Multi-mode laser radar networked atmospheric parameter collaborative monitoring system and monitoring method

Abstract

The invention discloses a multi-mode laser radar networked atmospheric parameter collaborative monitoring system and monitoring method, and relates to the technical field of atmospheric monitoring, and the method comprises the steps: deploying multi-mode laser radar nodes, synchronously collecting multi-source atmospheric parameters for preprocessing, and obtaining a fusion feature matrix; transmitting the fusion feature matrix to an edge computing node through dynamic networking, and aggregating the edge node through federal learning to realize cloud fusion; the method comprises the following steps: constructing a space-time dynamic graph based on laser radar nodes and a space-time dependency relationship between the nodes, decomposing space-time dynamic correlation into a stable item and a non-stable item through space-time decoupling, and updating space-time dynamic graph data by adopting an alternating optimization strategy through online learning; and constructing a prediction model, taking the real-time feature flow obtained by cloud fusion as input, detecting an abnormal event by predicting atmospheric parameters, and performing pollution traceability. According to the invention, high-precision and real-time monitoring of atmospheric parameters is realized through multi-mode cooperative sensing, intelligent fusion analysis and dynamic optimization decision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of atmospheric monitoring, more particularly, to a multi-modal laser radar networked atmospheric parameter collaborative monitoring system and a monitoring method. BACKGROUND

[0002] With the increasing problems of atmospheric environmental pollution, meteorological disasters and climate change, the demand for real-time and accurate monitoring of atmospheric parameters is becoming increasingly urgent. Traditional atmospheric monitoring methods (such as single-site measurement, satellite remote sensing or isolated sensor network) have problems such as insufficient spatial and temporal resolution, poor data collaboration, limited multi-parameter fusion capability, etc., and are difficult to meet the high-precision, multi-dimensional dynamic monitoring needs of complex atmospheric environment. Laser radar technology has become an important means of atmospheric parameter monitoring due to its high spatial and temporal resolution, high sensitivity and active detection capability, and can realize the inversion of key parameters such as aerosol, greenhouse gas, wind field, temperature and humidity. However, traditional single-mode laser radar systems usually only detect a single or a few atmospheric parameters, and are limited to fixed deployment or independent operation mode, making it difficult to achieve large-scale, networked collaborative monitoring. In addition, the data heterogeneity, calibration differences and communication protocol incompatibility between different laser radar systems further restrict the fusion and application efficiency of multi-source data.

[0003] In recent years, the development of multi-modal laser radar technology has provided a new way to break through the above bottlenecks. By integrating multiple laser radar detection modes (such as Mie scattering, differential absorption, Doppler wind measurement, etc.), multi-dimensional synchronous perception of atmospheric parameters can be achieved. At the same time, with the help of Internet of Things, edge computing and 5G / 6G communication technology, a networked collaborative monitoring system can be built, which can realize intelligent networking, real-time data interaction and collaborative processing of distributed nodes, thereby improving the monitoring coverage and data reliability. However, the existing technology still has challenges in multi-modal laser radar data fusion algorithm, networked collaborative optimization, dynamic calibration, etc., such as insufficient multi-source data fusion, low networked collaborative efficiency, and lack of dynamic calibration. Therefore, it is urgent to study a multi-modal laser radar networked collaborative monitoring method to build an efficient and accurate atmospheric parameter collaborative monitoring system through multi-source data fusion, intelligent networking optimization and dynamic calibration technology, and provide technical support for environmental governance and weather forecasting. SUMMARY

[0004] To solve the above technical problems, the present application proposes a multi-modal laser radar networked atmospheric parameter collaborative monitoring system and a monitoring method, which builds an efficient and accurate atmospheric parameter collaborative monitoring system through multi-source data fusion, intelligent networking optimization and dynamic calibration technology.

[0005] The first aspect of the application provides a multi-modal laser radar networked atmospheric parameter cooperative monitoring system, comprising a multi-modal laser radar module, a multi-source data fusion and preprocessing module, an intelligent networking optimization module, a spatio-temporal dynamic graph modeling module, a cooperative decision and visualization module;

[0006] The multi-modal laser radar module synchronously collects multi-source atmospheric parameters by distributedly deploying multi-modal laser radar nodes, and the multi-modal radar comprises Mie scattering radar, differential absorption radar and Doppler wind measurement radar;

[0007] The multi-source data fusion and preprocessing module calibrates, aligns and feature fuses the multi-source atmospheric parameters to generate a fusion feature matrix;

[0008] The intelligent networking optimization module dynamically networks by software-defined network, transmits the fusion feature matrix to the edge computing node based on the reinforcement learning route selection transmission path, and realizes cloud fusion by aggregating the data uploaded by the edge node through federated learning;

[0009] The spatio-temporal dynamic graph modeling module constructs a spatio-temporal dynamic graph based on the spatio-temporal dependence relationship between the laser radar nodes and the nodes in the cloud, decomposes the spatio-temporal dynamic correlation into stable items and non-stationary items through spatio-temporal decoupling, and updates the spatio-temporal dynamic graph data through online learning using an alternating optimization strategy;

[0010] A prediction model is constructed based on the spatio-temporal dynamic graph data, real-time feature flow obtained by cloud fusion is taken as input, and atmospheric parameters after a preset time are predicted;

[0011] The cooperative decision and visualization module detects abnormal events and performs pollution tracing according to the predicted atmospheric parameters, and dynamically visualizes and displays real-time data, abnormal events and pollution tracing results.

[0012] In the scheme, the multi-source data fusion and preprocessing module specifically comprises:

[0013] Multi-source atmospheric parameters synchronously collected by the multi-modal laser radar are obtained, the multi-source atmospheric parameters include aerosol data, greenhouse gas data and wind field data, error correction is performed on the multi-source atmospheric parameters using ground meteorological station data, data comparison is performed on the observation overlapping areas of different modal laser radars, and dynamic cross-calibration is performed;

[0014] The multi-source atmospheric parameters are spatio-temporally aligned, and outliers are removed through outlier processing, and the preprocessed multi-source atmospheric parameters are obtained by using adjacent node data for filling, and physical features, statistical features and spatio-temporal features are extracted based on the preprocessed multi-source atmospheric parameters;

[0015] The physical characteristics, statistical characteristics and space-time characteristics of the multi-source atmospheric parameters are organized as a multi-dimensional tensor according to a space-time network, specific feature coding is performed on different atmospheric parameters, a CNN network is used to process the spatial distribution image of aerosol data, an LSTM is used to process the time sequence of gas data, and a 3D CNN network is used to process the wind speed field;

[0016] An attention mechanism is introduced to dynamically allocate the weights of the features of each modality, attention fusion features are obtained through weighted fusion, feature splicing features are obtained by splicing the features of each modality, and tensor fusion features are obtained by modeling the interaction between the modalities through outer product;

[0017] A fusion feature matrix is generated based on the attention fusion features, the feature splicing features and the tensor fusion features.

[0018] In the scheme, the intelligent networking optimization module specifically comprises:

[0019] A software-defined network controller is deployed to centrally manage the topology network and traffic state of the multi-modality lidar node group, the lidar node is used as a software-defined network switch, and data is forwarded according to the instructions of the software-defined network controller;

[0020] Virtual network slices are divided for different data types to guarantee the transmission bandwidth requirement of sudden pollution events, and the software-defined network controller updates the topology network by detecting the link state between nodes in real time through the LLDP protocol;

[0021] A network state vector is obtained according to the link delay, bandwidth utilization, node load and data priority of the topology network, the next hop node of data propagation is used as an optional action, a forward reward is constructed based on low delay and high throughput, and a penalty term is constructed according to the packet loss rate exceeding a preset threshold;

[0022] The forward reward and the penalty term are used to train a deep Q network algorithm, and the software-defined network controller based on the trained deep Q network selects an action according to the real-time network state vector to adaptively select an optimal path to transmit the fusion feature matrix to an edge node;

[0023] After receiving the fusion feature matrix, the edge node performs anomaly detection using a lightweight CNN-LSTM model, marks pollution anomaly data, and performs feature dimension reduction using principal component analysis, temporarily stores the pollution anomaly data in the edge node for on-demand uploading when there is no pollution anomaly data, and immediately triggers transmission when pollution anomaly data appears;

[0024] Each edge node updates the model parameters with local data, the cloud receives all edge node model parameters using the FedAvg algorithm, and performs weighted average according to the data volume, the cloud distributes the aggregated global model to the edge node, the next round of training is based on the new model to continue optimization, and the cloud uses the aggregated global model to predict the pollution anomaly of the whole area data, and the pollution anomaly data of the edge node and the pollution anomaly prediction data of the global model are fused to obtain the spatio-temporal grid data.

[0025] In this scheme, in the spatio-temporal dynamic graph modeling module, a spatio-temporal dynamic graph is constructed based on the laser radar nodes and the spatio-temporal dependence relationship between the nodes in the cloud, and the spatio-temporal dynamic correlation is decomposed into stable items and non-stationary items by spatio-temporal decoupling, specifically:

[0026] According to the laser radar nodes as nodes, the spatial edges and the time edges are constructed according to the geographical distance and the historical data trend between the nodes, the spatio-temporal dynamic graph is constructed based on the node and edge structure, and the graph structure is updated using a sliding time window to represent the real-time change of the atmospheric parameter;

[0027] The preprocessed spatio-temporal dynamic graph data is obtained, a static graph adjacency matrix is constructed to calculate a graph Laplacian matrix, a graph Fourier transform is performed on a node feature matrix based on the graph Laplacian matrix, and a low-frequency component is reserved to reconstruct a signal;

[0028] Each time slice signal is transformed using the low-frequency component reconstructed signal, a reconstructed stable item is obtained by low-frequency reconstruction, and a Savitzky-Golay filter is used to eliminate noise fluctuations and retain seasonal trends;

[0029] According to the spatio-temporal dynamic graph data, the original observation value and the stable item residual are obtained as the initial node state, the edge weight is constructed according to the spatial weight, the wind field correction and the concentration driving, the dynamic adjacency matrix is constructed, the dynamic adjacency matrix is represented and learned by a dynamic graph neural network, and spatial propagation and time evolution are performed using spatio-temporal convolution.

[0030] Based on the EMA filter, the exponential average in a preset time in the past is obtained, the current time prediction value is obtained based on the dynamic graph neural network, and the mutation feature is identified according to the exponential average and the current time prediction value to obtain the non-stationary item.

[0031] In this scheme, the spatio-temporal dynamic graph data is updated by using an alternating optimization strategy for online learning, specifically:

[0032] The update of the spatiotemporal dynamic data is decomposed into two stages of stable item optimization and non-stable item optimization by using expectation maximization, first, fixing the stable item, optimizing the non-stationary item, locking the current stable item parameter, calculating the residual signal of the observation value and the stable item, inputting the residual signal and real-time multi-source atmospheric parameter into the dynamic graph neural network, and performing gradient descent through the adaptive dynamic learning rate to optimize the attention weight and the gating parameter, and obtaining the non-stationary item update result;

[0033] Secondly, fix the non-stationary item, optimize the stable item, fix the current non-stationary item detection result, aggregate the denoised data in the past preset period, perform exponential weighted average on the graph Fourier coefficient, retain the first k main vibration modes, and adjust the mode amplitude through the adaptive dynamic learning rate, without changing the vibration form, to obtain the stable item update result.

[0034] In the scheme, a prediction model is constructed based on the spatiotemporal dynamic graph data, and real-time feature flow obtained by cloud fusion is taken as input to predict atmospheric parameters after a preset time, specifically:

[0035] The prediction model adopts a double-path architecture to process the stable item and the non-stationary item respectively, receives the stable item data of the past preset time through the time convolution network in the stable item prediction path and organizes it into a time sequence matrix according to the node, extracts features through causal dilated convolution, and obtains a trend basis vector for the future preset time;

[0036] The spatial aggregation is performed through the graph attention network to construct a graph structure of the geographical distance between the laser radar nodes, the multi-head attention is introduced for node-level feature interaction to obtain a spatial correlation feature matrix, and the spatial correlation feature matrix and the trend basis vector are generated into stable item features;

[0037] In the non-stable item prediction path, a dynamic graph is constructed according to the real-time edge weight and the current non-stable item value, a gated graph neural network is used for dynamic propagation, message passing iteration is performed through the gating mechanism, and layer normalization operation is performed after each iteration;

[0038] A grid centered on the target node is constructed for relative position coding, the self-attention mechanism is introduced to highlight the gradient mutation area, and an abnormal propagation feature map is output by the gated graph neural network as the non-stable item feature;

[0039] The cross-attention mechanism is used for feature fusion of the stable item feature and the non-stable item feature, and the original stable item trend basis vector is reserved through the residual connection to generate a fusion feature, which is input into a multi-step prediction decoder constructed by the multilayer perception mechanism to output the predicted value of the atmospheric parameter at the future preset time.

[0040] In the scheme, the collaborative decision and visualization module specifically comprises:

[0041] The predicted atmospheric parameters are subjected to data standardization processing, potential feature extraction of the atmospheric parameter prediction values after data standardization processing is performed by constructing an encoder through 3 layers of 3D-CNN spatio-temporal convolution layers, and group normalization processing is introduced;

[0042] A decoder is constructed using 3 layers of transposed spatio-temporal convolution layers, reconstruction errors are obtained, a baseline threshold is obtained based on historical normal data, the baseline threshold is corrected according to temperature and humidity data, the reconstruction errors are compared with the corrected baseline threshold, and abnormal events are classified according to the comparison results;

[0043] Based on the abnormal events, an observation abnormal propagation path in a monitoring grid is obtained, candidate sources are obtained within a preset range of the abnormal events, candidate source simulation paths are generated based on Lagrangian particle tracking of wind field data, Frechet distances between the observation abnormal propagation path and the candidate source simulation paths are calculated, and the candidate source corresponding to the minimum Frechet distance is selected as a pollution source tracing result and is visually displayed.

[0044] The second aspect of the present application proposes a multi-modal laser radar networked atmospheric parameter collaborative monitoring method, characterized in that it is applied to a multi-modal laser radar networked atmospheric parameter collaborative monitoring system, and includes the following steps:

[0045] Multi-modal laser radar nodes are deployed, multi-source atmospheric parameters are synchronously collected, the multi-source atmospheric parameters are subjected to data preprocessing, and a fusion feature matrix is generated by performing multi-source feature fusion on the preprocessed multi-source atmospheric parameters;

[0046] Software-defined networks are used for dynamic networking, transmission paths are selected based on reinforcement learning routing to transmit the fusion feature matrix to edge computing nodes, and cloud fusion is realized by aggregating the data uploaded by the edge nodes through federated learning;

[0047] A spatio-temporal dynamic graph is constructed based on the spatio-temporal dependence relationship between the laser radar nodes and the nodes in the cloud, the spatio-temporal dynamic correlation is decomposed into a stable term and a non-stationary term through spatio-temporal decoupling, and the spatio-temporal dynamic graph data is updated through online learning using an alternating optimization strategy;

[0048] A prediction model is constructed based on the spatio-temporal dynamic graph data, real-time feature streams obtained through cloud fusion are used as input, atmospheric parameters after a preset time are predicted, abnormal events are detected according to the predicted atmospheric parameters, and pollution source tracing is performed.

[0049] Compared with the prior art, the present application has the following advantages:

[0050] The multi-modal laser radar networked atmospheric parameter collaborative monitoring method of the application realizes high-precision and real-time monitoring of atmospheric environmental parameters by deeply fusing multi-source observation data and intelligent analysis technology. First, the system synchronously collects multi-dimensional parameters such as aerosol, greenhouse gas and three-dimensional wind field by distributedly deploying Mie scattering radar, differential absorption radar and Doppler radar nodes, and constructs an observation network with high temporal and spatial resolution. Through deep feature extraction and fusion technology, efficient collaborative analysis of multi-modal data is realized. Further, the introduction of spatio-temporal dynamic graph modeling and online learning strategy enhances the adaptability of the system to complex atmospheric dynamics. The spatio-temporal decoupling module decomposes the correlation into stable and non-stationary terms, which are optimized by low-frequency and high-frequency update strategies respectively, so that the response speed of the model in sudden pollution events is improved to seconds, and the long-term trend prediction error is greatly reduced. The intelligent networking optimization and collaborative decision-making module realizes efficient use of resources, improves pollution traceability accuracy, and realizes real-time visual early warning, providing a scientific basis for environmental governance and public service. BRIEF DESCRIPTION OF DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments or examples of the present application, the drawings needed to be used in the embodiments or examples will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0052] Figure 1 A block diagram of a multi-modal laser radar networked atmospheric parameter collaborative monitoring system is shown.

[0053] Figure 2 A flowchart showing the decomposition of spatio-temporal dependence into stable and non-stationary terms by spatio-temporal decoupling is shown.

[0054] Figure 3 A flowchart showing the construction of a prediction model to predict atmospheric parameters is shown.

[0055] Figure 4 A flowchart showing the multi-modal laser radar networked atmospheric parameter collaborative monitoring method is shown. DETAILED DESCRIPTION

[0056] In order to more clearly illustrate the technical solutions in the embodiments or examples of the present application, the drawings needed to be used in the embodiments or examples will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0057] In the following description, a lot of specific details are set forth in order to provide a thorough understanding of the present application, however, the present application can be practiced without resorting to the other ways different from those described here, therefore, the scope of protection of the present application is not limited by the specific embodiments disclosed below.

[0058] Figure 1 A block diagram of a multi-modal laser radar networked atmospheric parameter collaborative monitoring system is shown.

[0059] As Figure 1 shown, the first embodiment of the present application provides a multi-modal laser radar networked atmospheric parameter collaborative monitoring system 1, comprising: a multi-modal laser radar module 101, a multi-source data fusion and preprocessing module 102, an intelligent networking optimization module 103, a spatio-temporal dynamic graph modeling module 104, a collaborative decision and visualization module 105;

[0060] The multi-modal laser radar module 101 synchronously collects multi-source atmospheric parameters by distributed deployment of multi-modal laser radar nodes, the multi-modal radar including Mie scattering radar, differential absorption radar and Doppler wind radar;

[0061] The multi-source data fusion and preprocessing module 102 calibrates, aligns and feature fuses the multi-source atmospheric parameters to generate a fusion feature matrix;

[0062] The intelligent networking optimization module 103 adopts software-defined network for dynamic networking, transmits the fusion feature matrix to the edge computing node based on reinforcement learning route selection transmission path, and realizes cloud fusion by federated learning to aggregate data uploaded by edge nodes;

[0063] The spatio-temporal dynamic graph modeling module 104 constructs a spatio-temporal dynamic graph based on the spatio-temporal dependence relationship between laser radar nodes and nodes in the cloud, decomposes the spatio-temporal dynamic correlation into stable items and non-stationary items through spatio-temporal decoupling, and updates the spatio-temporal dynamic graph data through online learning using an alternating optimization strategy;

[0064] Based on the spatio-temporal dynamic graph data, a prediction model is constructed, the real-time feature stream obtained by cloud fusion is taken as input, and the atmospheric parameters after a preset time are predicted;

[0065] The collaborative decision and visualization module 105 detects abnormal events and conducts pollution tracing according to the predicted atmospheric parameters, and dynamically visualizes and displays real-time data, abnormal events and pollution tracing results.

[0066] It should be noted that the laser radar nodes are deployed in a spatial grid in the target area (such as a city or an industrial area) to ensure coverage of monitoring requirements, and the nodes integrate multi-modal laser radars including Mie scattering radars, differential absorption radars, and Doppler radars. GPS or PTP is used to synchronize the clocks of the nodes to ensure consistency of data timestamps. The Mie scattering radar transmits 532nm / 1064nm laser and receives backscattering signals of aerosols to output extinction coefficients and cloud heights. The differential absorption radar transmits 1572nm (CO2) and 1650nm (CH4) laser to inversely calculate gas concentrations through absorption differences. The Doppler radar measures laser frequency shifts to calculate three-dimensional wind fields.

[0067] The multi-source data fusion and preprocessing module acquires multi-source atmospheric parameters synchronously collected by multi-modal laser radars, the multi-source atmospheric parameters including aerosol data, greenhouse gas data, and wind field data, corrects errors of the multi-source atmospheric parameters by using ground meteorological station data, and corrects baseline drift of the laser radars by Kalman filtering. Data comparison is performed on overlapping areas of different modal laser radars for dynamic cross-calibration. The multi-source atmospheric parameters are spatiotemporally aligned, data is interpolated to a unified latitude and longitude grid, and Kriging interpolation is used to fill in the gaps. Outliers are removed through outlier processing, and adjacent node data is used for filling to obtain preprocessed multi-source atmospheric parameters. Physical features, statistical features, and spatiotemporal features are extracted based on the preprocessed multi-source atmospheric parameters. The physical features are directly extracted from each modal atmospheric parameter, the statistical features are mean, variance, and gradient calculated in a sliding window, and the spatiotemporal features are periodic patterns extracted by Fourier transform.

[0068] The physical features, statistical features, and spatiotemporal features of the multi-source atmospheric parameters are organized as multi-dimensional tensors according to spatiotemporal networks, specific feature coding is performed on different atmospheric parameters, a CNN network is used to process spatial distribution images of aerosol data to obtain spatial feature maps, an LSTM is used to process time series of gas data to obtain time sequence feature vectors, and a 3D CNN network is used to process wind speed fields to obtain spatiotemporal joint features. An attention mechanism is introduced to dynamically allocate weights of features of each modal, attention fusion features are obtained through weighted fusion, features of each modal are spliced to obtain feature splicing features, and modal interaction is modeled through outer product to obtain tensor fusion features. A fusion feature matrix is generated based on the attention fusion features, the feature splicing features, and the tensor fusion features. This scheme avoids the limitations of manual feature design through automatic feature learning, adapts to complex atmospheric dynamics, and explicitly captures gas-particle-wind interactions through attention or tensor fusion to realize cross-modal correlation modeling. High information density input features are provided for subsequent spatiotemporal prediction, and monitoring accuracy is significantly improved.

[0069] It should be noted that the intelligent networking optimization module deploys a software defined network controller to centrally manage the topology network and traffic state of the multi-modal laser radar node group, takes the laser radar node as a software defined network switch, and forwards data according to the software defined network controller instruction; virtual network slicing is divided for different data types to guarantee the transmission bandwidth demand of the burst pollution event, the software defined network controller updates the topology network by real-time detection of the link state between nodes through the LLDP protocol; the network state vector is obtained according to the link delay, bandwidth utilization, node load and data priority of the topology network, the next hop node of data propagation is taken as an optional action, a forward reward is constructed based on low delay and high throughput, and a penalty term is constructed according to the packet loss rate exceeding a preset threshold; the forward reward and the penalty term are used to train a deep Q network algorithm, the software defined network controller selects an action based on the trained deep Q network according to the real-time network state vector, and adaptively selects an optimal path to transmit the fusion feature matrix to an edge node; the lightweight calculation of the feature matrix is completed at the edge side, and the cloud load is reduced. After the edge node receives the fusion feature matrix, an abnormality detection is performed using a lightweight CNN-LSTM model, pollution abnormal data is marked, and feature dimension reduction is performed using principal component analysis, in the absence of pollution abnormal data, the data is temporarily stored in the edge node and uploaded on demand, and in the presence of pollution abnormal data, transmission is immediately triggered.

[0070] Under the premise of data privacy protection, the edge node abnormal data is cooperatively modeled and fused with the cloud through federated learning. Each edge node updates the model parameters using local data, the cloud receives the model parameters of all edge nodes using the FedAvg algorithm, and performs weighted averaging according to the data volume, the cloud distributes the aggregated global model to the edge nodes, the next round of training is based on the new model for optimization, and the cloud uses the aggregated global model to predict pollution abnormal data for the entire region, the pollution abnormal data of the edge nodes and the pollution abnormal prediction data of the global model are fused to obtain spatiotemporal gridded data, and the differential privacy (DP) technology is introduced to add noise when uploading parameters to prevent data leakage. Through the deep cooperation of SDN, reinforcement learning and federated learning, efficient, secure and intelligent processing of multi-modal monitoring data is realized.

[0071] Figure 2 A flowchart showing the decomposition of spatiotemporal dependence into stable and non-stationary terms through spatiotemporal decoupling is shown.

[0072] According to the embodiment of the present application, in the spatiotemporal dynamic graph modeling module, a spatiotemporal dynamic graph is constructed at the cloud based on the spatiotemporal dependence relationship between laser radar nodes and nodes, and the spatiotemporal dynamic dependence is decomposed into stable and non-stationary terms through spatiotemporal decoupling, specifically:

[0073] S202, according to the laser radar node as a node, according to the geographical distance between nodes and historical data trends respectively construct space edges and time edges, construct a space-time dynamic graph based on node and edge structure, update the graph structure with a sliding time window, and represent the real-time changes of the atmospheric parameters;

[0074] S204, obtaining the preprocessed space-time dynamic graph data, constructing a static graph adjacency matrix to calculate a graph Laplacian matrix, performing graph Fourier transform on a node feature matrix based on the graph Laplacian matrix, retaining a preset proportion of low-frequency components to reconstruct a signal;

[0075] S206, using the low-frequency component reconstructed signal to transform each time slice signal, obtaining a reconstructed stable term through low-frequency reconstruction, and using a Savitzky-Golay filter to eliminate noise fluctuations and retain seasonal trends;

[0076] S208, obtaining the original observation value and the stable term residual as the initial node state according to the space-time dynamic graph data, constructing edge weight according to the space weight, wind field correction and concentration driving, constructing a dynamic adjacency matrix, and performing representation learning on the dynamic adjacency matrix through a dynamic graph neural network, and using space-time convolution for spatial propagation and time evolution;

[0077] S210, obtaining the exponential average in a preset time based on EMA filtering, obtaining the current time prediction value through a dynamic graph neural network, and obtaining the non-stationary term according to the exponential average and the current time prediction value.

[0078] It should be noted that the graph Fourier transform method is used for stable term extraction, and the graph Laplacian matrix is calculated according to the space-time dynamic graph, the degree matrix D (diagonal element is the number of node connections), the adjacency matrix W (space weight + static meteorological weight), and the Laplacian matrix is represented as L=D-W. The Fourier basis U is obtained by eigenvalue decomposition of L, wherein the low-frequency component corresponds to the large-scale spatial mode, and the high-frequency component reflects the local details. Projecting to the time series signal, transforming each time slice signal, retaining the low-frequency component to reconstruct the stable term, and applying sliding average to the low-frequency coefficient to remove the fluctuation components above the daily cycle. Preferably, the graph Fourier transform method is used for stable term modeling, and the physical mode can be analyzed as: the first main mode is the regional background concentration field, the second main mode is the dominant wind field transmission channel, and the third main mode is the urban heat island effect characteristic.

[0079] The non-stationary term is extracted based on a dynamic graph neural network, and a dynamic adjacency matrix is constructed based on a basic space weight, real-time meteorological correction and pollution transmission coefficient, the basic space weight is obtained based on a distance decay function of a Gaussian kernel, the real-time meteorological correction is obtained by using a cosine similarity of wind direction multiplied by a normalized value of wind speed, and the pollution transmission coefficient is adjusted by a concentration gradient direction. The original observation value and the stable term residual are obtained as the initial node state, the dynamic graph is constructed according to the real-time edge weight calculation and the node state initialization, in the spatial propagation of the spatio-temporal convolution operation, the first layer is a pure spatial relationship aggregation, the second layer is a spatial relationship combined with feature attention aggregation, the expansion convolution is used to capture multi-scale time sequence mode in time evolution, and the noise fluctuation is filtered through the gating mechanism, and the non-stationary term is calculated through the mutation feature output.

[0080] It should be noted that the expectation maximization is used to decompose the update of the spatio-temporal dynamic data into two stages of stable term optimization and non-stable term optimization, through the separation update strategy, the stability of the long-term regularity can be maintained, and the short-term sudden event can be quickly responded.

[0081] Firstly, the stable term is fixed, and the non-stationary term is optimized, the target is to quickly capture the sudden pollution event and short-term meteorological change. The current stable term parameter (frozen Fourier basis coefficient) is locked, the residual signal of the observation value and the stable term is calculated, the residual signal and the real-time multi-source atmospheric parameter are input into the dynamic graph neural network, the gradient descent is performed through the adaptive dynamic learning rate, for example, the fixed learning rate is 1e-4, the learning rate is preheated to make the learning rate linearly grow, the attention weight and the gating parameter are optimized, and the non-stationary term update result is obtained; secondly, the non-stationary term is fixed, and the stable term is optimized, the target is to smooth the long-term trend and periodic mode. The current non-stationary term detection result is fixed, the denoising data (observation value-non-stationary term) in the past preset period is aggregated, the graph Fourier coefficient is exponentially weighted and averaged, the first k main vibration modes are retained, the mode amplitude is adjusted through the adaptive dynamic learning rate, the vibration form is not changed, for example, the basic learning rate is set to 1e-3, the temporary learning rate is increased to 5e-3 when the abnormality is detected, and the L2 regularization is added to prevent overfitting, and the stable term update result is obtained.

[0082] Preferably, a typical scene sample library is maintained, pollution event samples (non-stationary term features are retained) and pollution event samples (mode migration processes are recorded) are retrieved and stored in the typical scene sample library, a preset playback period is set, the pollution event samples and the pollution event samples are played back according to the preset playback period, important modes are prevented from being forgotten by the model, and the pollution event samples and the pollution event samples are updated in real time, the compatibility of new and old nodes is maintained through periodic playback.

[0083] Figure 3 A flowchart for constructing a prediction model to predict an atmospheric parameter is shown.

[0084] According to an embodiment of the present application, a prediction model is constructed based on the spatio-temporal dynamic graph data, real-time feature flow obtained through cloud fusion is taken as input, and an atmospheric parameter after a preset time is predicted, specifically:

[0085] In S302, the prediction model adopts a double-path architecture and processes stable items and non-stationary items respectively. In the stable item prediction path, the time convolution network receives stable item data of a past preset time and organizes it into a time sequence matrix according to nodes, performs feature extraction through causal dilated convolution, and obtains a trend basis vector of a future preset time.

[0086] In S304, spatial aggregation is performed through a graph attention network, a graph structure of geographical distance between laser radar nodes is constructed, multi-head attention is introduced for node-level feature interaction, a spatial correlation feature matrix is obtained, and the spatial correlation feature matrix and the trend basis vector generate stable item features.

[0087] In S306, in the non-stable item prediction path, a dynamic graph is constructed according to real-time edge weights and current non-stable item values, dynamic propagation is performed using a gated graph neural network, message passing iteration is performed through a gating mechanism, and layer normalization operation is performed after each iteration.

[0088] In S308, a grid centered on a target node is constructed for relative position coding, a self-attention mechanism is introduced to highlight gradient mutation areas, and an abnormal propagation feature map is output through a gated graph neural network as non-stable item features.

[0089] In S310, a cross-attention mechanism is used to fuse stable item features and non-stable item features, a residual connection is used to retain the original stable item trend basis vector, fusion features are generated, the fusion features are input into a multi-step prediction decoder constructed by a multi-layer perception mechanism, and a prediction value of an atmospheric parameter of a future preset time is output.

[0090] It should be noted that the prediction model adopts a dual-path architecture, which respectively processes the stable term and the non-stationary term, and finally fuses to generate the prediction result of the atmospheric parameter in the future period (such as 6 hours). The model input is the real-time feature flow after cloud fusion, and the output is the parameter prediction matrix constructed by the predicted value of the atmospheric parameter at the preset time step. In the stable term prediction path, a time convolution network is used for long-term trend modeling, which includes 4 layers of causal dilated convolution, each layer including: convolution kernel, weight normalization, ReLU activation and dropout. The daily cycle is captured by the bottom layer convolution, and the weekly and monthly scale trends are identified by the upper layer convolution. Spatial aggregation is used by graph attention network, the first layer of graph convolution performs node-level feature interaction, and the second layer of graph convolution performs regional-level pattern integration to output a spatial correlation feature matrix. In the non-stationary term prediction path, dynamic propagation is performed by a gated graph neural network, a dynamic graph is constructed using real-time edge weights and current non-stationary term values, and in the gating mechanism, the update gate controls the information retention ratio, and the reset gate determines the historical memory attenuation degree. Perform 3 iterations of message passing, and apply layer normalization operation after each iteration. In the self-attention enhancement module, a 9x9 grid centered on the target station is constructed as a local focusing window, relative position encoding is performed, the signal of the gradient mutation area is amplified by self-attention calculation, and an abnormal propagation feature map is obtained. Through the decoupling modeling of the stable term and the non-stationary term, the time positioning accuracy of the sudden pollution event is greatly improved.

[0091] It should be noted that in the collaborative decision and visualization module, the predicted atmospheric parameters are standardized, and 3 layers of 3D-CNN spatio-temporal convolution layers are used to construct an encoder to extract latent features from the standardized atmospheric parameter prediction values, and group normalization is introduced; a decoder is constructed using 3 layers of transposed spatio-temporal convolution to obtain reconstruction error, a baseline threshold is obtained based on historical normal data, and the baseline threshold is corrected according to temperature and humidity data, the reconstruction error is compared with the corrected baseline threshold, and the abnormal event is classified according to the comparison result; based on the abnormal event, the concentration gradient field in the monitoring grid is obtained, the observation abnormal propagation path is obtained according to the concentration gradient field, the candidate source is matched and obtained within the preset range of the abnormal event using the industrial facility database, the Lagrangian particle tracking based on the wind field data is used, the candidate source simulation path is generated based on the chemical component attenuation model, and the simulation path is corrected according to dry and wet deposition. Calculate the Fréchet distance between the observation abnormal propagation path and the candidate source simulation path, select the candidate source corresponding to the minimum Fréchet distance as the pollution source tracing result, and display it visually, wherein the source tracing result is displayed using the candidate source probability bubble chart and the optimal path red highlight. Through the integration of detection, source tracing and visualization means, the response time of pollution events is enhanced, and the source tracing accuracy in complex terrain areas is significantly improved.

[0092] Figure 4 A flowchart of a multi-modal laser radar networked atmospheric parameter collaborative monitoring method is shown.

[0093] The second embodiment of the present application provides a multi-modal laser radar networked atmospheric parameter cooperative monitoring method, applied to a multi-modal laser radar networked atmospheric parameter cooperative monitoring system, comprising the following steps:

[0094] S402, deploying multi-modal laser radar nodes, synchronously collecting multi-source atmospheric parameters, data preprocessing the multi-source atmospheric parameters, and multi-source feature fusion to generate a fusion feature matrix;

[0095] S404, dynamic networking by software defined network, transmission path selection by reinforcement learning routing to transmit the fusion feature matrix to the edge computing node, and cloud fusion by federated learning to aggregate the data uploaded by the edge node;

[0096] S406, constructing a spatio-temporal dynamic graph based on the spatio-temporal dependence relationship between the laser radar nodes and the nodes in the cloud, decomposing the spatio-temporal dynamic correlation into a stable term and a non-stationary term by spatio-temporal decoupling, and updating the spatio-temporal dynamic graph data by online learning using an alternating optimization strategy;

[0097] S408, constructing a prediction model based on the spatio-temporal dynamic graph data, taking the real-time feature flow obtained by cloud fusion as input, predicting the atmospheric parameters after a preset time, detecting abnormal events according to the predicted atmospheric parameters, and performing pollution tracing.

[0098] The third embodiment of the present application provides a computer readable storage medium, which comprises a multi-modal laser radar networked atmospheric parameter cooperative monitoring method program, and when the multi-modal laser radar networked atmospheric parameter cooperative monitoring method program is executed by a processor, the steps of the multi-modal laser radar networked atmospheric parameter cooperative monitoring method are realized.

[0099] In several embodiments provided in the present application, it should be understood that the disclosed method can be implemented in other ways. The above described embodiments are only illustrative, for example, the division of the units is only a logical function division, and actual implementation can have another division mode, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed components can be through some interface, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0100] The above-mentioned integrated unit of the present application, if realized in the form of a software functional unit and sold or used as an independent product, can also be stored in a computer-readable storage medium. Based on such understanding, the technical solutions of the embodiments of the present application can be embodied in the form of a software product, which is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in the various embodiments of the present application. The aforementioned storage medium includes various media that can store program codes, such as mobile storage devices, ROMs, RAMs, magnetic disks, or optical disks.

[0101] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A multi-modal laser radar networked atmospheric parameter collaborative monitoring system, characterized in that, The application relates to a multi-source atmospheric parameter prediction method based on a multi-source atmospheric parameter prediction system. The multi-source atmospheric parameter prediction system comprises a multi-modal laser radar module, a multi-source data fusion and preprocessing module, an intelligent networking optimization module, a space-time dynamic graph modeling module, a collaborative decision and visualization module, and a prediction model. The multi-modal laser radar module synchronously collects multi-source atmospheric parameters through distributed deployment of multi-modal laser radar nodes. The multi-source data fusion and preprocessing module calibrates, aligns and feature-fuses the multi-source atmospheric parameters to generate a fusion feature matrix. The intelligent networking optimization module adopts software-defined networking for dynamic networking, selects a transmission path based on reinforcement learning to transmit the fusion feature matrix to an edge computing node, and aggregates data uploaded by the edge node through federated learning to realize cloud fusion. The space-time dynamic graph modeling module constructs a space-time dynamic graph based on the space-time dependence relationship between the laser radar nodes and the nodes in the cloud, decouples the space-time dynamic correlation into a stable term and a non-stationary term through space-time decoupling, and updates the space-time dynamic graph data through online learning by adopting an alternating optimization strategy. A prediction model is constructed based on the space-time dynamic graph data, real-time feature flow obtained through cloud fusion is taken as input, and atmospheric parameters after a preset time are predicted.

2. The multi-modal laser radar networked atmospheric parameter collaborative monitoring system according to claim 1, characterized in that, The collaborative decision and visualization module detects abnormal events and performs pollution tracing according to the predicted atmospheric parameters, and dynamically visualizes real-time data, abnormal events and pollution tracing results. The multi-source data fusion and preprocessing module specifically comprises the following steps. Multi-source atmospheric parameters synchronously collected by the multi-modal laser radar are acquired, the multi-source atmospheric parameters comprise aerosol data, greenhouse gas data and wind field data, error correction is performed on the multi-source atmospheric parameters by using ground meteorological station data, data comparison is performed on the observation overlapping areas of different modal laser radars, dynamic cross-calibration is performed, the multi-source atmospheric parameters are aligned in space-time, outliers are removed through abnormal value processing, missing values are filled by using adjacent node data, preprocessed multi-source atmospheric parameters are acquired, physical features, statistical features and space-time features are respectively extracted based on the preprocessed multi-source atmospheric parameters, the physical features, statistical features and space-time features of the multi-source atmospheric parameters are organized into a multi-dimensional tensor according to a space-time network, specific feature coding is performed on different atmospheric parameters, a CNN network is used to process spatial distribution images of aerosol data, an LSTM is used to process time series of gas data, and a 3D CNN network is used to process wind speed fields; An attention mechanism is introduced to dynamically allocate the weights of the features of different modes, attention fusion features are acquired through weighted fusion, feature splicing features are acquired by splicing the features of different modes, and tensor fusion features are acquired by modeling mode interaction through outer product; A fusion feature matrix is generated based on the attention fusion features, the feature splicing features and the tensor fusion features. The intelligent networking optimization module specifically comprises the following steps.

3. The multi-modal laser radar networked atmospheric parameter collaborative monitoring system according to claim 1, characterized in that, A software-defined network controller is deployed to centrally manage a topology network and a traffic state formed by the multi-modal laser radar nodes, the laser radar nodes are taken as software-defined network switches, and data is forwarded according to the instructions of the software-defined network controller. ​ The virtual network slices are divided for different data types to guarantee the transmission bandwidth demand of the burst pollution event, and a software-defined network controller updates a topology network in real time by detecting a link state between nodes through an LLDP protocol; A network state vector is obtained according to a link delay, a bandwidth utilization, a node load and data priority of the topology network, a next hop node of data propagation is taken as an optional action, a forward reward is constructed based on low delay and high throughput, and a penalty term is constructed according to a packet loss rate exceeding a preset threshold; A deep Q network algorithm is trained using the forward reward and the penalty term, and an optimal path is adaptively selected by a software-defined network controller based on the trained deep Q network to transmit a fusion feature matrix to an edge node; After the edge node receives the fusion feature matrix, a lightweight CNN-LSTM model is used to perform anomaly detection, mark pollution anomaly data, and perform feature dimension reduction using principal component analysis, and when there is no pollution anomaly data, the data is temporarily stored in the edge node and uploaded on demand, and when pollution anomaly data appears, transmission is triggered immediately; Each edge node updates model parameters using local data, a cloud end receives model parameters of all edge nodes using a FedAvg algorithm, and performs weighted average according to data volume, the cloud end distributes the aggregated global model to the edge nodes, the next round of training is based on the new model for optimization, and the cloud end uses the aggregated global model to predict pollution anomaly data in the whole region, and the pollution anomaly data of the edge nodes and the pollution anomaly prediction data of the global model are fused to obtain spatiotemporal grid data.

4. The multi-modal laser radar networked atmospheric parameter collaborative monitoring system according to claim 1, characterized in that, In the spatiotemporal dynamic graph modeling module, a spatiotemporal dynamic graph is constructed based on laser radar nodes and spatiotemporal dependence between nodes in the cloud end, and spatiotemporal dynamic correlation is decomposed into a stable term and a non-stationary term through spatiotemporal decoupling, specifically: According to the laser radar nodes as nodes, a spatial edge and a time edge are constructed according to the geographical distance between nodes and historical data trends, a spatiotemporal dynamic graph is constructed based on node and edge structures, and a sliding time window is used to update the graph structure to represent real-time changes in atmospheric parameters; A static graph adjacency matrix is constructed based on the preprocessed spatiotemporal dynamic graph data, a graph Laplacian matrix is calculated, a graph Fourier transform is performed on a node feature matrix based on the graph Laplacian matrix, and a low-frequency component is reserved to reconstruct a signal; Each time slice signal is transformed using the low-frequency component reconstruction signal, a reconstructed stable term is obtained through low-frequency reconstruction, and a Savitzky-Golay filter is used to eliminate noise fluctuations and retain seasonal trends; Original observation values and stable term residuals are obtained as initial node states according to the spatiotemporal dynamic graph data, edge weights are constructed based on spatial weights, wind field correction and concentration driving, a dynamic adjacency matrix is constructed, representation learning is performed on the dynamic adjacency matrix through a dynamic graph neural network, and spatial propagation and time evolution are performed using spatiotemporal convolution; An exponential average in a preset time in the past is obtained based on EMA filtering, a current time prediction value is obtained based on the dynamic graph neural network, and a mutation feature is identified based on the exponential average and the current time prediction value to obtain a non-stationary term.

5. The multi-modal laser radar networked atmospheric parameter collaborative monitoring system according to claim 4, characterized in that, The updating of the spatio-temporal dynamic graph data is performed by online learning using an alternating optimization strategy, specifically: The updating of the spatio-temporal dynamic data is decomposed into two stages of stable item optimization and non-stable item optimization using expectation maximization, first, fix the stable item, optimize the non-stationary item, lock the current stable item parameter, calculate the residual signal of the observation value and the stable item, input the residual signal and real-time multi-source atmospheric parameters into the dynamic graph neural network, and perform gradient descent through adaptive dynamic learning rate to optimize the attention weight and gate parameter, and obtain the non-stationary item update result; Second, fix the non-stationary item, optimize the stable item, fix the current non-stationary item detection result, aggregate the denoised data in the past preset period, perform exponential weighted average on the graph Fourier coefficient, retain the top k main vibration modes, and fine-tune the mode amplitude through adaptive dynamic learning rate, without changing the vibration form, to obtain the stable item update result.

6. The multi-modal laser radar networked atmospheric parameter collaborative monitoring system according to claim 4, characterized in that, Based on the spatio-temporal dynamic graph data, a prediction model is constructed, and the real-time feature flow obtained by cloud fusion is used as input to predict the atmospheric parameters after a preset time, specifically: The prediction model uses a double-channel architecture to process the stable item and the non-stationary item, respectively, in the stable item prediction channel, the past preset time stable item data is received through the time convolution network and organized as a time sequence matrix by node, the feature is extracted through the causal inflation convolution, and the trend basis vector of the future preset time is obtained; Through the graph attention network, the spatial aggregation is performed, the graph structure of the geographical distance between the laser radar nodes is constructed, the multi-head attention is introduced for node-level feature interaction, and the spatial correlation feature matrix is obtained, and the spatial correlation feature matrix and the trend basis vector are generated as stable item features; In the non-stationary item prediction channel, a dynamic graph is constructed according to the real-time edge weight and the current non-stationary item value, a gated graph neural network is used for dynamic propagation, message passing iteration is performed through the gating mechanism, and layer normalization operation is performed after each iteration; A grid centered on the target node is constructed for relative position coding, the self-attention mechanism is introduced to highlight the gradient mutation area, and the abnormal propagation feature map output by the gated graph neural network is used as the non-stationary item feature; The cross-attention mechanism is used to fuse the stable item features and the non-stationary item features, and the original stable item trend basis vector is retained through the residual connection, and the fusion features are generated, which are input into the multi-step prediction decoder constructed by the multilayer perception mechanism to output the predicted value of the atmospheric parameters at the future preset time.

7. The multi-modal laser radar networked atmospheric parameter collaborative monitoring system according to claim 1, characterized in that, The collaborative decision and visualization module, specifically: The predicted atmospheric parameters are standardized, the 3-layer 3D-CNN spatio-temporal convolution layer is used to construct an encoder to extract the latent features of the standardized atmospheric parameter prediction values, and group normalization is introduced; A 3-layer transpose spatio-temporal convolution layer is used to construct a decoder to obtain the reconstruction error, a baseline threshold is obtained based on historical normal data, and the baseline threshold is corrected according to the temperature and humidity data, the reconstruction error and the corrected baseline threshold are compared, and the abnormal event is classified according to the comparison result. Based on the abnormal event, the observation abnormal propagation path in the monitoring grid is obtained, the candidate source in the abnormal event preset range is obtained, the candidate source simulation path is generated based on the Lagrangian particle tracking of the wind field data, the Fréchet distance of the observation abnormal propagation path and the candidate source simulation path is calculated, and the candidate source corresponding to the minimum Fréchet distance is selected as the pollution source tracing result and is visually displayed.

8. A multi-modal laser radar networked atmospheric parameter collaborative monitoring method, characterized in that, The application is applied to the multi-modal laser radar networked atmospheric parameter collaborative monitoring system in any one of claims 1-7, comprising the following steps: Deploying multi-modal laser radar nodes, synchronously collecting multi-source atmospheric parameters, data preprocessing of the multi-source atmospheric parameters, multi-source feature fusion of the preprocessed multi-source atmospheric parameters to generate a fusion feature matrix; Using software-defined network for dynamic networking, transmitting the fusion feature matrix to the edge computing node based on the reinforcement learning routing selection transmission path, and realizing cloud fusion by aggregating the data uploaded by the edge node through federated learning; Based on the spatio-temporal dependence relationship between the laser radar nodes and the nodes, a spatio-temporal dynamic graph is constructed in the cloud, the spatio-temporal dynamic correlation is decomposed into a stable term and a non-stationary term through spatio-temporal decoupling, and the spatio-temporal dynamic graph data is updated through online learning using an alternating optimization strategy; Based on the spatio-temporal dynamic graph data, a prediction model is constructed, real-time feature flow obtained through cloud fusion is taken as input, atmospheric parameters after a preset time are predicted, and abnormal events are detected according to the predicted atmospheric parameters and pollution source tracing is performed.

Images