lightweight triggering module is an unsupervised learning model based on single class classification, is deployed on the local side of the end-side microcontroller after structured pruning optimization, adopts corresponding feature extraction sub-modules for different types of data, and adopts a dynamic adaptive threshold value adjusted based on historical baseline statistical characteristics and real-time data signal-to-noise ratio for judging the deviation degree of current data distribution and historical baseline distribution.
3. The AI analysis-based online monitoring system of claim 1, wherein the multi-source multi-modal data fusion operation performed by the side progressive inference module comprises performing space-time alignment of multi-source data using a dynamic time warping algorithm and a geographic hash code, extracting time-frequency domain features, visual features and semantic features from the digital, image and text data, respectively, and performing adaptive weighted fusion on each modal feature using an inter-modal multi-head cross-attention mechanism, wherein the inter-modal multi-head cross-attention mechanism assigns a corresponding attention weight to each modal feature, and wherein the attention weight is dynamically adjusted based on a real-time signal-to-noise ratio and a historical reasoning confidence of the corresponding modal data.
4. The online monitoring system based on AI analysis of claim 1, wherein the dynamic depth neural network comprises a plurality of convolution layer groups stacked layer by layer, the convolution layer groups adopt a multi-scale cavity convolution parallel structure and comprise conventional convolution parallel branches and cavity convolution parallel branches, a convolution kernel of each cavity convolution parallel branch is provided with a corresponding cavity rate, the cavity rate increases with the layer group depth, the branches are output and spliced and then input into a next layer group, the channel number of the convolution layer groups increases layer by layer, each layer group of the dynamic depth neural network is connected with an early exit branch formed by a global average pooling layer and a fully connected classifier, the early exit branch adopts an adaptive dropout regularization mechanism, and the adaptive dropout regularization mechanism is provided with a corresponding neuron rejection rate, and the rejection rate is dynamically adjusted based on the variance of an output characteristic diagram of the current layer group.
5. The AI analysis-based online monitoring system of claim 4, wherein the side-by-side progressive inference module further comprises a confidence coefficient dynamic calculation unit for calculating a comprehensive confidence coefficient score according to a feature vector output by a current layer group, and the formula is: , Wherein, the In order to integrate the confidence score(s), For the original confidence level, As a signal-to-noise ratio factor, In order for the rate of change of the confidence, For the output of the information entropy of the feature vector probability distribution, 、、 The comprehensive confidence score is used for classifying the event into three grades of confirmation, suspected and conventional as the basis for adjusting the data transmission strategy.
6. The AI analysis-based online monitoring system of claim 1, wherein the knowledge graph tracing engine comprises an environmental causal knowledge graph database, a graph reasoning adaptation unit, an interpretable path generation unit, a path confidence quantization unit and a decision suggestion matching unit, wherein the environmental causal knowledge graph database stores three types of related edges of monitoring indexes, abnormal events, pollution sources, environmental factors, treatment measures, five types of entities, causal, space and regulations, the related edges are configured with updatable confidence weights, the graph reasoning adaptation unit maps abnormal event information into graph query sentences, the interpretable path generation unit adopts a depth-first graph search algorithm to traverse a plurality of reasoning paths from overscale index nodes to potential pollution source nodes and performs pruning optimization based on space-time and causal correlations, the path confidence quantization unit calculates a confidence score of each path based on path topology length, intermediate node confidence and related edge weights, and the decision suggestion matching unit matches treatment measures according to a tracing result and outputs the treatment measures to a digital twin engine as boundary conditions.
7. The online monitoring system based on AI analysis of claim 1, wherein the digital twinning deduction engine comprises a four-dimensional environment digital twinning body, a lightweight environment dynamic proxy model, a trend prediction output unit and a visual rendering unit, wherein the four-dimensional environment digital twinning body is constructed and dynamically updated based on geographic information, hydrological weather and pollution source distribution data, the lightweight environment dynamic proxy model is deployed after model pruning and quantization, a knowledge graph tracing result is taken as a core boundary condition, real-time monitoring data is taken as a calibration parameter, a pollutant space-time diffusion process is quickly deduced on the digital twinning body, the trend prediction output unit generates a pollution diffusion track, an influence range and a theoretical attenuation curve and is synchronized to a closed-loop optimization feedback module, and the visual rendering unit presents deduction results.
8. The AI analysis-based online monitoring system of claim 1, wherein the closed-loop optimization feedback module comprises an actual data acquisition unit, a deviation calculation unit, a measure effectiveness evaluation unit and a model feedback updating unit, wherein the actual data acquisition unit continuously acquires monitoring data within a preset time period after emergency treatment, the deviation calculation unit compares the actual data with a theoretical attenuation curve deduced by digital twinning to calculate pollution reduction rate deviation, the measure effectiveness evaluation unit quantifies contribution degrees of each measure based on concentration changes before and after treatment measure execution to generate an evaluation report, and the model feedback updating unit integrates the evaluation report, the monitoring data and treatment records into a training sample with labels and feeds back the training sample to a cloud model training platform for updating dynamic deep neural networks, knowledge map association weights and lightweight agent model parameters.
9. The AI analysis-based online monitoring system of claim 7, wherein the method for constructing the lightweight environmental dynamics proxy model comprises using simulation results of a high-precision environmental dynamics numerical model in a multi-dimensional parameter space as a training data set, constructing a deep neural network regression model with pollution source parameters, weather hydrologic parameters, space-time coordinates as input and pollutant concentration as output, training the deep neural network regression model by adopting a data enhancement and regularization strategy, learning an end-to-end nonlinear mapping from input to output field, deploying the model after training convergence as a lightweight proxy model, improving the deduction speed by at least two orders of magnitude compared with an original high-precision model, and enabling deduction precision to meet preset correlation requirements.
10. The online monitoring system based on AI analysis of claim 9, wherein the lightweight environmental dynamics agent model supports an online incremental update mechanism, wherein when accumulation of labeled training samples fed back by the closed-loop optimization feedback module exceeds a preset number threshold, incremental training is triggered, transfer learning is adopted to finely adjust parameters of the last two layers of full-connection layers of the model, parameters of the other layers are frozen, the updated model is switched into an online service model through a dual-model alternating mechanism after preheating verification meets precision requirements, and the samples are pushed by the closed-loop optimization feedback module in real time, so that the model continuously adapts to environmental changes.

Description

Online monitoring system based on AI analysis Technical Field The invention relates to the technical field of environmental monitoring, in particular to an online monitoring system based on AI analysis. Background In the prior art of an on-line monitoring system based on AI analysis, a complete technical system is mainly built around an ecological environment monitoring scene, a monitoring analysis model is generally built by adopting a machine learning algorithm, and real-time acquisition analysis and abnormal state identification of environmental parameters are supported. In the prior art, a support vector machine algorithm is used as a core, and model training and parameter optimization are completed by combining environment monitoring data acquired by a multi-source sensor, so that conventional monitoring and exceeding early warning of various environment indexes are realized. Meanwhile, a multidimensional environment data acquisition system is built in a matched manner in the prior art, real-time acquisition and standardized transmission of relevant parameters of the ecological environment are completed through a front-end distributed sensor network, model reasoning operation and analysis result output are completed by combining a rear-end centralized computing platform, a basic monitoring and early warning whole-flow closed loop is formed, and the relevant technical scheme is widely applied to multiple ecological environment monitoring scenes such as river basin management and control and park supervision. Aiming at the prior related technology, the inventor considers that the prior art has poor adaptability to dynamic time sequence data of the ecological environment, has insufficient fitting precision to nonlinear and non-stationary ecological monitoring data, has high demand on model calculation power, is difficult to adapt to an edge end real-time monitoring scene, has the problem of model precision attenuation in long-term operation, and has insufficient precision degree of ecological risk management and control. Disclosure of Invention The technical problem to be solved by the invention is that the existing environment monitoring system has the core problems of insufficient end-side-cloud cooperation and difficult multi-source data fusion, namely, each level independently operates, the abnormal response is lagged due to data island, the decision is one-sided, and the model cannot continuously evolve, so that an online monitoring system based on AI analysis is provided. In order to achieve the above purpose, the application adopts the following technical scheme that the on-line monitoring system based on AI analysis comprises a three-layer framework consisting of end side sensing equipment, edge computing nodes and a cloud server, and comprises the following components: The system comprises an end-side light triggering module, an end-side sensing device, a monitoring site and a micro-capacity anomaly detection module, wherein the end-side light triggering module is used for acquiring multi-source multi-mode environment monitoring data in real time, running the micro-capacity anomaly detection module to judge whether current data deviate from a historical base line, triggering an uploading instruction only when deviation is detected, and uploading a data packet to an edge computing node; The system comprises an edge side progressive reasoning module, an edge computing node, a neural network output layer, a cloud end, a network node and a network node, wherein the edge side progressive reasoning module is used for receiving and fusing multi-source multi-mode data from a plurality of end side devices, generating a fused feature vector, inputting the fused feature vector into the dynamic deep neural network comprising a plurality of cascade layer groups and corresponding early exit branches for progressive computation, terminating computation and outputting an environment state recognition result when the output confidence level of any early exit branch meets a preset threshold, and reporting the feature vector of a high-confidence abnormal event and the initial reasoning result to the cloud end; The cloud dual-engine collaborative decision-making module is used for receiving the abnormal event feature vectors reported by the side progressive inference module, respectively inputting the abnormal event feature vectors to the knowledge graph tracing engine and the digital twin deduction engine for parallel processing, wherein the knowledge graph tracing engine generates an interpretation path from an abnormal phenomenon to a potential pollution source and a reliability score based on a pre-constructed environment causal knowledge graph, the digital twin deduction engine simulates a space-time diffusion process of the pollutant and generates trend prediction based on a lightweight environment dynamics model, and fuses the tracing result and the trend prediction to generate an intelligent decision package