A method and system for automatic prediction of multiple environmental parameters based on GIS maps

Abstract

This application relates to the technical field of environmental monitoring, and discloses an automatic prediction method for multiple environmental parameters based on GIS maps. The method includes: real-time collection of noise intensity, vibration spectrum, wastewater turbidity, light intensity, and PM2.5 concentration through a distributed sensor network; spatiotemporal alignment of this data with social behavior data to generate a multidimensional spatiotemporal matrix; calculation of energy overlap using a spatiotemporal graph attention network to mark high-risk co-polluting areas and predicting pollution diffusion paths; construction of an adaptive prediction model including physical and behavioral driving channels, dynamically adjusting weights to optimize prediction accuracy; back-inference of pollution sources in areas exceeding standards based on the optimized model, matching equipment characteristics and generating control commands; and opening an AR augmented reality interface to verify the governance effect. When the deviation between virtual and real data exceeds a data deviation threshold, federated learning is triggered to update the global model and generate an environmental compliance report. This application can improve the efficiency of accurate decision-making in environmental governance.

Description

Technical Field

[0001] This application relates to the technical field of environmental monitoring, and in particular to a method and system for automatic prediction of multiple environmental parameters based on GIS maps. Background Technology

[0002] Currently, the field of environmental monitoring generally suffers from insufficient integration of multi-source data and weak dynamic prediction capabilities. Traditional environmental monitoring systems mostly rely on single sensors to collect basic parameters such as noise and particulate matter, making it difficult to effectively integrate dynamic behavioral data such as traffic flow and social media sentiment, resulting in significant lags in the spatiotemporal correlation analysis of pollution events.

[0003] While existing GIS platforms can visualize basic data, they generally use static grid partitioning and linear prediction models, which cannot capture the nonlinear interaction characteristics between complex environmental parameters. Especially in sudden pollution events, existing systems lack the ability to dynamically correct pollution diffusion paths, resulting in large deviations between prediction results and measured data, which seriously restricts the efficiency of accurate decision-making in environmental governance.

[0004] As can be seen from the above, how to improve the efficiency of accurate decision-making in environmental governance remains to be solved. Summary of the Invention

[0005] To improve the efficiency of accurate decision-making in environmental governance, this application provides a method and system for automatic prediction of multiple environmental parameters based on GIS maps.

[0006] Firstly, this application provides an automatic prediction method for multiple environmental parameters based on GIS maps, employing the following technical solution:

[0007] An automatic prediction method for multiple environmental parameters based on GIS maps includes:

[0008] The system collects real-time data on noise intensity, vibration spectrum, sewage turbidity, light intensity, and PM2.5 concentration in the target area. This data is transmitted to a GIS platform via a distributed sensor network. Simultaneously, it acquires real-time changes in traffic flow, spatiotemporal distribution of social media keywords, and short-term fluctuations in meteorological parameters in the target area. Based on GIS spatial gridding technology, the environmental parameters and dynamic behavior data are spatiotemporally aligned according to a preset format to generate a multi-dimensional spatiotemporal matrix that integrates physical parameters and social behavior. The physical parameters include noise intensity, vibration spectrum, sewage turbidity, light intensity, and PM2.5 concentration data, while the social behavior includes real-time changes in traffic flow, spatiotemporal distribution of social media keywords, and short-term fluctuations in meteorological parameters.

[0009] Based on the multidimensional spatiotemporal matrix, the energy distribution characteristics of the noise main frequency band and the spectral energy distribution characteristics of the vibration sensor are extracted. The energy overlap between the energy distribution characteristics and the spectral energy distribution characteristics is calculated. When the energy overlap exceeds the overlap threshold, it is marked as a high-risk co-polluting area. The sewage flow rate, real-time wind speed, precipitation data and co-polluting characteristics corresponding to the high-risk co-polluting area are obtained. The pollution diffusion path and concentration gradient change are predicted by combining the sewage flow rate, real-time wind speed, precipitation data and co-polluting characteristics. The spatiotemporal graph attention network is used to dynamically learn the nonlinear correlation between parameters and output the co-influence weight.

[0010] Based on the aforementioned synergistic influence weights, an adaptive prediction model is constructed that includes physical driving channels and behavioral driving channels. The physical channel generates baseline prediction results based on noise attenuation equations and CFD fluid simulations. The behavioral channel generates prediction correction factors by analyzing the temporal correlation between traffic flow and social data through an LSTM network. When the deviation between physical predictions and measured data exceeds the data deviation threshold for multiple consecutive periods, the weight of the physical channel is reduced and the weight of the behavioral channel is increased. At the same time, a digital twin sandbox is launched in GIS to simulate pollution source migration scenarios and generate virtual training data to optimize the adaptive prediction model.

[0011] Based on the optimized adaptive prediction model, the pollution diffusion path of the area exceeding the standard is reversed, and the target equipment is located by combining sensor data; the equipment feature matching model is invoked to perform wavelet packet decomposition on the noise and vibration signals, and the vibration features in the equipment feature fingerprint database are matched. When the match is successful, a control command is generated and sent to the IoT terminal device to execute the operation of shutting down the pump that exceeds the standard or starting the noise reduction barrier.

[0012] Based on the execution results of IoT terminals and real-time monitoring data, an AR augmented reality verification interface is opened to retrieve the predicted and measured data from the GIS platform. A deviation heatmap is generated based on the predicted and measured data, and the corresponding virtual-real data deviation is obtained. When the virtual-real data deviation exceeds 20%, the federated learning framework is triggered to aggregate the local gradients of edge nodes to update the global prediction model. Historical versions are retained to build a model evolution tree for anomaly backtracking analysis. A corresponding environmental compliance report is generated, which includes details of parameters exceeding standards, regulatory correlation analysis, control effect assessment, actionable recommendations, and blockchain-stored evidence data.

[0013] By adopting the above technical solutions, a dynamic decision-making closed loop for environmental governance is constructed through the spatiotemporal fusion and intelligent modeling of multi-source heterogeneous data, significantly improving the efficiency of precise decision-making. The system integrates physical environmental parameters and social behavior data in real time, and through adaptive prediction models with dynamic weight adjustment and digital twin simulation, it can quickly locate pollution synergistic risks and accurately predict diffusion paths. Combined with equipment feature matching and AR verification feedback, targeted control commands can be quickly generated and governance strategies can be dynamically optimized. At the same time, through federated learning and model evolution tree mechanisms, the prediction accuracy is continuously iterated, ultimately forming an automated closed loop of "monitoring-prediction-control-verification-optimization". This greatly shortens the decision response time, reduces human intervention errors, and enables environmental governance to shift from passive response to proactive, precise, and dynamic intelligent regulation.

[0014] Optionally, the method further includes:

[0015] The energy distribution characteristics of the main frequency band of noise are extracted by short-time Fourier transform. The energy distribution characteristics include the peak energy of the main frequency, the energy ratio of the frequency band, and the slope of the energy distribution curve.

[0016] Vibration spectrum analysis is used to extract the spectral energy distribution characteristics of vibration sensors, including the peak frequency of vibration energy, the amplitude of energy fluctuation in the frequency band, and the spectral energy density.

[0017] The calculation of the dynamic overlap between the energy distribution of the noise main frequency band and the energy distribution of the vibration spectrum includes: standardizing the frequency domain energy distribution curve of the energy distribution characteristics and the frequency domain energy distribution curve of the spectrum energy distribution characteristics, and using the weighted cosine similarity calculation formula, where the weight of the high frequency band is higher than that of the low frequency band.

[0018] When the dynamic overlap exceeds the overlap threshold and the difference in the proportion of high-frequency band energy is less than the difference in the proportion of high-frequency band energy, it is marked as a high-risk co-contamination area.

[0019] By adopting the above technical solutions, and through multi-dimensional feature analysis and dynamic similarity assessment, the accurate identification and location of synergistic risks of environmental pollution are achieved, significantly improving the decision-making efficiency of environmental governance. Through weighted analysis of high-frequency and low-frequency bands, combined with dynamic matching of energy distribution characteristics, the pollution source area under the synergistic effect of noise and vibration can be quickly located, avoiding misjudgments caused by single parameters or static thresholds in traditional methods. This allows governance resources to be targeted to high-risk pollution sources, reducing the redundancy of large-scale investigations. At the same time, through real-time dynamic assessment and threshold constraints, the accuracy and timeliness of decision-making are ensured, ultimately achieving targeted, intelligent, and rapid response in pollution prevention and control.

[0020] Optionally, in the process of dynamically learning the nonlinear correlation between parameters using a spatiotemporal graph attention network, the method further includes:

[0021] The node feature matrix of the spatiotemporal graph is constructed. The node feature matrix includes environmental parameters, dynamic behavior parameters and spatial correlation parameters. Among them, the environmental parameters include noise intensity, vibration spectrum energy peak, sewage turbidity, PM2.5 concentration and light intensity; the dynamic behavior parameters include traffic flow density, social media keyword mention frequency, wind speed and precipitation intensity; and the spatial correlation parameters include pollution diffusion path prediction values ​​and concentration gradient change values ​​of adjacent grids.

[0022] A dual-channel attention mechanism is introduced into the graph attention network, where the spatial channel calculates spatial attention weights based on grid adjacency relationships, and the temporal channel calculates temporal attention weights based on time series autocorrelation.

[0023] When the noise intensity is below the lower noise threshold, the environmental parameter weights of the corresponding grid are dynamically reduced to 30% of their original weights.

[0024] By adopting the above technical solutions, through the dual-channel mechanism (spatial and temporal dimensions) and dynamic weight adjustment of the spatiotemporal graph attention network, the accuracy and robustness of complex environmental parameter correlation modeling are significantly improved. In particular, it effectively suppresses interference factors in low-noise scenarios, enabling the model to more accurately capture the spatiotemporal patterns of pollution diffusion, thereby providing more reliable decision support for environmental governance.

[0025] Optionally, in the process of launching the digital twin sandbox to simulate the migration scenario of pollution sources, the method further includes:

[0026] A multiphysics coupled simulation model is constructed, which includes an air pollution diffusion module and a ground vibration diffusion module. The air pollution diffusion module simulates the air propagation path of PM2.5 and noise based on CFD fluid simulation, and the ground vibration diffusion module simulates the vibration energy attenuation in high-risk areas based on the vibration wave equation.

[0027] A spatiotemporal decay factor is introduced into the GIS grid. The spatiotemporal decay factor is calculated based on an exponential decay function of distance and time.

[0028] When the deviation between the simulated path and the measured pollution diffusion path exceeds the path deviation threshold, the sandbox is triggered to adaptively adjust the migration probability distribution parameters. Specifically, this includes increasing the contribution weight of wind speed to pollutant diffusion by 20% and recalculating the nonlinear relationship between precipitation intensity and pollutant deposition rate based on historical data.

[0029] By adopting the above technical solutions, through multi-physics coupled simulation and adaptive parameter calibration, the accuracy and dynamic adaptability of pollution diffusion path simulation are significantly improved. When measured deviations occur, model parameters (such as wind speed weight and precipitation influence) can be quickly adjusted to ensure that the simulation results are highly consistent with the real scene, providing a reliable basis for accurate location of pollution sources and formulation of governance strategies.

[0030] Optionally, in the implementation of the device feature matching model, the method further includes:

[0031] Three-level wavelet packet decomposition is performed on noise and vibration signals to extract vibration features in the high-frequency, mid-frequency, and low-frequency bands;

[0032] The decomposed vibration characteristics are then matched with the device fingerprint database using a weighted voting method.

[0033] A successful match is determined when the similarity of all three frequency bands exceeds the similarity threshold and the weighted total similarity is greater than or equal to the weighted total similarity threshold. When the matching failure rate continuously exceeds the matching failure rate threshold, the fingerprint database incremental update mechanism is triggered, which automatically retrieves vibration data from a preset time period and updates the fingerprint database feature vector using an online learning algorithm.

[0034] By adopting the above technical solution, the weighted matching of multi-band vibration characteristics and the dynamic fingerprint database update mechanism significantly improve the accuracy of pollution source equipment identification, ensuring rapid location of pollution sources in complex environments. At the same time, the matching model is continuously optimized through adaptive learning, reducing misjudgments and improving the speed and accuracy of treatment response.

[0035] Optionally, in the process of back-calculating the pollution diffusion path of the area exceeding the standard based on the optimized adaptive prediction model, the method further includes:

[0036] A multi-scale spatiotemporal inversion model is constructed, dividing the pollution diffusion path into micro, meso, and macro scales. At the micro scale, the precise coordinates of pollution source equipment are located based on sensor grid data. At the meso scale, the propagation path of pollutants from the pollution source to the area exceeding the standard is inferred by combining traffic flow and wind speed data. At the macro scale, the spatiotemporal distribution pattern of regional pollution sources is analyzed through historical pollution event data.

[0037] When locating target devices, a dynamic confidence assessment mechanism is adopted. The device position is calculated based on the confidence level of sensor data, where the confidence level is the matching degree between noise intensity and vibration energy. When the confidence level is lower than 80%, the collaborative inversion of adjacent grids is triggered, and the positioning accuracy is improved through multi-grid data fusion.

[0038] When generating control commands, a graded response strategy is introduced. When the parameter exceeding the standard is noise and the duration exceeds the preset duration threshold, the noise reduction barrier is activated first. When the parameter exceeding the standard is sewage turbidity and the concentration exceeds the sewage concentration trigger value, the pump is shut down and the surrounding sewage treatment facilities are linked to start the purification process.

[0039] By adopting the above technical solutions, the accuracy and efficiency of pollution source tracing and control are significantly improved through multi-scale inversion and intelligent hierarchical response mechanisms: the multi-scale model combines micro-level positioning, meso-level path analysis and macro-level spatiotemporal patterns to achieve high-precision dynamic tracking of pollution sources; dynamic confidence assessment and collaborative inversion ensure the reliability of positioning, while the hierarchical response strategy automatically triggers targeted control measures according to the type and severity of pollution, greatly shortening decision-making time and improving control effectiveness.

[0040] Secondly, this application provides an automatic prediction system for multiple environmental parameters based on GIS maps, employing the following technical solution:

[0041] An automatic prediction system for multiple environmental parameters based on GIS maps includes:

[0042] The multidimensional spatiotemporal matrix generation module collects real-time data on noise intensity, vibration spectrum, sewage turbidity, light intensity, and PM2.5 concentration in the target area. This data is transmitted to the GIS platform via a distributed sensor network. Simultaneously, it acquires real-time changes in traffic flow, spatiotemporal distribution of social media keywords, and short-term fluctuations in meteorological parameters in the target area. Based on GIS spatial gridding technology, the environmental parameters and dynamic behavior data are spatiotemporally aligned according to a preset format to generate a multidimensional spatiotemporal matrix that integrates physical parameters and social behavior. The physical parameters include noise intensity, vibration spectrum, sewage turbidity, light intensity, and PM2.5 concentration data, while the social behavior includes real-time changes in traffic flow, spatiotemporal distribution of social media keywords, and short-term fluctuations in meteorological parameters.

[0043] The collaborative influence weight output module extracts the energy distribution characteristics of the main frequency band of noise and the spectral energy distribution characteristics of the vibration sensor based on the multi-dimensional spatiotemporal matrix, calculates the energy overlap between the energy distribution characteristics and the spectral energy distribution characteristics, and marks the area as a high-risk collaborative pollution area when the energy overlap exceeds the overlap threshold; it obtains the sewage flow rate, real-time wind speed, precipitation data and collaborative pollution characteristics corresponding to the high-risk collaborative pollution area, and predicts the pollution diffusion path and concentration gradient changes by combining the sewage flow rate, real-time wind speed, precipitation data and collaborative pollution characteristics; it uses a spatiotemporal graph attention network to dynamically learn the nonlinear correlation between parameters and uses it to output the collaborative influence weight.

[0044] The adaptive prediction model optimization module constructs an adaptive prediction model containing physical driving channels and behavioral driving channels based on the synergistic influence weights. The physical channel generates baseline prediction results based on noise attenuation equations and CFD fluid simulation. The behavioral channel generates prediction correction factors by analyzing the temporal correlation between traffic flow and social data through an LSTM network. When the deviation between physical prediction and measured data exceeds the data deviation threshold for multiple consecutive periods, the weight of the physical channel is reduced and the weight of the behavioral channel is increased. At the same time, a digital twin sandbox is launched in GIS to simulate pollution source migration scenarios and generate virtual training data for optimizing the adaptive prediction model.

[0045] The matching module, based on the optimized adaptive prediction model, reverses the pollution diffusion path in the area exceeding the standard and locates the target equipment by combining sensor data; it retrieves the equipment feature matching model, performs wavelet packet decomposition on the noise and vibration signals, and uses it to match the vibration features in the equipment feature fingerprint database. When the match is successful, it generates control commands and sends them to the IoT terminal device to execute the operation of shutting down the pump that exceeds the standard or activating the noise reduction barrier.

[0046] The environmental compliance report generation module, based on the execution results of IoT terminals and real-time monitoring data, opens an AR augmented reality verification interface, retrieves predicted and measured data from the GIS platform, generates a deviation heatmap based on the predicted and measured data, and obtains the corresponding virtual-real data deviation. When the virtual-real data deviation exceeds 20%, the federated learning framework is triggered to aggregate the local gradients of edge nodes to update the global prediction model, and historical versions are retained to build a model evolution tree for anomaly backtracking analysis. This module is used to generate the corresponding environmental compliance report, which includes details of parameters exceeding standards, regulatory correlation analysis, control effect assessment, actionable recommendations, and blockchain-stored evidence data.

[0047] Thirdly, this application provides an automatic prediction method for multiple environmental parameters based on GIS maps, employing the following technical solution:

[0048] An automatic prediction method for multiple environmental parameters based on GIS maps includes a processor, wherein the processor runs a program for the automatic prediction method for multiple environmental parameters based on GIS maps as described above.

[0049] Fourthly, this application provides a storage medium, which adopts the following technical solution:

[0050] A storage medium storing a program for the automatic prediction method of multiple environmental parameters based on GIS maps as described in any one of the above.

[0051] In summary, this application includes at least one of the following beneficial technical effects:

[0052] First, by integrating multi-source heterogeneous data in real time and using intelligent modeling, a dynamic decision-making closed loop for environmental governance is constructed, significantly improving the accuracy of pollution collaborative risk identification and location. Through high-frequency and low-frequency weighted analysis, multi-scale inversion models, and dynamic weight adjustment mechanisms, pollution sources can be quickly located and diffusion paths can be accurately predicted, avoiding misjudgments caused by traditional methods relying on single parameters or static thresholds. This allows for targeted allocation of governance resources, reduces redundancy in large-scale investigations, and significantly shortens pollution location time.

[0053] Secondly, through intelligent hierarchical response and adaptive model optimization, the transformation of environmental governance from passive response to proactive regulation has been achieved. Multi-scale inversion combined with dynamic confidence assessment ensures the reliability of pollution source location. The hierarchical control strategy automatically triggers targeted measures (such as noise reduction barriers and pump shutdown) based on the type of pollution (e.g., noise, wastewater turbidity) and severity, significantly shortening the decision-making response time. At the same time, the model continuously iterates through digital twin simulation and federated learning to ensure prediction accuracy and timeliness, reducing human intervention errors.

[0054] Finally, a closed-loop verification and continuous optimization mechanism ensures the long-term effectiveness of the governance strategy. Augmented reality feedback and deviation heatmap analysis can correct model biases in real time, while federated learning and model evolution tree mechanisms enable dynamic updates of the prediction model to adapt to environmental changes. This fully automated closed loop significantly improves the intelligence level of environmental governance, making decision-making more accurate and execution more efficient, ultimately achieving targeted, dynamic, and sustainable pollution prevention and control. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating an automatic prediction method for multiple environmental parameters based on a GIS map, according to an exemplary embodiment.

[0056] Figure 2 This is a structural block diagram of an automatic prediction system for multiple environmental parameters based on a GIS map, according to an exemplary embodiment. Detailed Implementation

[0057] The embodiments of this application are described in detail below, and examples of the embodiments are shown in the accompanying drawings.

[0058] In the description of this specification, the references to "certain embodiments," "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples" refer to specific features, structures, materials, or characteristics described in connection with the described embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0059] This application discloses a method for automatic prediction of multiple environmental parameters based on GIS maps, referring to... Figure 1 ,include:

[0060] The S100 collects real-time data on noise intensity, vibration spectrum, sewage turbidity, light intensity, and PM2.5 concentration in the target area. This data is transmitted to the GIS platform via a distributed sensor network. Simultaneously, it acquires real-time changes in traffic flow, the spatiotemporal distribution of social media keywords, and short-term fluctuations in meteorological parameters in the target area. Based on GIS spatial gridding technology, it aligns environmental parameters and dynamic behavior data in a preset format to generate a multi-dimensional spatiotemporal matrix that integrates physical parameters and social behavior.

[0061] The physical parameters include noise intensity, vibration spectrum, sewage turbidity, light intensity, and PM2.5 concentration data, while the social behaviors include real-time changes in traffic flow, spatiotemporal distribution of keywords in social media environments, and short-term fluctuations in meteorological parameters.

[0062] It should be noted that, in this embodiment of the application, the corresponding parameters are obtained through sensor network deployment or other methods, as detailed below:

[0063] Noise sensors: Real-time noise intensity is collected using sound level meters (such as MEMS microphone arrays), and vibration spectrometers are extracted using spectrum analyzers; Water quality sensors: Conductivity sensors monitor wastewater turbidity, and pH sensors assist in identifying water quality anomalies; Environmental parameter sensors: PM2.5 sensors (such as laser scattering sensors) and light intensity sensors (photoresistors or photodiodes); Traffic flow: Real-time traffic flow statistics are obtained through geomagnetic sensors or cameras (combined with image recognition); Social media data: The spatiotemporal distribution of environmental keywords (such as "noise" and "pollution") within the region is captured through API interfaces; Meteorological data: Short-term fluctuation data such as wind speed and precipitation intensity are obtained by integrating meteorological bureau APIs.

[0064] For spatiotemporal alignment and gridding, the first step is to perform corresponding GIS grid division, which involves dividing the target area into 50m×50m square grids, with each grid corresponding to a spatial unit. Then, time-slice design is implemented, with each time slice lasting 5 minutes to ensure the real-time nature of data collection and analysis. In other words, the preset format is a 50m×50m square grid and a time slice lasting 5 minutes.

[0065] The data alignment method in this application embodiment unifies the time resolution for data with different sampling frequencies (such as data collected by sensors every second and social media updates every minute) through interpolation or sliding window averaging. In addition, the spatial alignment method is achieved by merging sensor data into its respective grid and mapping social media data to the nearest grid according to the user's GPS location.

[0066] Finally, regarding the matrix dimensions, which include spatial dimensions (grid ID), temporal dimensions (time slices), and parameter dimensions (physical + social behavioral parameters), parameters of different dimensions (such as noise dB, PM2.5 μg / m³) are standardized (e.g., Z-score standardization).

[0067] By unifying the modeling of physical environment and social behavior data, for example, we can discover the correlation between abnormally high PM2.5 concentrations during peak traffic hours; and a 50m×50m grid can accurately locate pollution sources (such as noise sources at a construction site), and capture short-term pollution events (such as sudden sewage leaks) in 5-minute time slices.

[0068] S200: Based on a multi-dimensional spatiotemporal matrix, the energy distribution characteristics of the noise main frequency band and the spectral energy distribution characteristics of the vibration sensor are extracted. The energy overlap between the energy distribution characteristics and the spectral energy distribution characteristics is calculated. When the energy overlap exceeds the overlap threshold, it is marked as a high-risk co-polluting area. The sewage flow rate, real-time wind speed, precipitation data and co-polluting characteristics corresponding to the high-risk co-polluting area are obtained. The pollution diffusion path and concentration gradient change are predicted by combining the sewage flow rate, real-time wind speed, precipitation data and co-polluting characteristics. The spatiotemporal graph attention network is used to dynamically learn the nonlinear correlation between parameters and output the co-influence weight. In this embodiment, the overlap threshold is 70%.

[0069] It should be noted here that the method also includes:

[0070] S211 extracts the energy distribution characteristics of the main frequency band of noise through short-time Fourier transform. The energy distribution characteristics include the peak energy of the main frequency, the energy ratio of the frequency band, and the slope of the energy distribution curve.

[0071] In this embodiment, the noise sensor collects noise intensity in real time using a sound level meter (such as a MEMS microphone array) and extracts the vibration spectrum using a spectrum analyzer. To accurately capture high-frequency components (such as vibration noise from mechanical equipment), short-time Fourier transform (STFT) analysis is used: a 1-second analysis window is selected, and the edges are smoothed using a Hamming window to reduce spectral leakage. The window overlap rate is set to 50% to balance time and spectral resolution.

[0072] It should be noted that key features include: peak frequency energy, which locates the frequency with the highest energy in the STFT spectrum (e.g., 160Hz), reflecting the dominant frequency of the noise source; frequency band energy ratio, which is the proportion of energy in the high-frequency band (150-200Hz) and the low-frequency band (50-100Hz) to the total energy, used to distinguish between equipment noise and environmental noise; and energy distribution curve slope, which involves linearly fitting the spectrum curves of the high-frequency and low-frequency bands and calculating the slope difference to identify abnormal pollution sources (e.g., high-speed equipment).

[0073] S212 extracts the spectral energy distribution characteristics of the vibration sensor through vibration spectrum analysis. The spectral energy distribution characteristics include the peak frequency of vibration energy, the amplitude of frequency band energy fluctuation, and the spectral energy density.

[0074] In this embodiment, the time-domain signal acquired by the vibration sensor (such as an accelerometer) is converted into a frequency-domain energy distribution through a Fast Fourier Transform (FFT). To satisfy the Nyquist sampling theorem, the sampling rate is set to twice the highest target frequency (e.g., a target frequency of 200Hz corresponds to a sampling rate of 400Hz).

[0075] It should be noted that key features include: peak vibration energy frequency, which locates the maximum energy value (e.g., 180Hz) in the FFT spectrum, reflecting the inherent vibration frequency of the equipment; frequency band energy fluctuation amplitude, the standard deviation of energy in the high-frequency band (150-200Hz), used to assess the operational stability of the equipment; and spectral energy density, the energy value per unit frequency band, used for correlation analysis with the noise spectrum.

[0076] S213, calculate the dynamic overlap between the energy distribution of the main frequency band of noise and the energy distribution of the vibration spectrum, including: standardizing the frequency domain energy distribution curve of the energy distribution characteristics and the frequency domain energy distribution curve of the spectrum energy distribution characteristics, and calculating the weighted cosine similarity using a formula where the weight of the high frequency band is higher than that of the low frequency band. Specifically, 150-200Hz is the high frequency band with a weight of 0.6, and 50-100Hz is the high frequency band with a weight of 0.4. When the dynamic overlap exceeds the overlap threshold and the difference in the energy proportion of the high frequency band is less than the difference in the energy proportion of the high frequency band, it is marked as a high-risk co-polluting area.

[0077] In this embodiment of the application, the frequency domain energy distribution curves of noise and vibration are normalized (e.g., normalized to the [0,1] range) to eliminate sensor differences. The similarity between the high frequency band (weight 60%) and the low frequency band (weight 40%) can be calculated using weighted cosine similarity, and the total overlap is the weighted sum of the two.

[0078] It should be noted that the threshold for the difference in energy proportion in the high-frequency band is 15%. In other words, the threshold constraints include: dynamic overlap > 70%, indicating that the noise and vibration energy distributions are highly similar and may originate from the same pollution source (such as equipment malfunction); and high-frequency band energy proportion difference < 15%, ensuring high-frequency band consistency and avoiding sensor noise or environmental interference. For example, if the noise and vibration energy proportions in the high-frequency band in a certain area are 40% and 45% respectively, and the overlap reaches 75%, it is marked as a "high-risk co-contamination area".

[0079] By employing multi-dimensional feature extraction and dynamic correlation analysis, high-risk pollution areas are accurately identified: First, short-time Fourier transform (STFT) and Hamming window are used to capture high-frequency components of noise signals, extracting the peak energy of the dominant frequency, the energy proportion of the frequency band, and the slope of the energy distribution curve to quantify the noise spectrum characteristics. Second, FFT analysis of vibration sensors is used to extract the peak frequency of vibration energy, the amplitude of frequency band energy fluctuations, and the spectral energy density to locate the vibration source of the equipment and assess its stability. Finally, through standardization processing and weighted cosine similarity calculation (60% weight for high-frequency bands and 40% weight for low-frequency bands), the similarity between noise and vibration energy distribution is dynamically evaluated. Combined with dual threshold constraints of dynamic overlap >70% and high-frequency band energy proportion difference <15%, environmental noise interference is eliminated, and high-risk co-polluting areas (such as abnormal equipment operation) are accurately marked, enabling rapid location of pollution sources and suppression of misjudgments, thereby improving the accuracy and timeliness of environmental governance.

[0080] Furthermore, in the process of dynamically learning the nonlinear relationships between parameters using spatiotemporal graph attention networks, the method also includes:

[0081] S214. Construct the node feature matrix of the spatiotemporal graph. The node feature matrix includes environmental parameters, dynamic behavior parameters, and spatial correlation parameters. Among them, the environmental parameters include noise intensity, vibration spectrum energy peak, sewage turbidity, PM2.5 concentration, and light intensity. The dynamic behavior parameters include traffic flow density, social media keyword mention frequency, wind speed, and precipitation intensity. The spatial correlation parameters include the predicted pollution diffusion path and concentration gradient change values ​​of adjacent grids.

[0082] In this embodiment, multi-source data is integrated through a node feature matrix, specifically including: environmental parameters such as noise intensity, peak vibration spectrum energy, wastewater turbidity (via conductivity sensor), PM2.5 concentration (via laser scattering sensor), and light intensity; dynamic behavioral parameters such as traffic flow (via geomagnetic sensor or camera statistics), frequency of mentions of the keyword "pollution" on social media (via API crawling), and wind speed and precipitation intensity (via meteorological bureau API); and spatial correlation parameters such as predicted pollution diffusion paths and concentration gradient changes between adjacent grids.

[0083] S215 introduces a dual-channel attention mechanism in the graph attention network. The spatial channel calculates the spatial attention weight based on the grid adjacency relationship, and the temporal channel calculates the temporal attention weight based on the time series autocorrelation. When the noise intensity is lower than the noise threshold, the environmental parameter weight of the corresponding grid is dynamically reduced to 30% of the original weight.

[0084] It's important to note the dual-channel attention mechanism: the spatial channel calculates weights based on grid adjacency relationships (e.g., 4-neighborhood), for example, assigning higher weights to sudden increases in PM2.5 concentration in adjacent grids; the temporal channel calculates weights based on the autocorrelation of historical data, for example, assigning higher weights to surges in traffic flow over the past hour. The noise threshold is 50 dB. Regarding dynamic weight adjustment, when the grid noise intensity is below 50 dB, the environmental parameter weights dynamically decrease to 30% of their original values ​​to reduce interference from minor vibrations or environmental noise.

[0085] Next, high-risk area labeling and pollution diffusion path prediction are performed. This requires combining the dynamic overlap and high-frequency band proportion difference thresholds from the third step to label high-risk areas. It's important to note that the spatiotemporal graph attention network outputs the contribution of each parameter to pollution diffusion (e.g., the weight of noise on PM2.5 diffusion) and integrates parameters such as wastewater flow rate and wind speed. The spatiotemporal graph convolutional network (ST-GCN) then predicts pollution diffusion paths and concentration changes. For example, if the wind speed is towards residential areas, the model predicts that PM2.5 concentration will increase in that area, guiding the precise allocation of pollution control resources.

[0086] In addition, multi-dimensional data can be acquired through sensor network deployment: physical environment data, such as noise sensors, water quality sensors, and PM2.5 sensors; social behavior data, such as traffic flow, social media keywords, and meteorological data.

[0087] In this embodiment, the GIS grid is divided into 50m × 50m square grids, each corresponding to a spatial unit; the time slice is designed to update every 5 minutes to ensure real-time performance. It should be noted that regarding data alignment: for time alignment, interpolation or sliding window averaging is used to unify the time resolution for sensor data per second and social media data per minute; for spatial alignment, sensor data is merged into its respective grid, and social media data is mapped to the nearest grid based on the user's GPS. The multidimensional spatiotemporal matrix dimensions include space (grid ID), time (5 minutes), and parameters (standardized physical + social behavior parameters), such as Z-score standardization to handle the dimensional difference between noise dB and PM2.5 μg / m³.

[0088] It should be noted that in the spatiotemporal graph attention network, a dual-channel attention mechanism (spatial and temporal channels) is used to dynamically capture the nonlinear relationships between parameters and output the weights of the synergistic effects of each parameter on pollution diffusion. The specific implementation is as follows:

[0089] First, there's the spatial channel attention mechanism, which calculates spatial weights based on grid adjacency relationships (such as 4-neighborhoods or 8-neighborhoods). For example, by analyzing the similarity of node features between adjacent grids (such as the similarity of parameters like noise intensity and PM2.5 concentration), and combining this with the adjacency matrix (which defines the physical distance or pollution diffusion path between grids), the influence of neighboring grids on the current grid is weighted. If the PM2.5 concentration of an adjacent grid suddenly increases, the model assigns it a higher spatial weight to identify the direction of pollution diffusion. The spatial weights are dynamically calculated through a graph attention layer, specifically by performing softmax normalization on the product of node feature similarity and the adjacency matrix to obtain the spatial correlation strength between adjacent grids.

[0090] Next is the temporal attention mechanism, which calculates time weights based on the autocorrelation of time series data. For example, by analyzing the correlation between historical data (such as traffic flow and wind speed changes in the past hour) and current pollution parameters, and combining this with time decay coefficients (such as exponential decay factors to ensure that recent data has a greater impact on the prediction), the model captures the dynamic correlation of parameters over time. If a surge in traffic flow in a certain area leads to an increase in PM2.5 concentration, the model will assign higher time weights to historical traffic data. The time weights are dynamically calculated through the temporal attention layer, specifically by performing softmax normalization on the product of historical data correlation and the time decay coefficient to determine the contribution of historical data to the current prediction.

[0091] In low-noise scenarios (e.g., mesh noise intensity < 50 dB), the weights of environmental parameters (e.g., peak vibration energy, wastewater turbidity) are dynamically reduced to 30% of their original values ​​to suppress the interference of minor environmental noise or sensor errors on the model. For example, if the noise intensity of a mesh is 45 dB, the weight of its vibration energy parameter will be reduced from 0.5 to 0.15, reducing the noise interference of irrelevant parameters on the model.

[0092] Furthermore, the network fuses spatial and temporal attention weights through multi-layer graph convolution, ultimately outputting the synergistic influence weights of each parameter on pollution diffusion. For example, the model might output "noise contributes 40% to PM2.5 diffusion" or "wind speed contributes 60% to sewage leakage diffusion." These weights are used to quantify the nonlinear correlation between parameters: when the high-frequency energy overlap between noise and vibration is high, the model enhances the synergistic weights of both in locating pollution sources; while for parameters with low correlation (such as light intensity), the weights are significantly suppressed.

[0093] Furthermore, by combining parameters such as synergistic influence weights, wastewater flow rate, and wind speed, a spatiotemporal graph convolutional network (ST-GCN) is used to predict pollution diffusion paths and concentration gradient changes. For example, if the wind direction points towards residential areas and the synergistic weights show a high correlation between PM2.5 and traffic flow, the model will predict an increase in PM2.5 concentration in that area and generate a diffusion path map to guide the precise allocation of pollution control resources.

[0094] Through spatial and temporal attention mechanisms, it can identify implicit relationships between complex environmental parameters (such as the nonlinear dependence of wind speed on PM2.5 diffusion) rather than relying on predefined rules; it suppresses interference from irrelevant parameters in low-noise scenarios, improving the model's sensitivity to real pollution sources (such as accurately distinguishing between equipment vibration and environmental noise); the output weights provide interpretable evidence for pollution control, such as "70% of pollution in a certain area is caused by abnormal equipment vibration", guiding precise control; furthermore, the high resolution of 5-minute time slices and 50m×50m grids, combined with ST-GCN's spatiotemporal modeling capabilities, can quickly predict the diffusion path of short-term pollution events (such as sudden leaks), shortening the control response time.

[0095] S300 constructs an adaptive prediction model that includes physical and behavioral driving channels based on the weights of synergistic effects. The physical channel generates baseline prediction results based on noise attenuation equations and CFD fluid simulations. The behavioral channel generates prediction correction factors by analyzing the temporal correlation between traffic flow and social data through an LSTM network. When the deviation between the physical prediction and the measured data exceeds the data deviation threshold for multiple consecutive periods, the weight of the physical channel is reduced and the weight of the behavioral channel is increased. At the same time, a digital twin sandbox is launched in GIS to simulate pollution source migration scenarios and generate virtual training data to optimize the adaptive prediction model.

[0096] First, a noise attenuation equation for the physical driving channel is constructed. Based on acoustic principles, a mathematical model is used to describe the propagation and attenuation process of noise in the air, taking into account factors such as distance, obstacles, and reflective surfaces. Next, CFD fluid simulation is performed, requiring computational fluid dynamics (CFD) to simulate the diffusion path of PM2.5 particles in the air, considering the influence of factors such as wind speed and temperature gradient on pollutant diffusion.

[0097] By providing baseline predictions of pollution diffusion, a reference basis can be provided for subsequent revisions, and the accuracy based on physical laws helps to understand how environmental variables affect pollutant diffusion.

[0098] Next, a behavior-driven channel is constructed. Long Short-Term Memory (LSTM) networks are a special type of recurrent neural network (RNN) that excels at processing time-series data. In this embodiment, they are used to analyze the changing trends and correlations between traffic flow and the frequency of social media keyword mentions over time. The generated prediction correction factor is based on time-series patterns derived from historical data analysis, adjusting the results of the physical prediction model to reflect the actual impact of human activities on pollution diffusion.

[0099] Considering the impact of social behavior on pollution diffusion, such as increased PM2.5 concentrations during peak traffic hours, incorporating social behavior data can improve the predictive accuracy of the model, especially in such application scenarios in urban environments.

[0100] When the deviation between the physical prediction and the measured data exceeds a data deviation threshold for multiple consecutive periods (in this embodiment, the data deviation threshold is 15%, but it can be modified; these multiple periods can be three or more periods), the weight of the physical channel is reduced while the weight of the behavioral channel is increased. For example, the system automatically reduces the weight of the physical channel by 0.3 and correspondingly increases the weight of the behavioral channel to 0.7. This adjustment mechanism aims to optimize the model's prediction accuracy and reduce errors caused by the physical model's assumptions not fully conforming to reality. This improves the model's adaptability and flexibility, especially in complex and ever-changing urban environments, and ensures that the prediction results are closer to reality, enhancing the effectiveness of decision support.

[0101] When initiating the digital twin sandbox simulation, the simulation is launched within the GIS platform. This involves creating a virtual environment that simulates the migration of pollution sources in the real world. Then, by simulating pollution diffusion processes under different scenarios, a large amount of virtual data is collected to further optimize the adaptive prediction model. This provides a safe and controllable environment for testing and refining the model.

[0102] It should be noted here that the methods used in launching the digital twin sandbox to simulate pollution source migration scenarios also include:

[0103] S310 constructs a multiphysics coupled simulation model, which includes an air pollution diffusion module and a ground vibration diffusion module. The air pollution diffusion module simulates the air propagation path of PM2.5 and noise based on CFD fluid simulation, while the ground vibration diffusion module simulates the vibration energy attenuation in high-risk areas based on the vibration wave equation.

[0104] The air pollution dispersion module utilizes computational fluid dynamics (CFD) technology to accurately simulate the propagation paths of PM2.5 particles and noise in the atmosphere. This module considers the impact of meteorological factors such as wind speed, temperature gradient, and humidity on pollutant dispersion. For complex environment modeling, it incorporates topographic features such as buildings and terrain undulations, as well as factors like vegetation cover, on airflow and pollutant transport paths. Fine-grained mesh generation and boundary condition settings ensure that the simulation results are as close to reality as possible. It's important to note that the simulation not only simulates the dispersion of single pollutants but also analyzes the interactions between different pollutants, such as chemical reactions or synergistic effects, providing a more comprehensive environmental impact assessment.

[0105] The ground vibration propagation module is based on the vibration wave equation in elasticity to simulate how vibration energy in high-risk areas decays over time and space, taking into account the influence of geological structure, soil type, and other factors on vibration propagation. To distinguish between high-frequency vibrations (such as vibrations generated by mechanical equipment) and low-frequency vibrations (such as seismic waves), separate models and analyses are performed to accurately capture the energy distribution and propagation characteristics of different types of vibrations.

[0106] In other words, by combining noise and vibration data, potential environmental pollution sources within a region and their impact on the surrounding environment can be assessed, providing a scientific basis for accurately locating pollution sources. This provides a simulation platform that integrates multiple physical phenomena, capable of simultaneously considering the effects of air pollution and ground vibration. This is crucial for assessing the comprehensive impact of pollution sources within a specific region, helping to formulate more effective environmental protection strategies, thereby enhancing the understanding of pollutant diffusion mechanisms under complex environmental conditions and improving the accuracy of predictions.

[0107] S320 introduces a spatiotemporal decay factor into the GIS grid, which is calculated based on an exponential decay function of distance and time.

[0108] The spatiotemporal attenuation factor is calculated based on an exponential decay function of distance and time. As distance increases or time progresses, pollutant concentration gradually decreases. Specifically, for each GIS grid cell, the pollutant concentration is adjusted according to the distance from the pollution source and the length of time since the pollution occurred. Appropriate attenuation coefficients are determined based on historical monitoring data and experimental research results; these coefficients reflect the diffusion rate and attenuation patterns of different pollutants under different environmental conditions.

[0109] Applying the spatiotemporal decay factor to the GIS grid system enables the pollutant concentration value of each grid cell to reflect the actual diffusion situation. It not only considers the direct diffusion path, but also the influence of factors such as obstacles and terrain on the pollutant transmission path. This allows for a more accurate simulation of the changes in pollutants over time and space, especially the long-distance transmission effect, which helps to identify potential pollution hotspots and provides scientific decision support for environmental management departments.

[0110] S330: When the deviation between the simulated path and the measured pollution diffusion path exceeds the path deviation threshold, the sandbox is triggered to adaptively adjust the migration probability distribution parameters. Specifically, this includes increasing the contribution weight of wind speed to pollutant diffusion by 20% and recalculating the nonlinear relationship between precipitation intensity and pollutant deposition rate based on historical data.

[0111] In this embodiment of the application, the path deviation threshold is 10%. When the deviation between the simulated path and the measured pollution diffusion path exceeds 10%, the system automatically starts an adaptive adjustment mechanism. First, the system will identify the main cause of the deviation, which may be that some key environmental variables (such as wind speed and precipitation intensity) fail to accurately reflect the actual situation.

[0112] For example, if wind speed is found to be one of the main factors causing the deviation, the system will automatically increase the weight of wind speed's contribution to pollutant diffusion by 20% to better match the actual observation data.

[0113] Based on the latest measured data, the nonlinear relationship between precipitation intensity and pollutant deposition rate is recalculated. This process may involve the application of machine learning algorithms to uncover complex patterns hidden in large amounts of data. After each adjustment, the system records the performance indicators before and after the adjustment and performs comparative analysis. In this way, the model parameters are continuously optimized to better adapt to constantly changing environmental conditions.

[0114] The S400, based on an optimized adaptive prediction model, reverses the pollution diffusion path in areas exceeding standards and locates target equipment by combining sensor data; it retrieves the equipment feature matching model, performs wavelet packet decomposition on noise and vibration signals, and matches vibration features in the equipment feature fingerprint database. When a match is successful, it generates control commands and sends them to IoT terminal devices to execute operations such as shutting down the excessive pump or activating a noise reduction barrier.

[0115] First, an optimized adaptive prediction model, combined with real-time monitoring data from the GIS grid (such as PM2.5 concentration and noise intensity), quickly identifies areas where pollutant concentrations exceed standards. Then, the pollution diffusion path of these areas is reverse-engineered. This can be achieved by utilizing both physical driving channels (such as noise attenuation equations and CFD fluid simulations) and behavioral driving channels (such as the influence of traffic flow and social media data) to deduce the diffusion path of pollutants from their source to the affected areas. Specifically, CFD simulations are used to reverse-analyze the impact of factors such as wind speed and topography on pollutant transport; and a spatiotemporal attention network is used to analyze the correlation between pollution sources and affected areas in historical data, further validating the accuracy of the reverse-engineered path.

[0116] It should be noted that, since pollution spread may be affected by a variety of factors (such as changes in wind direction, precipitation deposition, etc.), the system will generate multiple possible spread paths and sort them by probability based on sensor data and model prediction results, and finally determine the most likely source of pollution.

[0117] Through the above steps, potential pollution sources in areas with excessive pollution can be quickly identified, providing direction for subsequent precise governance. Furthermore, the process of reverse-engineering the path not only relies on physical laws but also incorporates social behavioral data, improving the accuracy and reliability of source tracing. In addition, the multi-path probability assessment mechanism ensures that the optimal solution can be found even in complex environments, avoiding errors that may be caused by single-path inference.

[0118] Then, data from noise sensors, vibration sensors, and other environmental sensors (such as PM2.5 sensors) within the excessive area are fused. High-risk co-polluting areas are identified through dynamic overlap calculations (e.g., high-frequency similarity > 70% and energy percentage difference < 15%). Within these high-risk areas, the scope is further narrowed down by combining the noise and vibration spectrum characteristics collected by the sensors, pinpointing specific equipment or facilities (such as pumps, generators, etc.). For example, a vibration sensor detects a specific frequency vibration signal generated by a piece of equipment during operation; a noise sensor captures high-frequency noise emitted by the same equipment. By analyzing the vibration intensity and noise spectrum distribution of the equipment, it is determined whether it is operating abnormally (e.g., excessive speed or bearing wear), thus confirming whether it is a major source of pollution.

[0119] Furthermore, in the implementation of the device feature matching model, the method also includes:

[0120] S410 performs three-level wavelet packet decomposition on noise and vibration signals to extract vibration features in the high-frequency, mid-frequency, and low-frequency bands.

[0121] First, the noise and vibration signals of the target device are decomposed into three levels of wavelet packets, with each level further dividing the signal into finer frequency bands. After the three-level decomposition, the original signal is divided into eight sub-frequency bands, corresponding to the high-frequency band, mid-frequency band, and low-frequency band, respectively.

[0122] High frequency band: typically includes sharp noise and vibration signals generated by equipment failure or abnormal operation, such as vibration caused by bearing wear or imbalance; Mid frequency band: reflects the main vibration frequency under normal operating conditions of the equipment, such as vibration related to motor speed; Low frequency band: captures environmental noise and background vibration, and may also include some low-frequency mechanical fault signals.

[0123] Wavelet packet decomposition can provide information in both the time and frequency domains, making it particularly suitable for the analysis of non-stationary signals and helping to extract subtle features of equipment operating status. Through three-level decomposition, more detailed spectral information can be obtained, providing a rich data foundation for subsequent feature matching.

[0124] Within each sub-band, characteristic parameters such as energy distribution and dominant frequency peak of the vibration signal are calculated. For example, energy distribution calculates the total energy in each band and its proportion of the total signal energy; dominant frequency peak identifies the main vibration frequency and its corresponding amplitude in each band; other features, such as spectral shape and spectral slope, can also be extracted to comprehensively describe the characteristics of the vibration signal.

[0125] By extracting feature parameters from different frequency bands, it is helpful to distinguish different operating states of equipment, especially abnormal operating states; these feature parameters will serve as key inputs in the subsequent matching process to ensure the accuracy and reliability of the matching results.

[0126] S420 performs weighted voting matching between the decomposed vibration characteristics and the device fingerprint database.

[0127] The extracted vibration features are compared with feature vectors in the device fingerprint database to calculate similarity. Cosine similarity or other distance metrics are used to calculate similarity scores for high-frequency, mid-frequency, and low-frequency bands respectively. Different weights are assigned based on the importance of each frequency band; specifically, the similarity weight for high-frequency bands is 0.5, for mid-frequency bands it is 0.3, and for low-frequency bands it is 0.2. The similarity scores for each frequency band are multiplied by their corresponding weights, and then summed to obtain the total similarity score.

[0128] The weighted voting mechanism takes into account the differences in importance of different frequency bands in equipment fault diagnosis, improving the accuracy of matching; and by setting reasonable weights, it can better balance the sensitivity of high-frequency bands and the stability of low-frequency bands, ensuring the robustness of matching results.

[0129] S430: When the similarity of all three frequency bands exceeds the similarity threshold and the weighted total similarity is greater than or equal to the weighted total similarity threshold, a successful match is determined. When the matching failure rate continuously exceeds the matching failure rate threshold, the fingerprint database incremental update mechanism is triggered, which automatically retrieves vibration data from a preset time period and updates the fingerprint database feature vector using an online learning algorithm.

[0130] It should be noted that in this embodiment of the application, the similarity threshold is 80%, the weighted total similarity threshold is 85%, the matching failure rate threshold is 20%, and the preset similar time period is the most recent 30 days.

[0131] In other words, a match is considered successful when the similarity across all three frequency bands exceeds 80%, and the weighted total similarity is ≥85%. Specifically: high-frequency band similarity >80%, mid-frequency band similarity >80%, low-frequency band similarity >80%, and weighted total similarity ≥85%.

[0132] Strict matching conditions are set to ensure that a match is only considered successful when all frequency bands are highly similar, thereby avoiding false positives. By comprehensively considering the similarity of each frequency band, the reliability of the matching results is improved, and the probability of false positives or false negatives is reduced.

[0133] The system continuously monitors the success rate of device feature matching. If the matching failure rate exceeds 20% for an extended period, it triggers an incremental update mechanism for the fingerprint database. Simultaneously, it automatically retrieves vibration data from a preset time period, which includes the actual vibration of the equipment under different operating conditions.

[0134] By employing online learning algorithms (such as adaptive augmentation and incremental PCA) to process newly collected data, new feature vectors are generated and added to the existing device fingerprint database. By continuously updating the feature vectors in the fingerprint database, the model can adapt to changes in device operating status, thereby improving the accuracy of long-term prediction and matching.

[0135] The fingerprint database incremental update mechanism can dynamically adjust the model to adapt to changes in device operating status and extend the model's validity period. In addition, the application of online learning algorithms enables the fingerprint database to be gradually optimized without interrupting system operation, improving the system's flexibility and adaptability.

[0136] The S500, based on the execution results of IoT terminals and real-time monitoring data, opens an AR augmented reality verification interface, retrieves predicted and measured data from the GIS platform, generates a deviation heatmap based on the predicted and measured data, and obtains the corresponding virtual-real data deviation. When the virtual-real data deviation exceeds 20%, it triggers the federated learning framework to aggregate the local gradients of edge nodes to update the global prediction model, retains historical versions to build a model evolution tree for anomaly backtracking analysis, and generates a corresponding environmental compliance report. The environmental compliance report includes details of parameters exceeding standards, regulatory correlation analysis, control effect assessment, actionable recommendations, and blockchain-stored evidence data.

[0137] Based on the Internet of Things (IoT) platform, an augmented reality (AR) verification interface was developed and opened. This interface allows users to view and interact with pollution diffusion prediction results and actual measurement data through mobile devices or dedicated hardware. The execution results from IoT terminals (such as changes in noise levels after shutting down excessive pumps) are integrated with real-time monitoring data provided by the GIS platform into the AR environment, enabling users to intuitively see the comparison between predictions and actual conditions.

[0138] Pollution diffusion path prediction data (including pollutant concentration distribution, wind direction influence, etc.) are extracted from the GIS platform and compared with data from real-time monitoring sensors (such as PM2.5 concentration, noise intensity, etc.). Then, the difference between the predicted value and the measured value is calculated to generate a deviation heat map. This step not only considers the numerical difference, but also includes the consistency analysis of spatial distribution. The deviation heat map helps to identify areas with large prediction errors, which facilitates targeted optimization.

[0139] In this process, GIS tools are used to display the deviation between predicted and measured data in the form of a heat map. Different colors represent different degrees of deviation, with red indicating areas of high deviation and green indicating areas of low deviation. In addition to the visual heat map, it is also necessary to quantify these deviations, that is, to obtain the specific percentage of the difference between the predicted and measured data.

[0140] If the deviation heatmap shows that the deviation between virtual and real data in certain areas exceeds 20%, the federated learning framework is triggered. Each edge node (such as a sensor network distributed in different geographical locations) trains a model based on its local data and uploads the trained gradients to the central server. The central server is responsible for aggregating the gradient information of all edge nodes and updating the global prediction model. It is necessary to retain historical versions before each update and construct a model evolution tree for future anomaly backtracking analysis.

[0141] Federated learning technology can improve model accuracy without sharing the original data, protecting privacy while increasing prediction accuracy. Furthermore, the model evolution tree records the historical trajectory of model development, which helps to understand and analyze the changing trends of model performance.

[0142] Based on the latest forecasts, measured data, and deviation analysis, a detailed environmental compliance report is generated. The report includes: details of exceeding parameters, listing the specific types of pollutants exceeding standards and their concentrations; regulatory correlation analysis, comparing current environmental laws and regulations to identify which provisions are violated; control effectiveness assessment, evaluating the actual effect of implemented control measures on improving the pollution situation; actionable recommendations, proposing specific improvement suggestions for the current situation; and blockchain-based data storage, using blockchain technology to store key data in the report, ensuring data immutability and enhancing the report's credibility.

[0143] In this embodiment of the application, the method further includes, during the process of reverse-engineering the pollution diffusion path of the area exceeding the standard based on the optimized adaptive prediction model:

[0144] First, a multi-scale spatiotemporal inversion model is constructed, dividing the pollution diffusion path into microscale, mesoscale, and macroscale.

[0145] The process involves using a 50m×50m sensor grid to precisely locate the coordinates of pollution source equipment. This step relies on a high-resolution sensor network capable of capturing subtle changes within a local area. At the mesoscale, combined with environmental data such as traffic flow and wind speed, the specific path of pollutants from the pollution source to the area exceeding the standard is deduced. This scale considers environmental factors over a wider range, such as the impact of wind direction and topography on pollutant diffusion. At the macroscale, by analyzing data from historical pollution events, the spatiotemporal distribution patterns of pollution sources within the region are studied. This long-term data analysis helps identify potential pollution hotspots and trends.

[0146] Then, when locating the target device, a dynamic confidence assessment mechanism is adopted to calculate the device position based on the confidence level of the sensor data. The confidence level is the matching degree between noise intensity and vibration energy. When the confidence level is lower than 80%, the collaborative inversion of adjacent grids is triggered to improve the positioning accuracy through multi-grid data fusion.

[0147] The system assigns a confidence score to each sensor grid based on the quality of the sensor data (such as the matching degree between noise intensity and vibration energy). A higher confidence score indicates more reliable data, which can be used for more accurate positioning. When the confidence score of a grid falls below 80%, the system triggers data fusion from neighboring grids to improve the accuracy of the target device's location. For example, data from multiple neighboring grids can be combined to correct the initial estimate. This dynamic confidence assessment mechanism ensures that only high-quality data is used for the final positioning decision, reducing the possibility of misjudgment. Through collaborative inversion, it can compensate for the insufficiency of data from a single grid, thereby improving positioning accuracy.

[0148] Finally, when generating control commands, a graded response strategy is introduced. When the parameter exceeding the standard is noise and the duration exceeds the preset duration threshold, the noise reduction barrier is activated first. When the parameter exceeding the standard is sewage turbidity and the concentration exceeds the sewage concentration trigger value, the pump is shut down and the surrounding sewage treatment facilities are linked to start the purification process.

[0149] The duration threshold is 1 hour, and the wastewater concentration trigger value is twice the set threshold. When noise is detected to exceed the standard and the duration exceeds 1 hour, a noise reduction barrier is activated first. This method can effectively reduce noise levels and reduce the impact on the surrounding environment without interrupting equipment operation. If wastewater turbidity is detected to exceed the standard and the concentration exceeds twice the set threshold, the relevant pumps are immediately shut down, and the surrounding wastewater treatment facilities are linked to start the purification process. This step aims to quickly stop the spread of pollution and take measures to purify the affected area.

[0150] The tiered response strategy adopts targeted measures based on different types of pollution, which improves the efficiency and effectiveness of the response. It not only ensures the ability to respond quickly in emergencies, but also takes into account the characteristics of different types of pollution, thus achieving precise governance.

[0151] Based on the above steps, the entire solution can not only quickly and accurately identify pollution sources and locate target equipment, but also take the most appropriate control measures according to the specific situation, thereby effectively managing and controlling environmental pollution problems.

[0152] This application discloses an automatic prediction system for multiple environmental parameters based on GIS maps, referring to... Figure 2 ,include:

[0153] The multidimensional spatiotemporal matrix generation module 001 collects real-time data on noise intensity, vibration spectrum, sewage turbidity, light intensity, and PM2.5 concentration in the target area. This data is transmitted to the GIS platform via a distributed sensor network. Simultaneously, it acquires real-time changes in traffic flow, spatiotemporal distribution of social media keywords, and short-term fluctuations in meteorological parameters in the target area. Based on GIS spatial gridding technology, it aligns environmental parameters and dynamic behavior data in a preset format to generate a multidimensional spatiotemporal matrix that integrates physical parameters and social behavior. The physical parameters include noise intensity, vibration spectrum, sewage turbidity, light intensity, and PM2.5 concentration data, while the social behavior includes real-time changes in traffic flow, spatiotemporal distribution of social media keywords, and short-term fluctuations in meteorological parameters.

[0154] The collaborative influence weight output module 002 extracts the energy distribution characteristics of the main frequency band of noise and the spectral energy distribution characteristics of vibration sensors based on a multi-dimensional spatiotemporal matrix. It calculates the energy overlap between the energy distribution characteristics and the spectral energy distribution characteristics. When the energy overlap exceeds the overlap threshold, it marks the area as a high-risk collaborative pollution area. It obtains the sewage flow rate, real-time wind speed, precipitation data and collaborative pollution characteristics corresponding to the high-risk collaborative pollution area. It combines the sewage flow rate, real-time wind speed, precipitation data and collaborative pollution characteristics to predict the pollution diffusion path and concentration gradient changes. It uses a spatiotemporal graph attention network to dynamically learn the nonlinear correlation between parameters and uses it to output the collaborative influence weight.

[0155] The adaptive prediction model optimization module 003 constructs an adaptive prediction model that includes physical driving channels and behavioral driving channels based on the weights of synergistic effects. The physical channel generates baseline prediction results based on noise attenuation equations and CFD fluid simulation. The behavioral channel generates prediction correction factors by analyzing the temporal correlation between traffic flow and social data through an LSTM network. When the deviation between physical prediction and measured data exceeds the data deviation threshold for multiple consecutive periods, the weight of the physical channel is reduced and the weight of the behavioral channel is increased. At the same time, a digital twin sandbox is launched in GIS to simulate pollution source migration scenarios and generate virtual training data for optimizing the adaptive prediction model.

[0156] The matching module 004, based on the optimized adaptive prediction model, reverses the pollution diffusion path in the area exceeding the standard and locates the target equipment by combining sensor data; it calls up the equipment feature matching model, performs wavelet packet decomposition on the noise and vibration signals, and uses it to match the vibration features in the equipment feature fingerprint database. When the match is successful, it generates a control command and sends it to the IoT terminal device to execute the operation of shutting down the pump that exceeds the standard or starting the noise reduction barrier.

[0157] The environmental compliance report generation module 005, based on the execution results of IoT terminals and real-time monitoring data, opens an AR augmented reality verification interface, retrieves predicted and measured data from the GIS platform, generates a deviation heatmap based on the predicted and measured data, and obtains the corresponding virtual-real data deviation. When the virtual-real data deviation exceeds 20%, it triggers the federated learning framework to aggregate the local gradients of edge nodes to update the global prediction model, retains historical versions to build a model evolution tree for anomaly backtracking analysis; and generates the corresponding environmental compliance report, which includes details of parameters exceeding standards, regulatory correlation analysis, control effect assessment, actionable recommendations, and blockchain-stored evidence data.

[0158] This application also discloses an automatic prediction system for multiple environmental parameters based on GIS maps, including a processor, in which a program of any one of the above-mentioned automatic prediction methods for multiple environmental parameters based on GIS maps is running.

[0159] This application also discloses a storage medium storing a program for the automatic prediction method of multiple environmental parameters based on GIS maps as described in any one of the above embodiments.

[0160] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A GIS map-based multi-class environmental parameter automatic prediction method, characterized by, The method comprises the following steps: Real-time acquisition of noise intensity, vibration spectrum, sewage turbidity, light intensity and PM2.5 concentration data of the target area, transmission of the data to the GIS platform through a distributed sensor network, synchronous acquisition of real-time traffic flow changes, social media environmental keyword spatiotemporal distribution and short-term fluctuation data of meteorological parameters of the target area, spatiotemporal alignment of environmental parameters and dynamic behavior data according to a preset format based on GIS spatial gridding technology, generation of a multi-dimensional spatiotemporal matrix integrating physical parameters and social behavior, wherein the physical parameters include noise intensity, vibration spectrum, sewage turbidity, light intensity and PM2.5 concentration data, and the social behavior includes real-time traffic flow changes, social media environmental keyword spatiotemporal distribution and short-term fluctuation data of meteorological parameters; Based on the multi-dimensional spatiotemporal matrix, the energy distribution characteristics of the noise main frequency band and the frequency spectrum energy distribution characteristics of the vibration sensor are extracted, the energy overlap degree of the energy distribution characteristics and the frequency spectrum energy distribution characteristics is calculated, and when the energy overlap degree exceeds the overlap degree threshold, it is marked as a high-risk synergistic pollution area; the sewage flow, real-time wind speed, precipitation data and synergistic pollution characteristics corresponding to the high-risk synergistic pollution area are obtained, and the pollution diffusion path and concentration gradient change are predicted by combining the sewage flow, real-time wind speed, precipitation data and synergistic pollution characteristics, and the spatiotemporal graph attention network is used to dynamically learn the non-linear correlation between parameters and output the synergistic influence weight; According to the synergistic influence weight, an adaptive prediction model including a physical driving channel and a behavior driving channel is constructed, the physical channel generates a baseline prediction result based on a noise attenuation equation and a CFD fluid simulation, the behavior channel generates a prediction correction factor by analyzing the time sequence correlation of traffic flow and social data through an LSTM network, when the deviation between the physical prediction and the measured data exceeds the data deviation threshold for continuous multiple periods, the weight of the physical channel is reduced and the weight of the behavior channel is increased, and at the same time, a digital twin sandbox is started in GIS to simulate pollution source migration scenarios to generate virtual training data and optimize the adaptive prediction model; Based on the optimized adaptive prediction model, the pollution diffusion path is backstepped for the over-standard area, and the target equipment is located combined with the sensor data; the device feature matching model is called to perform wavelet packet decomposition on the noise and vibration signals, and match the vibration features in the device feature fingerprint library, when the matching is successful, a control command is generated and sent to the Internet of Things terminal equipment, and the operation of shutting down the over-standard pump or starting the noise reduction barrier is performed; According to the execution result of the Internet of Things terminal and the real-time monitoring data, an AR augmented reality verification interface is opened, the prediction data and the measured data of the GIS platform are called, a deviation heat map is generated based on the prediction data and the measured data, and the corresponding virtual-real data deviation is obtained, when the virtual-real data deviation exceeds 20%, a federal learning framework is triggered to aggregate the local gradient of the edge node to update the global prediction model, a historical version is reserved to construct a model evolution tree for abnormal backtracking analysis; a corresponding environmental protection compliance report is generated, which includes over-standard parameter details, regulation correlation analysis, control effect evaluation, executable suggestions and blockchain notarization data. 2.The GIS map-based multi-class environmental parameter automatic prediction method of claim 1, wherein, The method further comprises the following steps: The energy distribution characteristics of the main frequency band of the noise are extracted by short-time Fourier transform, and the energy distribution characteristics include the main frequency energy peak value, the frequency band energy proportion, and the energy distribution curve slope; The frequency spectrum energy distribution characteristics of the vibration sensor are extracted by vibration spectrum analysis, and the frequency spectrum energy distribution characteristics include the vibration energy peak frequency, the frequency band energy fluctuation amplitude, and the frequency spectrum energy density; The dynamic overlap degree of the noise main frequency band energy distribution and the vibration frequency spectrum energy distribution is calculated, including: the frequency energy distribution curve of the energy distribution characteristics is standardized with the frequency energy distribution curve of the frequency spectrum energy distribution characteristics, and the weighted cosine similarity calculation formula is used, wherein the high frequency band weight is higher than the low frequency band weight; When the dynamic overlap degree exceeds the overlap degree threshold and the high frequency band energy proportion difference is less than the high frequency band energy proportion difference threshold, the high-risk collaborative pollution area is marked. 3.The GIS map-based multi-class environmental parameter automatic prediction method of claim 2, wherein, In the process of dynamically learning the non-linear correlation between parameters by using the spatio-temporal graph attention network, the method further includes: A node feature matrix of the spatio-temporal graph is constructed, and the node feature matrix includes environmental parameters, dynamic behavior parameters, and spatial correlation parameters, wherein the environmental parameters include noise intensity, vibration frequency spectrum energy peak value, sewage turbidity, PM2.5 concentration, and light intensity, the dynamic behavior parameters include traffic flow density, social media keyword mention frequency, wind speed, and precipitation intensity, and the spatial correlation parameters include pollution diffusion path prediction values and concentration gradient change values of adjacent grids; A double-channel attention mechanism is introduced in the graph attention network, wherein the spatial channel calculates the spatial attention weight based on the grid adjacency relationship, and the time channel calculates the time attention weight based on the time series autocorrelation; When the noise intensity is lower than the noise lower threshold, the environmental parameter weight of the corresponding grid is dynamically reduced to 30% of the original weight. 4.The GIS map-based multi-class environmental parameter automatic prediction method of claim 3, wherein, In the process of starting the digital twin sandbox to simulate the migration scenario of the pollution source, the method further includes: A multi-physics field coupling simulation model is constructed, and the multi-physics field coupling simulation model includes an air pollution diffusion module and a ground vibration diffusion module, the air pollution diffusion module simulates the air propagation path of PM2.5 and noise based on a CFD fluid simulation, and the ground vibration diffusion module simulates the vibration energy attenuation of the high-risk area based on a vibration wave equation; A space-time attenuation factor is introduced in the GIS grid, and the space-time attenuation factor is calculated based on an exponential attenuation function of distance and time; When the simulation path deviates from the measured pollution diffusion path by more than the path deviation threshold, the sandbox adaptively adjusts the migration probability distribution parameters, specifically including: increasing the contribution weight of wind speed to pollution diffusion by 20%, and recalculating the non-linear relationship between precipitation intensity and pollution settling rate according to historical data. 5.The GIS map-based multi-class environmental parameter automatic prediction method of claim 4, wherein, In the implementation process of the equipment feature matching model, the method further includes: The noise and vibration signals are subjected to three-layer wavelet packet decomposition, and vibration features of high, medium, and low frequency bands are extracted; The decomposed vibration features are weighted and voted to match with the equipment fingerprint library; When the three-frequency band similarities all exceed the similarity threshold and the weighted total similarity is greater than or equal to the weighted total similarity threshold, it is determined that the matching is successful; when the matching failure rate continuously exceeds the matching failure rate threshold, a fingerprint library incremental update mechanism is triggered, preset vibration data in a similar time period is automatically called, and an online learning algorithm is used to update the fingerprint library feature vector. 6.The GIS map-based multi-class environmental parameter automatic prediction method of claim 5, wherein, In the process of reversely deducing the pollution diffusion path of the over-standard area based on the optimized adaptive prediction model, the method further comprises: A multi-scale space-time inversion model is constructed to divide the pollution diffusion path into micro-scale, meso-scale and macro-scale, wherein the micro-scale locates the precise coordinates of the pollution source equipment based on the sensor grid data, the meso-scale reversely deduces the propagation path of the pollutant from the pollution source to the over-standard area in combination with the traffic flow and wind speed data, and the macro-scale analyzes the space-time distribution law of the regional pollution source through historical pollution event data; In positioning the target equipment, a dynamic confidence evaluation mechanism is used to calculate the equipment position by weighting the confidence of the sensor data, wherein the confidence is the matching degree of noise intensity and vibration energy; when the confidence is lower than 80%, the cooperative inversion of adjacent grids is triggered to improve the positioning accuracy through multi-grid data fusion; In generating the control instruction, a hierarchical response strategy is introduced, the noise reduction barrier is preferentially started when the over-standard parameter is noise and the duration exceeds the preset duration threshold, and the pump is closed and the surrounding sewage treatment facilities are started to start the purification process when the over-standard parameter is sewage turbidity and the concentration exceeds the sewage concentration trigger value.

7. A GIS map-based multi-class environmental parameter automatic prediction system, characterized by, It comprises: A multi-dimensional space-time matrix generation module, which collects noise intensity, vibration spectrum, sewage turbidity, light intensity and PM2.5 concentration data of the target area in real time, transmits them to the GIS platform through a distributed sensor network, synchronously acquires traffic flow real-time changes, social media environmental keyword space-time distribution and meteorological parameter short-term fluctuation data of the target area, and aligns the environmental parameters and dynamic behavior data in space-time according to a preset format based on GIS spatial gridding technology, to generate a multi-dimensional space-time matrix that fuses physical parameters and social behaviors, wherein the physical parameters include noise intensity, vibration spectrum, sewage turbidity, light intensity and PM2.5 concentration data, and the social behaviors include traffic flow real-time changes, social media environmental keyword space-time distribution and meteorological parameter short-term fluctuation data; A cooperative influence weight output module, which extracts the energy distribution characteristics of the noise main frequency band and the frequency spectrum energy distribution characteristics of the vibration sensor based on the multi-dimensional space-time matrix, calculates the energy overlap degree of the energy distribution characteristics and the frequency spectrum energy distribution characteristics, marks the high-risk cooperative pollution area when the energy overlap degree exceeds the overlap degree threshold, acquires the sewage flow, real-time wind speed, precipitation data and cooperative pollution characteristics corresponding to the high-risk cooperative pollution area, predicts the pollution diffusion path and concentration gradient change in combination with the sewage flow, real-time wind speed, precipitation data and cooperative pollution characteristics, dynamically learns the non-linear correlation between parameters by using a space-time graph attention network, and outputs the cooperative influence weight. An adaptive prediction model optimization module constructs an adaptive prediction model containing a physical driving channel and a behavior driving channel according to the synergistic influence weight, the physical channel generates a baseline prediction result based on a noise attenuation equation and a CFD fluid simulation, and the behavior channel generates a prediction correction factor by analyzing the time series correlation between traffic flow and social data through an LSTM network; when the deviation between the physical prediction and the measured data exceeds the data deviation threshold for multiple consecutive periods, the weight of the physical channel is reduced and the weight of the behavior channel is increased, and at the same time, a digital twin sandbox is started in GIS to simulate the migration scenario of the pollution source to generate virtual training data for optimizing the adaptive prediction model; A matching module, based on the optimized adaptive prediction model, reverses the pollution diffusion path of the over-standard area and locates the target device combined with sensor data; a device feature matching model is called to perform wavelet packet decomposition on noise and vibration signals and match vibration features in the device feature fingerprint library; when the matching is successful, a control instruction is generated and sent to the IoT terminal device to execute the operation of shutting down the over-standard pump or starting the noise reduction barrier; An environmental compliance report generation module, according to the execution result of the IoT terminal and the real-time monitoring data, opens an AR augmented reality verification interface, calls the prediction data and measured data of the GIS platform, generates a deviation heat map based on the prediction data and the measured data, and obtains the corresponding virtual-real data deviation; when the virtual-real data deviation exceeds 20%, a federal learning framework is triggered to aggregate edge node local gradient updates to update the global prediction model, and historical versions are retained to build a model evolution tree for abnormal backtracking analysis; used to generate a corresponding environmental compliance report, the environmental compliance report includes over-standard parameter details, regulation correlation analysis, control effect evaluation, executable suggestions and blockchain notarization data.

8. A GIS map-based multi-class environmental parameter automatic prediction system, characterized by, A processor having a program of the GIS map-based multi-class environmental parameter automatic prediction method according to any one of claims 1-6 running therein.

9. A storage medium, characterized by A storage having a program of the GIS map-based multi-class environmental parameter automatic prediction method according to any one of claims 1-6.

Images