Multi-source sensing storage environment cooperative monitoring and early warning method

Abstract

The invention relates to the technical field of material physicochemical property monitoring of a multi-source sensing network, in particular to a multi-source sensing storage environment collaborative monitoring and early warning method, which comprises the following steps of: dividing grid units according to material storage types, binding sensor node coordinates, constructing a grid region reference model, and constructing a grid region reference model; integrating the volatilization characteristic parameters analyzed by the laboratory, historical monitoring data and an environment threshold value to generate a substance and environment relation matrix; the method comprises the following steps of: calling model parameters to invert a theoretical gas concentration value, and triggering finite element physical field simulation through residual analysis to generate a three-dimensional simulation field data set containing a temperature gradient and a diffusion path. And the dynamic correlation analysis precision and the early warning reliability of storage environment monitoring are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of material physical and chemical property monitoring using a multi-source sensing network, and in particular to a storage environment collaborative monitoring and early warning method using a multi-source sensing network. Background Art

[0002] The multi-source sensing collaborative monitoring and early warning method for warehouse environments integrates multiple heterogeneous sensor nodes, such as temperature, humidity, gas concentration, and video images, to build a distributed sensing network. It relies on data fusion algorithms to perform spatiotemporal correlation analysis of environmental parameters. The system uses a layered processing architecture that combines Kalman filtering with deep learning models to extract features and identify abnormal patterns in multimodal sensor data. It optimizes information complementarity between sensors through a dynamic weight allocation strategy, reducing the risk of misjudgment caused by noise interference and data loss from a single sensor. Based on preset environmental safety thresholds and historical operating patterns, the system establishes a multi-dimensional risk assessment model and utilizes edge computing nodes to analyze data streams in real time. When monitoring indicators exceed preset confidence intervals or exhibit temporal mutation characteristics, a hierarchical early warning mechanism is triggered and optimization and control recommendations are generated, effectively maintaining the stability of the warehouse environment and the safety of material storage.

[0003] In the process of heterogeneous multi-source sensor data fusion, existing methods suffer from collaborative deviations in measurement parameters caused by spatiotemporal calibration errors of sensor nodes and environmental interference, resulting in a decrease in the accuracy of dynamic correlation analysis of physical quantities such as gas concentration, temperature and humidity in multidimensional risk assessment models. In particular, when sensor nodes produce inconsistent sampling due to differences in deployment locations or sudden environmental changes, the failure of spatiotemporal alignment of cross-modal data will significantly weaken the confidence of abnormal pattern recognition. This problem directly affects the reliability of warehouse environmental safety assessment in scenarios involving monitoring of the physical and chemical properties of materials. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a multi-source sensing warehouse environment collaborative monitoring and early warning method, which is used to solve the technical problems in the existing technology that the reliability of monitoring the physical and chemical properties of stored materials is reduced due to the failure of cross-modal data spatiotemporal alignment and the decrease in dynamic correlation analysis accuracy caused by spatiotemporal calibration errors and environmental interference of multi-source heterogeneous sensor nodes.

[0005] In order to solve the above technical problems, the specific technical solutions of the present invention are as follows: The multi-source sensing storage environment collaborative monitoring and early warning method provided by the present invention includes: Collect historical environmental monitoring data and laboratory material chemical composition data of the storage area, divide the grid cells according to the material storage type, combine the binding relationship between sensor node coordinates and grid identifiers, receive and set environmental thresholds, and build a gridded regional benchmark model; Calling the material volatility characteristic parameters of the corresponding grid in the gridded regional benchmark model, inverting the theoretical gas concentration value through the material volatility rate equation, performing residual sequence calculation on the real-time collected sensor monitoring data and the theoretical gas concentration value, and triggering a physical field simulation based on the finite element method when the residual continuously exceeds a preset tolerance to generate a three-dimensional simulation field data set; The real-time collected sensor monitoring data and the three-dimensional simulation field data set are processed through a preset dual-channel processing process, wherein: The first processing channel performs Kalman filtering on the sensor monitoring data collected in real time for state prediction; The second processing channel extracts physical field features from the three-dimensional simulation field dataset through a graph neural network; Dynamically optimize the covariance matrix weights of the Kalman filter and the sparsity of the adjacency matrix of the graph neural network using a genetic algorithm, and output an environment state vector; The environmental state vector is split into a real-time data stream and a simulation data stream, which are respectively input into a time series convolutional network to extract short-term fluctuation characteristics and a long short-term memory network to capture the physical field evolution trend. The short-term fluctuation feature vector of the real-time data stream and the physical field evolution trend feature vector of the simulation data stream are weightedly fused through an attention mechanism, and coupled with the material thermodynamic parameters in the gridded regional benchmark model to generate a risk deviation index; Dynamically correct the material diffusion coefficient and thermodynamic critical value in the gridded regional benchmark model based on the risk deviation index, and update the environmental threshold; When the risk deviation index exceeds a preset level, the updated gridded regional benchmark model parameters are called to construct sensor data nodes, material state deduction nodes and equipment control nodes, and generate equipment control instructions.

[0006] Furthermore, in the multi-source sensing warehouse environment collaborative monitoring and early warning method of the present invention, the construction of the gridded regional benchmark model includes: Divide the storage area into grid units according to material storage types, and bind sensor node coordinates to grid identifiers; The correlation coefficients between the volatility characteristics of substances and temperature and humidity in the grid cells were extracted by laboratory gas chromatography-mass spectrometry analysis to generate a substance-environment relationship matrix. The substance-environment relationship matrix is ​​associated with historical monitoring data in a preset database according to grid coding to generate a gridded regional benchmark model library including sensor topology, environmental thresholds and substance volatility characteristic parameters.

[0007] Furthermore, in the multi-source sensing storage environment collaborative monitoring and early warning method of the present invention, the residual sequence calculation between the real-time collected sensor monitoring data and the theoretical gas concentration value includes: Calling the substance volatilization characteristic parameters of the corresponding grid in the gridded regional benchmark model library, and inverting the theoretical gas concentration value according to the substance volatilization rate equation; Calculating a residual sequence between the real-time collected sensor monitoring data and the theoretical gas concentration value, and triggering a physical field simulation based on the finite element method when the residual continues to exceed a preset tolerance; Combining historical data in the gridded regional benchmark model library, a three-dimensional simulation field data set including temperature gradients and gas diffusion paths is generated.

[0008] Furthermore, in the multi-source sensing warehouse environment collaborative monitoring and early warning method of the present invention, the processing of the real-time collected sensor monitoring data and the three-dimensional simulation field data set through a preset dual-channel processing process includes: Performing Kalman filtering on the sensor monitoring data collected in real time to predict the state and generate a real-time data state estimate; Extracting physical field features from the three-dimensional simulation field dataset through a graph neural network to generate a simulation data feature vector; Based on the real-time data state estimation value and the simulation data feature vector, calculating a similarity measure between the real-time data state estimation value and the simulation data feature vector in a latent space; Using the similarity metric as an optimization goal, driving a genetic algorithm to iteratively adjust the covariance matrix weights of the Kalman filter and the sparsity of the adjacency matrix of the graph neural network; According to the optimized covariance matrix weights and adjacency matrix sparsity, the real-time data state estimation value and the simulated data feature vector are fused to output an environment state vector with a confidence label.

[0009] Furthermore, in the multi-source sensing warehouse environment collaborative monitoring and early warning method of the present invention, generating the risk deviation index includes: Splitting the environmental state vector into a real-time data stream and a simulation data stream; Inputting the real-time data stream into a time series convolutional network to extract a first eigenvector representing short-term fluctuations of environmental parameters; Inputting the simulation data stream into a long short-term memory network to extract a second eigenvector representing an evolution trend of the physical field; Performing weighted fusion of the first feature vector and the second feature vector through an attention mechanism to generate a fused feature vector; The fused feature vector is coupled with the material thermodynamic parameters in the gridded regional benchmark model library for analysis, and a risk deviation index representing the degree to which the environmental state deviates from the material stability condition is output.

[0010] Furthermore, in the multi-source sensing warehouse environment collaborative monitoring and early warning method of the present invention, the calling of updated gridded regional benchmark model parameters, the construction of sensor data nodes, material state deduction nodes, and equipment control nodes, and the generation of equipment control instructions include: constructing sensor data nodes, material state deduction nodes, and equipment control nodes in a Bayesian network based on the updated gridded regional benchmark model parameters; Initializing the conditional probability table of the Bayesian network nodes using the Kalman filter covariance matrix weights and the graph neural network adjacency matrix sparsity optimized by the genetic algorithm; The generated three-dimensional simulation field data set and the updated environmental threshold are input into the Bayesian network, and the environmental state change trajectory under different control strategies is simulated through Monte Carlo sampling; screening a control instruction combination that causes the physical field parameters in the three-dimensional simulation field data set to converge with the updated environmental threshold; Device control instructions are generated according to the screening results, and dynamic thresholds in the gridded regional benchmark model parameter library are synchronously updated.

[0011] Furthermore, the multi-source sensing warehouse environment collaborative monitoring and early warning method of the present invention further includes: collecting environmental parameter feedback data after the execution of the equipment control instruction in real time through sensors, including temperature change rate, gas concentration gradient and humidity distribution state; Input the feedback data into the residual analysis module, recalculate the theoretical gas concentration value, and perform residual sequence comparison with the measured value in the feedback data; When the statistical characteristics of the residual sequence exceed the environmental threshold of the gridded regional benchmark model library, a data calibration process is triggered to dynamically adjust the material diffusion coefficient and thermodynamic critical value in the model library; Synchronizing the updated model parameters to the dual-channel processing and the physical field simulation module, reconstructing the data fusion strategy and the three-dimensional simulation field generation logic; Through the iterative adaptation of the updated model parameters and the equipment control instructions, a closed-loop feedback link of gridded regional benchmark model parameter optimization and storage environment equipment control is formed.

[0012] Furthermore, in the multi-source sensing warehouse environment collaborative monitoring and early warning method of the present invention, generating a three-dimensional simulation field data set includes: generating initial three-dimensional simulation field data including temperature gradients and gas diffusion paths based on the finite element physical field simulation results; Adding a grid code mark to the initial three-dimensional simulated field data, wherein the grid code is aligned with the grid unit identifier and the bound sensor node coordinates according to a spatial topological relationship; In the dual-channel processing, the material volatility characteristic parameters in the gridded regional benchmark model library are matched based on the grid code; Using the substance volatility characteristic parameters as constraints, calculating a similarity measure between the real-time data state estimate and the simulated data feature vector in a latent space; According to the similarity metric optimization result, the physical field parameters of the three-dimensional simulation field data set are synchronously corrected to generate a calibrated three-dimensional simulation field data set that is spatially and temporally aligned with the real-time monitoring data.

[0013] Furthermore, in the multi-source sensing warehouse environment collaborative monitoring and early warning method of the present invention, the dynamically correcting the material diffusion coefficient and thermodynamic critical value in the gridded area benchmark model based on the risk deviation index and updating the environmental threshold value includes: extracting material thermodynamic equation parameters associated with the current environmental state in the gridded area benchmark model library based on the generated risk deviation index; According to the value of the risk deviation index, the diffusion coefficient and the thermodynamic critical value in the thermodynamic equation of the substance are iteratively adjusted through a gradient descent algorithm; The updated diffusion coefficient and thermodynamic critical value are transmitted back to the physical field simulation module to reconstruct the boundary conditions of the gas diffusion equation in the process of generating the three-dimensional simulation field data set; a preset finite element physical field simulation method is called to recalculate the gas diffusion path based on the updated diffusion coefficient to generate a revised three-dimensional simulation field data set; The modified three-dimensional simulation field data set is input into the dual-channel processing process, the latent space similarity measure is recalculated, and the data fusion weight is optimized simultaneously.

[0014] Furthermore, in the multi-source sensing warehouse environment collaborative monitoring and early warning method described in the present invention, the method of dynamically optimizing the covariance matrix weights of the Kalman filter and the sparsity of the adjacency matrix of the graph neural network using a genetic algorithm to output an environmental state vector includes: using the Euclidean distance in the latent space between the real-time collected sensor monitoring data and the generated three-dimensional simulation field data set as a fitness function, setting the covariance matrix weights of the Kalman filter and the sparsity of the adjacency matrix of the graph neural network as parameters to be optimized; Iteratively adjusting the parameters to be optimized through a genetic algorithm to minimize the latent space distance, and generating an optimized joint parameter set of the Kalman filter and the graph neural network; Feeding back the optimized joint parameter set to the volatilization rate equation inversion process to correct the calculation accuracy of the theoretical gas concentration value; Based on the corrected theoretical gas concentration value, the residual sequence calculation and physical field simulation are re-performed, and the three-dimensional simulation field generation logic of the three-dimensional simulation field data set is updated.

[0015] Beneficial effects of the present invention: The present invention constructs a gridded regional benchmark model, divides the storage area into grid units according to the material storage type and binds the sensor node coordinates, integrates the volatile characteristic parameters analyzed in the laboratory, historical monitoring data and environmental thresholds to generate a material-environment relationship matrix, provides a spatiotemporal alignment benchmark and physical law constraints for multi-source heterogeneous sensor data, and effectively solves the spatiotemporal calibration error problem of cross-modal data caused by deployment differences of sensor nodes; dynamically monitors the deviation between real-time data and theoretical values ​​through the residual analysis module, triggers finite element physical field simulation to generate a three-dimensional simulation field data set, combines the Kalman filter and graph neural network of the dual-channel processing module to optimize the fusion weights of real-time data and simulation data respectively, uses the genetic algorithm to minimize the latent space similarity measure, and suppresses environmental noise interference and data distribution deviation; the risk deviation indicator drives the dynamic correction of the material diffusion coefficient and thermodynamic critical value, combines the Bayesian network deduction and Monte Carlo sampling to screen the optimal control instructions, forms a closed-loop feedback link, and realizes the dynamic adaptation of model parameter self-correction and equipment control. Each module uses grid code as index, and through virtual and real data complementarity, parameter collaborative optimization and iterative update mechanism, it significantly improves the dynamic correlation analysis accuracy and early warning reliability of warehouse environment monitoring, and reduces the interference of environmental mutations on the monitoring of physical and chemical properties of materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solution of the present invention, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, for ordinary technicians in this field, other drawings can be obtained based on the drawings without paying any creative labor.

[0017] Figure 1 This is a flow chart of a multi-source sensing warehouse environment collaborative monitoring and early warning method provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. The technical solutions provided by each embodiment of the present invention are described in detail below in conjunction with the drawings. In order to better understand the purpose of the present invention, the present invention is further described in detail below.

[0019] See also Figure 1 The multi-source sensing storage environment collaborative monitoring and early warning method provided by the present invention includes: S101, collect historical environmental monitoring data and laboratory material chemical composition data of the storage area, divide the grid cells according to the material storage type, combine the binding relationship between sensor node coordinates and grid identifiers, receive and set environmental thresholds, and build a gridded regional benchmark model; S102, calling the material volatility characteristic parameters of the corresponding grid in the gridded regional benchmark model, inverting the theoretical gas concentration value through the material volatility rate equation, performing residual sequence calculation between the real-time sensor monitoring data and the theoretical gas concentration value, and triggering a physical field simulation based on the finite element method when the residual continuously exceeds a preset tolerance to generate a three-dimensional simulation field data set; S103, processing the real-time collected sensor monitoring data and the three-dimensional simulation field data set through a preset dual-channel processing process, wherein: The first processing channel performs Kalman filtering on the sensor monitoring data collected in real time for state prediction; The second processing channel extracts physical field features from the three-dimensional simulation field dataset through a graph neural network; S104, dynamically optimizing the covariance matrix weights of the Kalman filter and the adjacency matrix sparsity of the graph neural network using a genetic algorithm, and outputting an environment state vector; S105: Split the environmental state vector into a real-time data stream and a simulation data stream, input the data into a time series convolutional network to extract short-term fluctuation characteristics and a long short-term memory network to capture the physical field evolution trend, respectively, perform weighted fusion of the short-term fluctuation feature vector of the real-time data stream and the physical field evolution trend feature vector of the simulation data stream through an attention mechanism, and perform coupling analysis with the material thermodynamic parameters in the gridded regional benchmark model to generate a risk deviation index; S106, dynamically correcting the material diffusion coefficient and thermodynamic critical value in the gridded regional benchmark model based on the risk deviation index, and updating the environmental threshold; S107, when the risk deviation index exceeds a preset level, calling the updated gridded regional benchmark model parameters, constructing sensor data nodes, material state deduction nodes and equipment control nodes, and generating equipment control instructions.

[0020] The multi-source sensing warehouse environment collaborative monitoring and early warning method of the present invention implements dynamic environmental status assessment based on grid modeling and data fusion optimization. Its technical solution includes the following steps: collecting historical environmental monitoring data and laboratory material chemical composition data of the storage area, dividing the storage area into grid cells according to the material storage type, binding the coordinates of the sensor node and the grid identifier to each grid cell, extracting the volatility characteristic parameters and temperature and humidity correlation coefficients of the materials within the grid cell through laboratory gas chromatography-mass spectrometry analysis, generating a material-environment relationship matrix, and combining historical monitoring data in a preset database to establish a gridded regional benchmark model library containing sensor topology, environmental thresholds, and material volatility characteristics. This model library provides spatial alignment benchmarks and physical law constraints for subsequent data fusion.

[0021] The system uses the material volatility characteristic parameters for the corresponding grid in the gridded regional benchmark model library, inverts the theoretical gas concentration value using the material volatility rate equation, and calculates the residual sequence between the real-time sensor monitoring data and the theoretical gas concentration value. When the mean or variance of the residual sequence continuously exceeds the preset tolerance range, the finite element physical field simulation module is triggered. Combined with the historical data in the gridded model library, it generates a three-dimensional simulation field dataset containing temperature gradients and gas diffusion paths. The residual analysis module uses dynamic thresholds to determine abnormal data distribution, ensuring that the trigger conditions of the physical field simulation adapt to the dynamic changes in the storage environment.

[0022] Real-time sensor monitoring data and three-dimensional simulation field datasets are collaboratively optimized through a dual-channel processing module. The first processing channel uses a Kalman filter to predict the state of the real-time data, generating a noise-reduced real-time data state estimate. The second processing channel uses a graph neural network to extract physical field characteristics such as temperature and concentration fields from the three-dimensional simulation field data and generate a simulation data feature vector. The real-time data state estimate output by the dual channels and the simulation data feature vector are similarly measured in the latent space. A genetic algorithm is used to iteratively optimize the covariance matrix weights of the Kalman filter and the sparsity of the graph neural network adjacency matrix to minimize the latent space similarity measure. The fused output is an environmental state vector with a confidence label.

[0023] The environmental state vector is split into a real-time data stream and a simulated data stream. The real-time data stream is fed into a time-series convolutional network to extract short-term fluctuation characteristics, while the simulated data stream is fed into a long-short-term memory network to capture the evolution of the physical field. The two feature vectors are dynamically weighted and fused using an attention mechanism to generate a fused feature vector. This fused feature vector is then coupled with the thermodynamic parameters of the materials in the gridded model library for analysis. The degree of deviation between the current environmental state and the material's stability conditions is calculated, and a risk deviation index is output. This index drives the gradient descent optimization of the material diffusion coefficient and thermodynamic critical value in the gridded model library to update the environmental threshold parameters.

[0024] When the risk deviation index exceeds the preset level, the updated grid model parameters are used to construct a Bayesian network. Based on Monte Carlo sampling, the environmental evolution trajectory under different control strategies is simulated. The control instruction combination that makes the physical parameters of the three-dimensional simulation field converge with the environmental threshold is selected to generate equipment control instructions. The equipment control results are fed back to the residual analysis module through sensors. The theoretical gas concentration value is recalculated and residual calibration is performed. The material diffusion coefficient is dynamically adjusted and then returned to the physical field simulation module. The three-dimensional simulation field generation logic is reconstructed, forming a closed-loop feedback link between grid model parameter self-correction and equipment control. Each step achieves data spatiotemporal alignment through grid encoding. The latent space similarity measurement and genetic algorithm optimization suppress cross-modal data deviations, improving the accuracy of dynamic correlation analysis of warehouse environment monitoring.

[0025] Specifically, the multi-source sensing warehouse environment collaborative monitoring and early warning method of the present invention, wherein the gridded regional benchmark model is constructed includes: Divide the storage area into grid units according to material storage types, and bind sensor node coordinates to grid identifiers; The correlation coefficients between the volatility characteristics of substances and temperature and humidity in the grid cells were extracted by laboratory gas chromatography-mass spectrometry analysis to generate a substance-environment relationship matrix. The substance-environment relationship matrix is ​​associated with historical monitoring data in a preset database according to grid coding to generate a gridded regional benchmark model library including sensor topology, environmental thresholds and substance volatility characteristic parameters.

[0026] When constructing a gridded regional benchmark model, the storage area is divided into multiple grid cells based on material storage type. Material storage type includes physical and chemical properties such as flammability, volatility, and corrosiveness. The grid cells are divided based on the physical layout of the storage space and the chemical properties of the stored materials. Each grid cell is assigned a unique identifier. Sensor node coordinates are bound to the grid identifier through a spatial mapping algorithm. Coordinate transformation techniques are used to map sensor deployment locations to the center point or key monitoring location of the corresponding grid cell. A one-to-one or one-to-many association is established between sensor nodes and grid cells, providing a benchmark framework for the spatial alignment of multi-source sensor data.

[0027] Laboratory gas chromatography-mass spectrometry was used to determine the volatility parameters of substances stored within each grid cell, including volatility rate, saturated vapor pressure, and diffusion coefficient. Temperature and humidity control experiments were combined to obtain dynamic data on substance volatility under different environmental conditions, and the nonlinear correlation coefficient between temperature and humidity and volatility rate was extracted. After normalization of the experimental data, a substance-environment relationship matrix was constructed. The row vectors of the matrix represent the grid cell identifiers, and the column vectors contain volatility parameters, temperature and humidity influencing factors, and time series correlation weights. The matrix elements quantitatively reflect the volatility behavior of substances under specific environmental conditions.

[0028] The substance-environment relationship matrix is ​​imported into a pre-set database, which stores historical monitoring data, including temperature, humidity, gas concentration, and timestamp information collected by sensors. Experimental data is linked to historical data through grid coding, which is generated by combining grid cell identifiers, substance category codes, and time interval identifiers. This enables multidimensional index matching of experimental data and historical monitoring data. During the association process, a data fusion algorithm is used to integrate sensor topology, environmental safety thresholds, and substance volatility parameters to generate a gridded regional benchmark model library. This model library includes a spatial distribution model, a volatility dynamics model, and an environmental risk threshold table, providing a unified parameter benchmark for the collaborative analysis of real-time monitoring data and simulation data.

[0029] Specifically, the multi-source sensing storage environment collaborative monitoring and early warning method of the present invention, wherein the residual sequence calculation of the real-time collected sensor monitoring data and the theoretical gas concentration value includes: Calling the substance volatilization characteristic parameters of the corresponding grid in the gridded regional benchmark model library, and inverting the theoretical gas concentration value according to the substance volatilization rate equation; Calculating a residual sequence between the real-time collected sensor monitoring data and the theoretical gas concentration value, and triggering a physical field simulation based on the finite element method when the residual continues to exceed a preset tolerance; Combining historical data in the gridded regional benchmark model library, a three-dimensional simulation field data set including temperature gradients and gas diffusion paths is generated.

[0030] When calculating the residual series between real-time sensor monitoring data and theoretical gas concentration values, the system first uses the material volatility characteristic parameters for the corresponding grid cells in the gridded regional benchmark model library, including the volatility rate constant, diffusion coefficient, and saturated vapor pressure. Using the material volatility rate equation, a kinetic model of the time-varying gas concentration is established, and the theoretical gas concentration value is inverted. The volatility rate equation is constructed based on Fick's diffusion law and the Arrhenius equation. The input parameters include the temperature and humidity data of the current grid, the material's physical and chemical properties, and historical volatility trends. The output is a time series forecast of the theoretical gas concentration, providing a benchmark reference for residual analysis.

[0031] Real-time sensor monitoring data is aligned with theoretical gas concentration values ​​using a sliding window method, and the residual sequence for each sampling point is calculated. The statistical characteristics of the residual sequence include mean, variance, and range. A preset tolerance range is set based on the confidence interval of historical monitoring data. When the statistics of the residual sequence continuously exceed the tolerance range, the environmental state is determined to be abnormal, triggering the finite element physical field simulation module. The finite element simulation initializes boundary conditions based on historical data from the grid model library. Combined with the real-time temperature and humidity field distribution, it solves the gas diffusion partial differential equation to generate a three-dimensional simulation field data set containing temperature gradients and gas diffusion paths.

[0032] During the generation of the 3D simulation field dataset, historical data from the gridded model library is used through a spatiotemporal interpolation algorithm to fill in the spatial blind spots of the real-time monitoring data. This data is then integrated with heterogeneous data from temperature, humidity, and gas concentration sensors to construct a continuous 3D physical field distribution model. The simulation results are output as discretized grid node data, each containing temperature, concentration gradient, and diffusion flux parameters. A data reconstruction algorithm is then used to generate a smooth 3D simulation field dataset, which provides simulation data input for the dual-channel data processing module. A residual triggering mechanism forms a closed-loop feedback loop with the simulation data generation to dynamically optimize the prediction accuracy of the theoretical model.

[0033] Specifically, the multi-source sensing warehouse environment collaborative monitoring and early warning method of the present invention processes the real-time collected sensor monitoring data and the three-dimensional simulation field data set through a preset dual-channel processing process, including: Performing Kalman filtering on the sensor monitoring data collected in real time to predict the state and generate a real-time data state estimate; Extracting physical field features from the three-dimensional simulation field dataset through a graph neural network to generate a simulation data feature vector; Based on the real-time data state estimation value and the simulation data feature vector, calculating a similarity measure between the real-time data state estimation value and the simulation data feature vector in a latent space; Using the similarity metric as an optimization goal, driving a genetic algorithm to iteratively adjust the covariance matrix weights of the Kalman filter and the sparsity of the adjacency matrix of the graph neural network; According to the optimized covariance matrix weights and adjacency matrix sparsity, the real-time data state estimation value and the simulated data feature vector are fused to output an environment state vector with a confidence label.

[0034] Real-time sensor monitoring data and a three-dimensional simulation field dataset are collaboratively optimized through a dual-channel processing module. The first processing channel applies Kalman filtering to the real-time sensor data. The Kalman filter's state equation is constructed based on a dynamic model of the warehouse environment. The state transition matrix and observation matrix are calibrated using historical monitoring data. After the real-time data is input, a prediction-correction cycle is used to generate a noise-reduced state estimate, suppressing the impact of sensor noise and short-term environmental disturbances on data quality. The second processing channel uses a graph neural network to extract features from the three-dimensional simulation field data. The graph structure is constructed based on the spatial topological relationships of grid cells. Node features include temperature, gas concentration, and diffusion flux parameters. Edge weights are determined by the correlation of physical field gradients. Graph convolutional and pooling layers are used to extract multi-scale physical field feature vectors to characterize the global evolution trend of the warehouse environment.

[0035] Real-time data state estimates and simulated data feature vectors are mapped to a latent space through an embedding layer. The latent space dimension is adaptively determined by the complexity of the data modality. Cosine similarity or Euclidean distance is used to calculate the similarity measure between the two in the latent space, measuring the consistency of the spatial distribution of real-time monitoring data and simulated data. The similarity measure serves as the fitness function of the genetic algorithm. The parameters to be optimized include the covariance matrix weights of the Kalman filter and the sparsity threshold of the graph neural network adjacency matrix. The genetic algorithm iteratively adjusts the parameter combination through selection, crossover, and mutation operations to converge the latent space similarity measure to a minimum value. The optimized parameter combination constrains the weight distribution strategy of data fusion.

[0036] During the fusion process, the real-time data state estimate and the simulated data feature vector are fused through a dynamic weighted summation. The weight coefficient is determined by a combination of a similarity metric and optimized parameters. The fused data is mapped into an environmental state vector through a fully connected layer. The output vector contains parameters such as temperature stability, gas concentration anomaly index, and diffusion risk level. Each dimension is labeled with a confidence level, which is calculated by combining the residual variance of the Kalman filter and the feature reconstruction error of the graph neural network. The environmental state vector serves as the input to the risk assessment module and undergoes multidimensional correlation analysis with thermodynamic parameters in the gridded model library, supporting the subsequent calculation of risk deviation indicators and dynamic correction of model parameters.

[0037] Specifically, in the multi-source sensing warehouse environment collaborative monitoring and early warning method of the present invention, generating the risk deviation index includes: Splitting the environmental state vector into a real-time data stream and a simulation data stream; Inputting the real-time data stream into a time series convolutional network to extract a first eigenvector representing short-term fluctuations of environmental parameters; Inputting the simulation data stream into a long short-term memory network to extract a second eigenvector representing an evolution trend of the physical field; Performing weighted fusion of the first feature vector and the second feature vector through an attention mechanism to generate a fused feature vector; The fused feature vector is coupled with the material thermodynamic parameters in the gridded regional benchmark model library for analysis, and a risk deviation index representing the degree to which the environmental state deviates from the material stability condition is output.

[0038] When generating risk deviation indicators, the environmental state vector is split into a real-time data stream and a simulated data stream, based on differences in the data's spatiotemporal characteristics. The real-time data stream contains time-series environmental parameters collected by sensors, while the simulated data stream originates from gridded spatially distributed data generated by physical field simulation. The splitting process is implemented using a data channel separation algorithm. The separated data stream retains the original timestamp and grid encoding information, providing an independent input source for subsequent feature extraction. The real-time data stream is input into a time-series convolutional network. The network structure uses a stack of dilated causal convolutional layers. The convolution kernel width is adaptively adjusted according to the sensor sampling frequency to capture short-term fluctuations in environmental parameters. The output is the first eigenvector representing instantaneous changes in temperature and gas concentration. The feature dimension is consistent with the number of grid cells.

[0039] The simulated data stream is fed into a long-short-term memory (LSTM) network. The network's input layer receives a time series of three-dimensional simulated field data. Each time step corresponds to the temperature gradient, gas diffusion flux, and humidity distribution parameters of the grid nodes. A gating mechanism filters historical state information, and the output is a second eigenvector representing the evolutionary trend of the physical field. The number of hidden layer nodes in the LSTM network matches the spatial resolution of the grid cells, and the eigenvector dimension reflects the multi-scale nature of the physical field evolution. The first and second eigenvectors are dynamically fused using a multi-head attention mechanism. Attention weights are calculated based on the temporal and spatial correlations of the eigenvectors. The weighted fusion eigenvector contains information about both short-term anomalies and long-term evolution of the environmental state.

[0040] The fused feature vectors are coupled with the material thermodynamic parameters in the gridded regional benchmark model library for analysis. The material thermodynamic parameters include the critical value of the volatilization rate, the diffusion coefficient threshold, and the thermal stability boundary conditions. The coupling analysis uses a tensor product operation to map the fused feature vectors to the material thermodynamic parameter space, calculates the deviation of the environmental state of each grid cell from the stability condition, and converts the deviation into a risk deviation index through normalization. The indicator output is a multidimensional vector, where the dimensions correspond to the safety level of different material categories, and the numerical range is aligned with the preset environmental threshold range, which is used to trigger graded warning signals and equipment control instructions. The feature fusion and parameter coupling process is implemented through a differentiable computational graph, supporting end-to-end model training and parameter optimization.

[0041] Specifically, the multi-source sensing warehouse environment collaborative monitoring and early warning method of the present invention, wherein the calling of updated gridded regional benchmark model parameters, the construction of sensor data nodes, material state deduction nodes, and equipment control nodes, and the generation of equipment control instructions include: constructing sensor data nodes, material state deduction nodes, and equipment control nodes in a Bayesian network based on the updated gridded regional benchmark model parameters; Initializing the conditional probability table of the Bayesian network nodes using the Kalman filter covariance matrix weights and the graph neural network adjacency matrix sparsity optimized by the genetic algorithm; The generated three-dimensional simulation field data set and the updated environmental threshold are input into the Bayesian network, and the environmental state change trajectory under different control strategies is simulated through Monte Carlo sampling; screening a control instruction combination that causes the physical field parameters in the three-dimensional simulation field data set to converge with the updated environmental threshold; Device control instructions are generated according to the screening results, and dynamic thresholds in the gridded regional benchmark model parameter library are synchronously updated.

[0042] When generating risk deviation indicators, the environmental state vector is split into a real-time data stream and a simulated data stream based on the differences in the spatiotemporal characteristics of the data sources. The real-time data stream contains time-series environmental parameters collected by sensors, while the simulated data stream consists of gridded physical field parameters from a three-dimensional simulated field dataset. The splitting process is implemented using a data channel separation algorithm, preserving the original timestamp and grid encoding information. The real-time data stream is input into a time-series convolutional network. The network adopts a dilated convolution structure, and the convolution kernel width is dynamically adjusted based on the sensor sampling frequency and the periodic characteristics of the data. Multi-level convolution operations are used to capture short-term fluctuation patterns in temperature and gas concentration. The output is the first eigenvector representing the instantaneous changes in environmental parameters. The feature dimension matches the number of grid cells, and each dimension corresponds to the abnormal fluctuation intensity of a specific grid.

[0043] The simulation data stream is fed into the LSTM network. The network input layer receives a time series of three-dimensional simulated field data. The input data for each time step includes the temperature gradient, gas diffusion flux, and humidity distribution parameters of the grid nodes. The LSTM network filters historical state information through forget gates and input gates. The memory cells store the evolution trend of the physical field, and the output is a second eigenvector representing the diffusion path and long-term changes in the temperature field. The eigenvector dimension matches the spatial resolution of the grid, reflecting the multi-scale spatiotemporal correlation characteristics of the physical field parameters.

[0044] The first and second eigenvectors are dynamically fused using a multi-head attention mechanism. Attention weights are calculated based on the temporal and spatial correlations of the eigenvectors. Temporal correlation is determined by the eigenvector's self-attention score, while spatial correlation is weighted by the topological distance of the grid code. The weighted fused eigenvector integrates short-term anomalies of environmental parameters with the long-term evolution of the physical field, generating a fused eigenvector containing multidimensional information in time and space. The fused eigenvector is then tensor-multiplied with the material thermodynamic parameters from the gridded regional benchmark model library. These parameters include the critical value for volatilization rate, the threshold for diffusion coefficient, and the boundary conditions for thermal stability. The resulting eigenvector is then mapped to the same parameter space.

[0045] Through normalization, the mapping results are converted into environmental state deviations for each grid cell. The deviations represent the degree of difference between the current environmental parameters and the material's stability conditions. The deviations are weighted and aggregated, and then output as a risk deviation index. The index dimensions correspond to the safety levels of different material categories, and the numerical range aligns with the preset environmental threshold range. The risk deviation index triggers a graded warning signal and serves as the basis for generating equipment control instructions. The feature fusion and parameter coupling process is implemented through a differentiable computational graph, supporting end-to-end gradient backpropagation and dynamically optimizing the collaborative efficiency of feature extraction and parameter analysis.

[0046] Specifically, the multi-source sensing warehouse environment collaborative monitoring and early warning method of the present invention further includes: collecting environmental parameter feedback data after the execution of the equipment control instruction in real time through sensors, including temperature change rate, gas concentration gradient and humidity distribution state; Input the feedback data into the residual analysis module, recalculate the theoretical gas concentration value, and perform residual sequence comparison with the measured value in the feedback data; When the statistical characteristics of the residual sequence exceed the environmental threshold of the gridded regional benchmark model library, a data calibration process is triggered to dynamically adjust the material diffusion coefficient and thermodynamic critical value in the model library; Synchronizing the updated model parameters to the dual-channel processing and the physical field simulation module, reconstructing the data fusion strategy and the three-dimensional simulation field generation logic; Through the iterative adaptation of the updated model parameters and the equipment control instructions, a closed-loop feedback link of gridded regional benchmark model parameter optimization and storage environment equipment control is formed.

[0047] After the equipment control instructions are executed, multi-source sensors deployed in the storage area collect environmental parameter feedback data in real time. The feedback data includes the temperature change rate, gas concentration gradient, and humidity distribution status. Sensor types include thermocouples, gas concentration sensors, and humidity sensors. The data collection frequency is synchronized with the response cycle of the equipment control action. After preprocessing, the feedback data is input into the residual analysis module. The preprocessing includes noise filtering, timestamp alignment, and grid code matching to ensure that the data format is consistent with the input requirements of the gridded regional benchmark model library. The residual analysis module calls the material volatility characteristic parameters in the current model library and recalculates the theoretical gas concentration value based on the volatility rate equation. The time decay factor is introduced into the calculation process of the theoretical value to reflect the dynamic changes in environmental conditions after equipment control.

[0048] The recalculated theoretical gas concentration values ​​are compared with the measured values ​​in the feedback data using a sliding window method for residual sequence comparison. The window length is dynamically adjusted according to the effective time of the equipment control. The statistical characteristics of the residual sequence include mean, variance, and autocorrelation coefficient. After the statistics are calculated, they are compared with the environmental thresholds preset in the gridded regional benchmark model library. The environmental thresholds are set according to the thermodynamic stability level of the material. When the statistical characteristics of the residual sequence exceed the threshold range, the data calibration process is triggered. The calibration process iteratively adjusts the material diffusion coefficient and thermodynamic critical value in the model library through the gradient descent algorithm. The optimization goal is to make the statistics of the residual sequence converge to the threshold range.

[0049] The updated model parameters are synchronized to the dual-channel processing and physical field simulation modules via a parameter publish-and-subscribe mechanism. The dual-channel processing receives the adjusted Kalman filter covariance matrix weights and graph neural network adjacency matrix sparsity and reconstructs the weight distribution strategy for data fusion. The physical field simulation module reinitializes the boundary conditions of the gas diffusion partial differential equation based on the revised diffusion coefficient and thermodynamic critical value, generating an updated 3D simulation field dataset containing the revised temperature gradient and gas diffusion path. Version control technology is used during parameter synchronization to avoid time series data conflicts caused by model parameter updates.

[0050] The updated model parameters and equipment control commands form a closed-loop feedback loop through an iterative adaptation mechanism. This adaptation mechanism, built on a reinforcement learning framework, uses an environmental state deviation indicator as a reward function and the execution effect of the equipment control commands as a state transition signal. After each parameter update, the efficiency of the control commands in correcting the environmental state is reassessed, and the command generation strategy is optimized. Residual analysis, parameter calibration, and model update cycles within the closed-loop link are aligned with the sensor data acquisition frequency, enabling dynamic optimization of the gridded regional benchmark model parameters and real-time response coordination of the warehouse environment control equipment. Data and control flows communicate asynchronously through a message queue middleware, ensuring system stability in high-concurrency scenarios.

[0051] Specifically, in the multi-source sensing warehouse environment collaborative monitoring and early warning method of the present invention, generating a three-dimensional simulation field data set includes: generating initial three-dimensional simulation field data including temperature gradients and gas diffusion paths based on the finite element physical field simulation results; Adding a grid code mark to the initial three-dimensional simulated field data, wherein the grid code is aligned with the grid unit identifier and the bound sensor node coordinates according to a spatial topological relationship; In the dual-channel processing, the material volatility characteristic parameters in the gridded regional benchmark model library are matched based on the grid code; Using the substance volatility characteristic parameters as constraints, calculating a similarity measure between the real-time data state estimate and the simulated data feature vector in a latent space; According to the similarity metric optimization result, the physical field parameters of the three-dimensional simulation field data set are synchronously corrected to generate a calibrated three-dimensional simulation field data set that is spatially and temporally aligned with the real-time monitoring data.

[0052] When generating a 3D simulation field data set, the 3D simulation field data is initialized based on the calculation results of the finite element physical field simulation module. The simulation module solves the coupled partial differential equations for the temperature field and gas diffusion field, combining the boundary conditions and initial parameters in the gridded regional benchmark model library to generate the initial 3D simulation field data, including temperature gradients and gas diffusion paths. During the finite element simulation process, the spatial discretization of the grid cells uses tetrahedral or hexahedral elements. The unit node coordinates match the sensor deployment locations, and the boundary conditions are obtained by interpolation of real-time monitoring data to ensure the physical consistency of the simulation results.

[0053] The initial 3D simulation field data is tagged with a grid code. The grid code is generated by combining a grid cell identifier, a material category code, and a timestamp. The coding rules align with the spatial topology of the sensor node coordinates. The grid cell identifier is uniquely determined based on the storage area division rules. The material category code corresponds to the physical and chemical properties of the stored material, and the timestamp indicates the timing of data generation. The code attachment process is implemented using a hash mapping algorithm, which uniquely binds the data and code of each grid node, providing an index basis for subsequent parameter matching.

[0054] During the dual-channel processing, material volatility characteristic parameters are matched from a gridded regional benchmark model library based on grid codes. This matching process utilizes a key-value query mechanism, where the key is the grid code and the value is the material volatility rate constant, diffusion coefficient, and thermodynamic critical value for the corresponding grid cell. The matched parameters undergo data normalization and are converted to the same dimensional system as the real-time data state estimate and the simulated data feature vector, serving as constraints for the latent space similarity metric.

[0055] In calculating the similarity metric between the real-time data state estimates and the simulated data feature vectors in the latent space, the material volatility parameters are introduced as constraints in the loss function. The loss function is composed of a linear combination of the Euclidean distance and the parameter constraints, and the weight coefficient is dynamically adjusted based on the material category corresponding to the grid code. The similarity metric calculation results are used to evaluate the consistency between the simulated data and the real-time monitoring data. Lower metric values ​​indicate closer distributions in the latent space and higher confidence in the data fusion.

[0056] Based on the similarity metric optimization results, a gradient descent algorithm is used to simultaneously correct the physical field parameters of the 3D simulation field dataset. These corrections include the temperature gradient coefficient, gas diffusion rate, and thermal conductivity coefficient. These corrections are then fed back to the finite element simulation module, where the temperature and concentration field distributions at the mesh nodes are recalculated to generate a calibrated 3D simulation field dataset. This calibrated dataset is aligned with the timestamps of the real-time monitoring data using a spatiotemporal interpolation algorithm, eliminating timing deviations between the simulation data and the sensor data, enabling dynamic visualization of the warehouse environment and supporting risk assessment.

[0057] Specifically, in the multi-source sensing warehouse environment collaborative monitoring and early warning method of the present invention, the method of dynamically correcting the material diffusion coefficient and thermodynamic critical value in the gridded area benchmark model based on the risk deviation index and updating the environmental threshold value includes: extracting material thermodynamic equation parameters associated with the current environmental state in the gridded area benchmark model library based on the generated risk deviation index; According to the value of the risk deviation index, the diffusion coefficient and the thermodynamic critical value in the thermodynamic equation of the substance are iteratively adjusted through a gradient descent algorithm; The updated diffusion coefficient and thermodynamic critical value are transmitted back to the physical field simulation module to reconstruct the boundary conditions of the gas diffusion equation in the process of generating the three-dimensional simulation field data set; a preset finite element physical field simulation method is called to recalculate the gas diffusion path based on the updated diffusion coefficient to generate a revised three-dimensional simulation field data set; The modified three-dimensional simulation field data set is input into the dual-channel processing process, the latent space similarity measure is recalculated, and the data fusion weight is optimized simultaneously.

[0058] After generating the risk deviation index, the diffusion coefficient and thermodynamic critical value of the material in the gridded regional benchmark model are dynamically modified based on this index. First, the material thermodynamic equation parameters associated with the current environmental state are extracted from the gridded regional benchmark model library, including the diffusion coefficient, critical temperature, and volatilization rate constant. The association conditions are determined by mapping the numerical range of the risk deviation index to the material category. A key-value indexing mechanism is used during the extraction process, with the key being the grid code and the value being the set of thermodynamic parameters for the corresponding grid cell. Indexing efficiency is optimized through hash tables, supporting fast queries in high-concurrency scenarios.

[0059] Based on the value of the risk deviation index, a gradient descent algorithm is used to iteratively adjust the diffusion coefficient and thermodynamic critical value in the material thermodynamic equation. The objective function is composed of the weighted squared error between the risk deviation index and the parameter adjustment amount. The weight coefficient is dynamically assigned based on the material type of the grid cell. The step size parameter of the gradient descent is adaptively adjusted. The initial step size is set based on historical optimization records. During the iteration process, the step size is dynamically reduced based on the convergence rate of the objective function until the parameter adjustment amount falls below the preset threshold or the maximum number of iterations is reached, achieving a balance between stability and efficiency in parameter optimization.

[0060] The updated diffusion coefficient and thermodynamic critical value are transmitted back to the physical field simulation module via a parameter synchronization interface. This interface uses a message queue middleware to implement asynchronous data transmission, avoiding main thread blocking. After receiving the new parameters, the physical field simulation module reconstructs the boundary conditions of the gas diffusion equation, which include the initial concentration distribution, temperature gradient, and diffusion flux constraints. This reconstruction process is completed by updating the node parameters of the finite element mesh. The preset finite element physical field simulation method is called to resolve the gas diffusion partial differential equation based on the updated diffusion coefficient, calculate the concentration field and temperature field distribution of the mesh nodes, and generate a revised three-dimensional simulation field dataset. The dataset contains the revised diffusion path and temperature gradient parameters.

[0061] The revised three-dimensional simulated field dataset is input into the dual-channel processing module. The dual-channel processing module receives the updated simulation data and recalculates the similarity measure between the real-time data state estimate and the simulated data feature vector in the latent space. The similarity measure calculation introduces the updated diffusion coefficient as a constraint term, and data consistency is assessed through a weighted combination of cosine similarity and Euclidean distance. The optimized similarity measure drives the genetic algorithm to adjust the Kalman filter covariance matrix weights and the sparsity of the graph neural network adjacency matrix, dynamically assigning fusion weights for real-time and simulated data, and improving the confidence level of data fusion. The coordinated optimization of parameter correction and data fusion forms a closed-loop iterative chain, supporting the adaptive calibration of grid model parameters and the dynamic accuracy improvement of warehouse environment monitoring.

[0062] Specifically, the multi-source sensing warehouse environment collaborative monitoring and early warning method described in the present invention uses a genetic algorithm to dynamically optimize the covariance matrix weights of the Kalman filter and the sparsity of the adjacency matrix of the graph neural network, and outputs the environmental state vector, including: using the Euclidean distance in the latent space between the real-time collected sensor monitoring data and the generated three-dimensional simulation field data set as the fitness function, setting the covariance matrix weights of the Kalman filter and the sparsity of the adjacency matrix of the graph neural network as parameters to be optimized; Iteratively adjusting the parameters to be optimized through a genetic algorithm to minimize the latent space distance, and generating an optimized joint parameter set of the Kalman filter and the graph neural network; Feeding back the optimized joint parameter set to the volatilization rate equation inversion process to correct the calculation accuracy of the theoretical gas concentration value; Based on the corrected theoretical gas concentration value, the residual sequence calculation and physical field simulation are re-performed, and the three-dimensional simulation field generation logic of the three-dimensional simulation field data set is updated.

[0063] When dynamically optimizing the parameters of the Kalman filter and graph neural network using a genetic algorithm, real-time sensor data and a three-dimensional simulated field dataset are first mapped into a latent space through an embedding layer. The latent space dimension is adaptively determined by the complexity of the data modality. The Euclidean distance in the latent space serves as the fitness function, reflecting the distribution consistency of real-time and simulated data. The parameters to be optimized include the covariance matrix weights of the Kalman filter and the sparsity threshold of the graph neural network adjacency matrix. The parameter initialization range is set based on historical optimization records to avoid the loss of convergence efficiency caused by an overly large search space.

[0064] The genetic algorithm iteratively adjusts parameter combinations through selection, crossover, and mutation operations. The selection mechanism is based on a roulette wheel strategy, where individuals with lower fitness values ​​are more likely to be selected. The crossover operation uses a single-point crossover method, randomly selecting a crossover point within the parameter encoding sequence to exchange gene segments. The mutation operation perturbs the parameter values ​​using Gaussian noise to increase population diversity. The optimal individual from each generation is retained during the iteration process until the Euclidean distance in the latent space converges to a preset threshold or the maximum number of iterations is reached. This generates the optimized joint parameter set for the Kalman filter and graph neural network.

[0065] The optimized joint parameter set is fed back to the volatilization rate equation inversion process via a message queue middleware. The inversion module receives the updated covariance matrix weights, adjusts the noise covariance estimate of the state prediction model, and reconstructs the input parameters of the volatilization rate equation. A dynamic attenuation factor is introduced into the calculation of the corrected theoretical gas concentration value to reflect the dynamic characteristics of the substance volatilization after environmental control, improving the time series matching accuracy between the theoretical and measured values.

[0066] Based on the corrected theoretical gas concentration value, the residual sequence calculation is re-executed, and the residual window length is dynamically adjusted according to the response time of the volatilization rate. The residual statistical characteristics are input into the finite element physical field simulation module, triggering the update logic of the three-dimensional simulation field data set. The simulation module solves the coupled partial differential equations of the temperature field and gas diffusion field based on the new parameters to generate an updated three-dimensional simulation field data set. The updated data set is aligned with the real-time monitoring data in time and space through grid encoding, and input into the dual-channel processing module to recalculate the latent space similarity measure, forming a closed-loop feedback link of parameter optimization, data correction and model iteration, supporting the dynamic accuracy improvement and stability enhancement of the warehouse environment monitoring system.

[0067] The names of the technical features of the present invention are explained as follows: The gridded regional benchmark model construction algorithm divides grid cells based on material storage type, binds sensor node coordinates to grid identifiers, extracts volatility characteristic parameters through laboratory gas chromatography-mass spectrometry analysis, generates a material-environment relationship matrix, and combines historical monitoring data to establish a model library that includes sensor topology, environmental thresholds, and material volatility characteristics. The algorithm achieves spatiotemporal alignment of multi-source data through grid encoding, providing a unified physical law constraint and spatial index benchmark for subsequent analysis.

[0068] The residual analysis algorithm uses volatility characteristic parameters from a gridded model library to invert theoretical gas concentrations based on Fick's diffusion law and the Arrhenius equation. A sliding window method is then used to calculate the residual sequence between real-time sensor data and theoretical values. When residual statistics (mean, variance) consistently exceed preset tolerances, a finite element simulation is triggered, combining historical data to generate a 3D simulated field dataset, filling monitoring blind spots and suppressing environmental noise interference.

[0069] In the dual-channel processing algorithm, the first channel uses a Kalman filter to predict the state of real-time data. This filter suppresses sensor noise through the state equation and the observation matrix, generating a de-noised state estimate. The second channel extracts the physical field characteristics of the three-dimensional simulation field through a graph neural network. The graph structure is constructed based on a grid topology, and node features include parameters such as temperature gradient and diffusion flux. After the outputs of both channels are mapped to the latent space, a genetic algorithm is used to optimize the weights of the Kalman filter covariance matrix and the sparsity of the graph neural network adjacency matrix, minimize the Euclidean distance in the latent space, dynamically adjust the data fusion weights, and output the environmental state vector with confidence labels.

[0070] The risk deviation generation algorithm splits the environmental state vector into real-time and simulated data streams, feeding them into a time-series convolutional network and a long-short-term memory network, respectively. It extracts short-term fluctuation characteristics and long-term evolution trends, then performs weighted fusion using a multi-head attention mechanism to generate a fused feature vector. This vector is then tensor-multiplied with thermodynamic parameters from the gridded model library (such as diffusion coefficient thresholds and volatilization rate critical values). After normalization, it outputs a risk deviation index, quantifying the degree of deviation of the environmental state from material stability conditions.

[0071] The dynamic parameter correction algorithm uses a gradient descent method to iteratively optimize the material diffusion coefficient and thermodynamic critical value based on the risk deviation index. The updated parameters are synchronized to the physical field simulation module via a message queue, reconstructing the boundary conditions of the gas diffusion equation and generating a corrected three-dimensional simulation field dataset. The parameter correction process is linked to the dual-channel processing module, optimizing the data fusion strategy through latent space similarity measurement, and forming a parameter self-correction mechanism.

[0072] The closed-loop feedback algorithm utilizes a Bayesian network to construct probabilistic relationships between sensor data nodes, material state deduction nodes, and device control nodes. Combined with Monte Carlo sampling, it simulates the environmental evolution trajectories of different control strategies, selecting command combinations that converge physical field parameters with environmental thresholds. Device control results are fed back to the residual analysis module via sensors, recalibrating model parameters and updating simulation logic, achieving iterative adaptation of model optimization and device control.

[0073] The gridded regional benchmark model divides storage areas into grid cells based on material storage type. Each cell is bound to sensor node coordinates and generates a unique grid code. Laboratory gas chromatography-mass spectrometry is used to extract parameters such as volatilization rate and diffusion coefficient. Combined with historical monitoring data, a matrix of material-environment relationships is generated, and a model library is constructed that includes sensor topology, environmental thresholds, and volatilization characteristics. This model achieves spatiotemporal alignment of multi-source data through grid coding, providing a unified spatial index and physical constraints for residual analysis and physical field simulation.

[0074] The residual analysis model uses volatility characteristic parameters from a gridded model library, inverts theoretical gas concentrations based on Fick's diffusion law and the Arrhenius equation, and calculates the residual sequence between real-time data and theoretical values ​​using a sliding window method. When the residual mean or variance exceeds preset tolerances, a finite element physical field simulation is triggered, combining historical data to generate a 3D simulation field dataset, filling sensor monitoring blind spots and suppressing environmental noise interference.

[0075] In this dual-channel processing model, the first channel uses a Kalman filter to predict the state of real-time data. This filter suppresses sensor noise through the state equation and the observation matrix, and outputs a de-noised real-time data state estimate. The second channel uses a graph neural network to extract physical field features such as temperature gradients and diffusion flux from the three-dimensional simulated field data. The graph structure is constructed based on grid topology, with node features corresponding to grid codes. After mapping the two-channel data through the latent space, a genetic algorithm is used to optimize the Kalman filter covariance matrix weights and the graph neural network adjacency matrix sparsity, minimize the latent space Euclidean distance, dynamically adjust the fusion weights, and output the environment state vector with confidence labels.

[0076] The risk deviation generation model splits the environmental state vector into a real-time data stream and a simulated data stream, respectively feeding them into a time-series convolutional network and a long short-term memory network to extract short-term fluctuation characteristics and long-term evolution trends. A multi-head attention mechanism performs a weighted fusion of the two feature vectors to generate a fused feature vector. This fused feature vector is then combined with thermodynamic parameters from the gridded model library (such as diffusion coefficient threshold and volatilization rate critical value) for tensor operations. After normalization, it outputs a risk deviation indicator that quantifies the degree of deviation of the environmental state from the material's stability conditions.

[0077] The dynamic parameter correction model uses a gradient descent method to iteratively optimize the material diffusion coefficient and thermodynamic critical value based on the risk deviation indicator. The updated parameters are synchronized to the physical field simulation module via a message queue, reconstructing the boundary conditions of the gas diffusion equation and generating a revised three-dimensional simulation field dataset. The parameter correction process is linked to the dual-channel processing module, optimizing the data fusion strategy through latent space similarity measurement, and forming a parameter self-correction mechanism.

[0078] The closed-loop feedback model utilizes a Bayesian network to construct probabilistic relationships between sensor data nodes, material state deduction nodes, and device control nodes. Combined with Monte Carlo sampling, it simulates environmental evolution trajectories under different control strategies, selecting command combinations that converge physical field parameters with environmental thresholds. Device control results are fed back to the residual analysis module via sensors, recalibrating model parameters and updating simulation logic, achieving iterative adaptation of model optimization and device control.

[0079] The above model uses grid coding as the data link. Through the complementarity of virtual and real data, parameter collaborative optimization and closed-loop iteration mechanism, it solves the problem of spatiotemporal alignment failure of multi-source heterogeneous sensor data, improves the accuracy of dynamic correlation analysis, and ensures the reliability of warehouse environment monitoring and real-time early warning.

[0080] The specific implementation of the present invention is based on the demand for collaborative monitoring of multi-source sensor data in the storage environment, combined with the physical and chemical properties of material storage and dynamic environmental changes, and realizes the monitoring and early warning functions through the following steps: First, the grid units are divided according to the material storage type of the storage area, each grid unit is bound to the sensor node coordinates and a unique grid code is generated. The grid code consists of a material category code, a spatial location identifier and a timestamp. The volatility characteristic parameters of the material in each grid unit are analyzed by laboratory gas chromatography and mass spectrometry technology, including the volatilization rate constant, saturated vapor pressure and diffusion coefficient. The dynamic correlation data of the volatilization of the material and the environment are obtained by combining temperature and humidity control experiments, and a material and environment relationship matrix is ​​constructed. The experimental data and historical monitoring data are associated through grid coding, and the sensor topology, environmental threshold and volatility characteristic parameters are integrated to generate a gridded regional benchmark model library, providing a unified benchmark for the spatiotemporal alignment of multi-source data.

[0081] Real-time sensor monitoring data retrieves the volatility characteristic parameters of the corresponding grid in the grid model library. Theoretical gas concentration values ​​are inverted based on Fick's diffusion law and the Arrhenius equation. The residual sequence between the real-time data and the theoretical values ​​is calculated using a sliding window method. If the residual mean or variance consistently exceeds preset tolerances, the finite element physical field simulation module is triggered. Combining historical data from the grid model library with real-time temperature and humidity distributions, it solves the gas diffusion partial differential equation and generates a three-dimensional simulated field dataset containing temperature gradients and diffusion paths. The simulated data is then applied using a spatiotemporal interpolation algorithm to fill in monitoring blind spots and reconstruct a smooth three-dimensional physical field distribution, providing simulation input for subsequent data fusion.

[0082] Real-time sensor data and three-dimensional simulated field data are fed into a dual-channel processing module. The first channel uses a Kalman filter to predict the state of the real-time data, suppressing noise interference and generating a state estimate. The second channel extracts the physical field characteristics of the simulated data using a graph neural network. The graph structure is constructed based on grid-encoded topological relationships, and node features include temperature, concentration, and diffusion flux parameters. The output data from both channels are mapped into a latent space. The Euclidean distance between the real-time state estimate and the simulated feature vector is calculated as a similarity measure. A genetic algorithm is used to iteratively optimize the covariance matrix weights of the Kalman filter and the sparsity of the graph neural network adjacency matrix to minimize the latent space distance. The resulting state vector is then fused to generate a confidence-labeled environmental state vector. This vector is split into real-time and simulated data streams, respectively, and fed into a temporal convolutional network and a long short-term memory network to extract short-term fluctuations and long-term evolution characteristics. After weighted fusion using an attention mechanism, it is coupled with thermodynamic parameters from the gridded model library for analysis to generate a risk deviation indicator.

[0083] The risk deviation indicator drives gradient descent optimization of the material diffusion coefficient and thermodynamic critical value, updates model library parameters, and synchronizes them with the physical field simulation module. This reconstructs the boundary conditions of the gas diffusion equation and generates corrected three-dimensional simulation field data. Simultaneously, based on Bayesian networks and Monte Carlo sampling, the environmental evolution trajectories of different control strategies are simulated. Command combinations that converge the physical field parameters with the environmental thresholds are selected to generate device control commands. After the commands are executed, sensors provide feedback on environmental parameter changes, recalculate residuals, and calibrate model parameters. Parameter synchronization and dynamic adjustment of data fusion strategies are achieved through message queue middleware, forming a closed-loop feedback link between model self-correction and device control. The grid-coded spatiotemporal indexing mechanism and closed-loop iterative optimization suppress cross-modal data bias, ensuring the dynamic accuracy of warehouse environmental monitoring and the reliability of early warnings.

[0084] The present invention constructs a gridded regional benchmark model, divides the storage area into grid units according to the material storage type and binds the sensor node coordinates, integrates the material volatility characteristic parameters, historical monitoring data and environmental thresholds obtained by laboratory analysis, and generates a material and environment relationship matrix. Grid coding is used as a spatiotemporal alignment benchmark to associate the collection location of multi-source heterogeneous sensor data with the simulation data space topology, solving the problem of spatiotemporal calibration errors of cross-modal data caused by deployment differences of sensor nodes. The residual analysis module dynamically monitors the deviation between real-time sensor data and theoretical gas concentration values. When the residual continues to exceed the preset tolerance, the finite element physical field simulation is triggered to generate a three-dimensional simulation field data set. The historical data in the gridded model is combined to fill the monitoring blind spots, suppressing data loss and noise accumulation caused by environmental interference.

[0085] The dual-channel processing module uses Kalman filtering and graph neural networks to process real-time and simulated data, respectively. A genetic algorithm dynamically optimizes covariance matrix weights and adjacency matrix sparsity, minimizing the latent space similarity metric and improving the fusion consistency between real-time and simulated data. A risk deviation indicator drives gradient descent optimization of material diffusion coefficients and thermodynamic critical values. This combines Bayesian networks with Monte Carlo sampling to select optimal control instructions, forming a closed-loop feedback loop. Parameter updates are synchronized with the physical field simulation and data fusion modules, reconstructing the boundary conditions of the gas diffusion equation and the fusion weight distribution strategy to achieve self-correction of model parameters and dynamic adaptation of environmental thresholds.

[0086] Through a grid-coded spatiotemporal indexing mechanism and closed-loop iterative optimization, this invention effectively suppresses the failure of spatiotemporal alignment of cross-modal data, enhancing the accuracy of dynamic correlation analysis of physical quantities such as temperature and concentration fields. A real-time response mechanism combining virtual and real data complementation, coordinated parameter correction, and equipment control reduces the impact of sudden environmental changes on monitoring reliability, improving the stability and early warning accuracy of monitoring the physical and chemical properties of stored materials.

Claims

1. A multi-source sensing warehouse environment collaborative monitoring and early warning method, characterized by: include: Collect historical environmental monitoring data and laboratory material chemical composition data of the storage area, divide the grid cells according to the material storage type, combine the binding relationship between sensor node coordinates and grid identifiers, receive and set environmental thresholds, and build a gridded regional benchmark model; Calling the material volatility characteristic parameters of the corresponding grid in the gridded regional benchmark model, inverting the theoretical gas concentration value through the material volatility rate equation, performing residual sequence calculation on the real-time collected sensor monitoring data and the theoretical gas concentration value, and triggering a physical field simulation based on the finite element method when the residual continuously exceeds a preset tolerance to generate a three-dimensional simulation field data set; The real-time collected sensor monitoring data and the three-dimensional simulation field data set are processed through a preset dual-channel processing process, wherein: The first processing channel performs Kalman filtering on the sensor monitoring data collected in real time for state prediction; The second processing channel extracts physical field features from the three-dimensional simulation field dataset through a graph neural network; Dynamically optimize the covariance matrix weights of the Kalman filter and the sparsity of the adjacency matrix of the graph neural network using a genetic algorithm, and output an environment state vector; The environmental state vector is split into a real-time data stream and a simulation data stream, which are respectively input into a time series convolutional network to extract short-term fluctuation characteristics and a long short-term memory network to capture the physical field evolution trend. The short-term fluctuation feature vector of the real-time data stream and the physical field evolution trend feature vector of the simulation data stream are weightedly fused through an attention mechanism, and coupled with the material thermodynamic parameters in the gridded regional benchmark model to generate a risk deviation index; Dynamically correct the material diffusion coefficient and thermodynamic critical value in the gridded regional benchmark model based on the risk deviation index, and update the environmental threshold; When the risk deviation index exceeds a preset level, the updated gridded regional benchmark model parameters are called to construct sensor data nodes, material state deduction nodes and equipment control nodes, and generate equipment control instructions.

2. The multi-source sensing storage environment collaborative monitoring and early warning method according to claim 1 is characterized in that: The constructing of the gridded regional benchmark model comprises: Divide the storage area into grid units according to material storage types, and bind sensor node coordinates to grid identifiers; The correlation coefficients between the volatility characteristics of substances and temperature and humidity in the grid cells were extracted by laboratory gas chromatography-mass spectrometry analysis to generate a substance-environment relationship matrix. The substance-environment relationship matrix is ​​associated with historical monitoring data in a preset database according to grid coding to generate a gridded regional benchmark model library including sensor topology, environmental thresholds and substance volatility characteristic parameters.

3. The multi-source sensing storage environment collaborative monitoring and early warning method according to claim 2 is characterized in that: The performing residual sequence calculation on the real-time collected sensor monitoring data and the theoretical gas concentration value includes: Calling the substance volatilization characteristic parameters of the corresponding grid in the gridded regional benchmark model library, and inverting the theoretical gas concentration value according to the substance volatilization rate equation; Calculating a residual sequence between the real-time collected sensor monitoring data and the theoretical gas concentration value, and triggering a physical field simulation based on the finite element method when the residual continues to exceed a preset tolerance; Combining historical data in the gridded regional benchmark model library, a three-dimensional simulation field data set including temperature gradients and gas diffusion paths is generated.

4. The multi-source sensing storage environment collaborative monitoring and early warning method according to claim 3 is characterized in that: The processing of the real-time collected sensor monitoring data and the three-dimensional simulation field data set through a preset dual-channel processing process includes: Performing Kalman filtering on the sensor monitoring data collected in real time to predict the state and generate a real-time data state estimate; Extracting physical field features from the three-dimensional simulation field dataset through a graph neural network to generate a simulation data feature vector; Based on the real-time data state estimation value and the simulation data feature vector, calculating a similarity measure between the real-time data state estimation value and the simulation data feature vector in a latent space; Using the similarity metric as an optimization goal, driving a genetic algorithm to iteratively adjust the covariance matrix weights of the Kalman filter and the sparsity of the adjacency matrix of the graph neural network; According to the optimized covariance matrix weights and adjacency matrix sparsity, the real-time data state estimation value and the simulated data feature vector are fused to output an environment state vector with a confidence label.

5. The multi-source sensing storage environment collaborative monitoring and early warning method according to claim 4 is characterized in that: The generating risk deviation index includes: Splitting the environmental state vector into a real-time data stream and a simulation data stream; Inputting the real-time data stream into a time series convolutional network to extract a first eigenvector representing short-term fluctuations of environmental parameters; Inputting the simulation data stream into a long short-term memory network to extract a second eigenvector representing an evolution trend of the physical field; Performing weighted fusion of the first feature vector and the second feature vector through an attention mechanism to generate a fused feature vector; The fused feature vector is coupled with the material thermodynamic parameters in the gridded regional benchmark model library for analysis, and a risk deviation index representing the degree to which the environmental state deviates from the material stability condition is output.

6. The multi-source sensing warehouse environment collaborative monitoring and early warning method according to claim 5 is characterized in that: The calling of the updated gridded regional benchmark model parameters to construct the sensor data nodes, the material state deduction nodes, and the device control nodes to generate the device control instructions includes: constructing the sensor data nodes, the material state deduction nodes, and the device control nodes in a Bayesian network based on the updated gridded regional benchmark model parameters; Initializing the conditional probability table of the Bayesian network nodes using the Kalman filter covariance matrix weights and the graph neural network adjacency matrix sparsity optimized by the genetic algorithm; The generated three-dimensional simulation field data set and the updated environmental threshold are input into the Bayesian network, and the environmental state change trajectory under different control strategies is simulated through Monte Carlo sampling; screening a control instruction combination that causes the physical field parameters in the three-dimensional simulation field data set to converge with the updated environmental threshold; Device control instructions are generated according to the screening results, and dynamic thresholds in the gridded regional benchmark model parameter library are synchronously updated.

7. The multi-source sensing storage environment collaborative monitoring and early warning method according to claim 6 is characterized in that: Also includes: Collecting environmental parameter feedback data after the execution of the device control instructions in real time through sensors, including temperature change rate, gas concentration gradient and humidity distribution status; Input the feedback data into the residual analysis module, recalculate the theoretical gas concentration value, and perform residual sequence comparison with the measured value in the feedback data; When the statistical characteristics of the residual sequence exceed the environmental threshold of the gridded regional benchmark model library, a data calibration process is triggered to dynamically adjust the material diffusion coefficient and thermodynamic critical value in the model library; The updated model parameters are synchronized to the dual-channel processing process and the physical field simulation module, and the data fusion strategy and three-dimensional simulation field generation logic are reconstructed.

8. The multi-source sensing storage environment collaborative monitoring and early warning method according to claim 7 is characterized in that: Generating the three-dimensional simulation field data set includes: generating initial three-dimensional simulation field data including temperature gradient and gas diffusion path based on the finite element physical field simulation results; Adding a grid code mark to the initial three-dimensional simulated field data, wherein the grid code is aligned with the grid unit identifier and the bound sensor node coordinates according to a spatial topological relationship; In the dual-channel processing, the material volatility characteristic parameters in the gridded regional benchmark model library are matched based on the grid code; Using the substance volatility characteristic parameters as constraints, calculating a similarity measure between the real-time data state estimate and the simulated data feature vector in a latent space; According to the similarity metric optimization result, the physical field parameters of the three-dimensional simulation field data set are synchronously corrected to generate a calibrated three-dimensional simulation field data set that is spatially and temporally aligned with the real-time monitoring data.

9. The multi-source sensing storage environment collaborative monitoring and early warning method according to claim 8 is characterized in that: The dynamically correcting the material diffusion coefficient and thermodynamic critical value in the gridded regional benchmark model based on the risk deviation index and updating the environmental threshold comprises: extracting material thermodynamic equation parameters associated with the current environmental state in the gridded regional benchmark model library based on the generated risk deviation index; According to the value of the risk deviation index, the diffusion coefficient and the thermodynamic critical value in the thermodynamic equation of the substance are iteratively adjusted through a gradient descent algorithm; The updated diffusion coefficient and thermodynamic critical value are transmitted back to the physical field simulation module to reconstruct the boundary conditions of the gas diffusion equation in the process of generating the three-dimensional simulation field data set; a preset finite element physical field simulation method is called to recalculate the gas diffusion path based on the updated diffusion coefficient to generate a revised three-dimensional simulation field data set; The modified three-dimensional simulated field data set is input into the dual-channel processing process to recalculate the latent space similarity measure.

10. The multi-source sensing storage environment collaborative monitoring and early warning method according to claim 9 is characterized in that: The method of dynamically optimizing the covariance matrix weights of the Kalman filter and the sparsity of the adjacency matrix of the graph neural network using a genetic algorithm to output an environmental state vector includes: using the Euclidean distance between the real-time collected sensor monitoring data and the generated three-dimensional simulation field data set in the latent space as a fitness function, and setting the covariance matrix weights of the Kalman filter and the sparsity of the adjacency matrix of the graph neural network as parameters to be optimized; Iteratively adjusting the parameters to be optimized through a genetic algorithm to minimize the latent space distance, and generating an optimized joint parameter set of the Kalman filter and the graph neural network; Feeding back the optimized joint parameter set to the volatilization rate equation inversion process to correct the calculation accuracy of the theoretical gas concentration value; Based on the corrected theoretical gas concentration value, the residual sequence calculation and physical field simulation are re-performed, and the three-dimensional simulation field generation logic of the three-dimensional simulation field data set is updated.

Images