Urban complex environment state prediction method and device based on dynamic multi-map fusion

Abstract

The application relates to a city complex environment state prediction method and device based on dynamic multi-graph fusion. The method proposes a network architecture including a dynamic multi-graph fusion unit, a cross-scale time modeling unit and a joint space-time prediction unit. The dynamic multi-graph fusion unit models the intra-graph and inter-graph dependency relationship in a dynamic manner through a double attention mechanism, integrates multiple graphs including a distance graph, an environment similarity graph, a concentration correlation graph and a learned adaptive graph, the cross-scale time modeling unit models the seasonal change and long-term trend on different time scales, and finally, the joint space-time prediction module effectively captures the complex space-time dependency relationship by using a graph convolution network and a time convolution network. The method can effectively improve the gas concentration prediction accuracy of the comprehensive pipe gallery in the complex environment.

Description

Technical Field

[0001] This application relates to the field of computer information technology, and in particular to a method and apparatus for predicting the state of complex urban environments based on dynamic multi-graph fusion. Background Technology

[0002] As a crucial component of new urban infrastructure, underground utility tunnels are being used more and more widely. They centrally lay municipal trunk lines such as gas, water, electricity, and telecommunications within a single underground corridor. Due to their significant advantages in resource integration, land conservation, and centralized maintenance, they are being adopted by most cities worldwide, redefining urban development patterns. However, the enclosed and narrow environment of these tunnels presents serious safety hazards. In the event of a leak, flammable, explosive, or toxic gases can rapidly accumulate, severely endangering public safety and the stable operation of urban services.

[0003] Historically, past accident cases have shown that traditional passive, reactive safety management measures are often inadequate to handle such emergencies. Therefore, accurately predicting gas concentrations within integrated utility tunnels has become a key technology for proactive disaster prevention and safety assurance. Its aim is to strive for proactive disaster prevention, provide a scientific basis for disaster response decisions, and fundamentally safeguard this vital lifeline of urban stability and safety.

[0004] In recent years, the rapid development of deep learning technology has provided new solutions for the fusion of large-scale, multi-dimensional data. Especially in the processing of unstructured data, large-scale time-series data, and network relationship data, methods based on attention mechanisms and graph neural networks have achieved widespread application and good results. In the scenario of e-commerce and logistics integration, by performing spatiotemporal fusion modeling of multi-source data such as merchants, products, logistics, and industry characteristics, it is expected to more accurately predict and quantify merchant claims risks, providing decision-making references for e-commerce platforms, logistics companies, and insurance institutions. Recent research on gas concentration prediction has increasingly adopted deep learning models, especially using graph neural networks (GNNs) to model the spatiotemporal dependencies of sensor nodes in complex environments (such as urban underground utility tunnels). However, existing methods have two major limitations: First, the static graph structure is flawed, generally relying on fixed distance-based graph structures, making it difficult to capture dynamic, multifaceted spatial interactions caused by factors such as ventilation systems and pipe layouts. Although some studies use multiple predefined graphs to represent different relationships, their structures remain fixed during training, limiting the model's adaptability to dynamic spatial patterns. Secondly, there is a lack of time-series modeling. Although GNN models perform well in traffic and air quality prediction, they mainly focus on autoregressive time dynamics, often ignoring key changes such as short-term fluctuations and long-term trends. This narrow time-series modeling capability weakens the model's robustness in highly dynamic and heterogeneous environments. Summary of the Invention

[0005] Therefore, it is necessary to provide a method and device for predicting the state of complex urban environments based on dynamic multi-graph fusion, which can make accurate predictions in complex environments, in order to address the above-mentioned technical problems.

[0006] A method for predicting the state of complex urban environments based on dynamic multi-graph fusion, the method comprising: Acquire multimodal real-time data, which is time-series data received by multiple sensor nodes installed under the urban underground integrated pipe gallery; Based on real-time multimodal data, distance graph structure, environmental similarity graph structure, and concentration correlation graph structure are constructed respectively. All constructed graph structures are input into the dynamic multi-graph fusion unit. A global graph-level embedding is constructed for each graph structure, and a graph-level embedding is constructed for each node. A graph-aware representation of each node is constructed based on the graph-level embedding and the global graph-level embedding. Intra-graph attention weights and cross-graph attention weights are generated through the spatial attention layer and the graph attention layer based on the graph-aware representation and the global graph-level embedding, respectively. Then, the intra-graph attention weights and cross-graph attention weights are fused using a gating fusion mechanism to generate a fusion similarity matrix. The original gas concentration sequence is input into a cross-scale time modeling unit for scale processing. In the cross-scale time modeling unit, an average pooling layer is used to generate multiple sequence sets with different time resolutions from the original gas concentration sequence. A shared embedding layer is used to map each sequence set into a depth representation. Each depth representation is then split into seasonal and trend components, and a differentiated fusion strategy is used to fuse them to obtain seasonal and trend features. After merging the seasonal and trend features, a time representation feature that comprehensively encodes multi-scale seasonal fluctuations and macro trends is generated. The fused similarity matrix and time representation features are input into the spatiotemporal prediction unit to predict gas concentration, and the gas concentration prediction result is obtained.

[0007] In one embodiment, the multimodal real-time data includes data related to the environment in which each sensor is located and gas concentration data; Before constructing the graph structure, the position data of each sensor that receives the multimodal real-time data are also acquired.

[0008] In one embodiment, when constructing each graph structure based on multimodal real-time data: Based on the location data of each sensor, the adjacency matrix is ​​calculated using a thresholded Gaussian kernel to construct the distance graph structure representing the spatial correlation between sensor nodes based on physical distance; The environmental association is quantified by using the cosine similarity of the data related to the environment of each sensor, and the environmental similarity graph structure that captures the similarity of environmental conditions between nodes is constructed. The data related to the environment of each sensor includes temperature data and humidity data. Using the Pearson correlation of the gas concentration data, a concentration correlation graph structure representing the similarity of gas concentrations between capture nodes is constructed.

[0009] In one embodiment, the construction of a global graph-level embedding for each graph structure and a graph-level embedding for each node includes: By using two fully connected layers to aggregate the node features in each graph structure, a global graph-level embedding for each graph structure is obtained. By using Node2Vec to encode the local neighborhood structure of each node in each graph structure, the graph-level embedding corresponding to each node is obtained.

[0010] In one embodiment, in the dynamic multi-graph fusion unit, after fusing the intra-graph attention weights and cross-graph attention weights using a gating fusion mechanism, the fused data is converted into a compact representation using a dense projection layer, and then the fused data in the compact representation is multiplied by its own transpose matrix to obtain the fusion similarity matrix.

[0011] In one embodiment, in the cross-scale time modeling unit, a differentiated fusion strategy is used to fuse the seasonal component and the trend component respectively, including: For seasonal components, a bottom-up approach is used for mixing, aggregating fine-grained changes to guide coarser-grained seasonal characteristics. For trend components, a top-down approach is used for mixing, utilizing higher-level stable macro trends to guide local time structures.

[0012] In one embodiment, the spatiotemporal prediction unit includes a plurality of spatiotemporal blocks stacked sequentially and a multilayer perceptron, wherein each spatiotemporal block consists of a graph convolutional network and a temporal convolutional network. The preset spatial representation features and the temporal representation features are extended into spatiotemporal representation features. In the spatiotemporal block, the spatiotemporal representation features and the fused similarity matrix are processed at each time step using a graph convolutional network to obtain spatially enhanced representation features. The spatial augmented representation features of each node are processed along the time axis using the temporal convolutional network to obtain the output data of the spatiotemporal block; After flattening the final output data of the spatiotemporal blocks of multiple stacked structures, the data is input into the multilayer sensor to obtain the gas concentration prediction result.

[0013] In one embodiment, in the spatiotemporal prediction unit: The graph convolutional network is constructed from hybrid jump propagation layers; The temporal convolutional network is constructed from a dilated convolutional initial layer.

[0014] This application also provides a device for predicting the state of complex urban environments based on dynamic multi-graph fusion, the device comprising: The real-time data acquisition module is used to acquire multimodal real-time data, which is time-series data received by multiple sensor nodes set under the urban underground integrated pipe gallery; The graph structure construction module is used to construct distance graph structure, environmental similarity graph structure, and concentration correlation graph structure based on multimodal real-time data. The dynamic multi-graph fusion module is used to input all the constructed graph structures into the dynamic multi-graph fusion unit. It constructs a global graph-level embedding for each graph structure and a graph-level embedding for each node. It constructs a graph-aware representation for each node based on the graph-level embedding and the global graph-level embedding. It generates intra-graph attention weights and cross-graph attention weights based on the graph-aware representation and the global graph-level embedding through the spatial attention layer and the graph attention layer, respectively. Then, it uses a gating fusion mechanism to fuse the intra-graph attention weights and cross-graph attention weights to generate a fusion similarity matrix. The cross-scale time modeling module is used to input the original gas concentration sequence into the cross-scale time modeling unit for scale processing. In the cross-scale time modeling unit, the original gas concentration sequence is used to generate a series of sequences with different time resolutions using an average pooling layer. Each series of sequences is mapped to a depth representation using a shared embedding layer. Each depth representation is then split into seasonal and trend components, and these components are fused using a differentiated fusion strategy to obtain seasonal and trend features. The seasonal and trend features are then merged to generate a time representation feature that comprehensively encodes multi-scale seasonal fluctuations and macro trends. The gas concentration prediction module is used to input the fused similarity matrix and time representation features into the spatiotemporal prediction unit to predict the gas concentration and obtain the gas concentration prediction result.

[0015] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the specific steps in the above-described method for predicting the state of complex urban environments based on dynamic multi-graph fusion.

[0016] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the specific steps of the above-described method for predicting the state of complex urban environments based on dynamic multi-graph fusion.

[0017] The aforementioned method and apparatus for predicting the state of complex urban environments based on dynamic multi-graph fusion utilizes multimodal real-time data received from multiple sensor nodes located under urban underground utility tunnels. Based on this data, distance graph structures, environmental similarity graph structures, concentration correlation graph structures, and adaptive graph structures are constructed. All constructed graph structures are input into a dynamic multi-graph fusion unit. A global graph-level embedding is constructed for each graph structure, and a graph-level embedding is constructed for each node. A graph-aware representation of each node is constructed based on the graph-level embedding and the global graph-level embedding. Intra-graph attention weights and cross-graph attention weights are generated through a spatial attention layer and a graph attention layer, respectively, based on the graph-aware representation and the global graph-level embedding. Finally, a gating fusion mechanism is used to adjust the intra-graph attention weights and cross-graph attention weights. After attention weights are fused, a fusion similarity matrix is ​​generated. The original gas concentration sequence is then input into a cross-scale time modeling unit for scaling. Within this unit, an average pooling layer generates multiple sequence sets with different time resolutions from the original gas concentration sequence. A shared embedding layer maps each sequence set to a depth representation, which is then decomposed into seasonal and trend components. These components are fused using a differentiated fusion strategy to obtain seasonal and trend features. These features are then merged to generate a time representation feature that comprehensively encodes multi-scale seasonal fluctuations and macro-trends. Finally, the fusion similarity matrix and the time representation feature are input into a spatiotemporal prediction unit for gas concentration prediction, yielding the predicted gas concentration result. The spatiotemporal architecture proposed in this method enables the invention to comprehensively capture the complex dynamic patterns of gas concentration evolution in underground utility tunnels across space and time, thereby achieving accurate gas concentration prediction. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a method for predicting the state of complex urban environments based on dynamic multi-graph fusion in one embodiment. Figure 2 This is a schematic diagram of the overall framework of the neural network architecture proposed by the method in one embodiment. Figure 3 This is a schematic diagram of a dynamic multi-graph attention mechanism in one embodiment. Figure 3 (a) is a schematic diagram of the spatial attention layer. Figure 3 (b) is a schematic diagram of the structure of the attention layer; Figure 4 This is a schematic diagram of a cross-scale seasonal and trend blending learning framework in one embodiment; Figure 5 This is a schematic diagram of the structure of the extended inception layer and the hybrid jump propagation layer in one embodiment, wherein, Figure 5(a) shows a schematic diagram of the multi-scale dilated convolutional layer structure in a temporal convolutional network. Figure 5 (b) shows a schematic diagram of the hybrid jump propagation layer structure in a graph convolutional network; Figure 6 This is a schematic diagram comparing the prediction curves of AirPhyNet and our method on hydrogen sulfide test data in an experiment. Figure 6 (a) and Figure 6 (b) Schematic diagrams showing the test data of sensor 2 and sensor 17 respectively; Figure 7 This is a structural block diagram of an urban complex environment state prediction device based on dynamic multi-graph fusion in one embodiment; Figure 8 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0020] In this embodiment, as Figure 1 As shown, a method for predicting the state of complex urban environments based on dynamic multi-graph fusion is provided, which specifically includes the following steps: Step S100: Obtain multimodal real-time data, which is time-series data received by multiple sensor nodes set up under the urban underground integrated pipe gallery.

[0021] Step S110: Construct distance graph structure, environmental similarity graph structure, and concentration correlation graph structure based on multimodal real-time data.

[0022] Step S120: Input all the constructed graph structures into the dynamic multi-graph fusion unit. Construct a global graph-level embedding for each graph structure and a graph-level embedding for each node. Construct a graph-aware representation for each node based on the graph-level embedding and the global graph-level embedding. Generate intra-graph attention weights and cross-graph attention weights based on the graph-aware representation and the global graph-level embedding through the spatial attention layer and the graph attention layer, respectively. Then, use the gating fusion mechanism to fuse the intra-graph attention weights and cross-graph attention weights to generate a fusion similarity matrix.

[0023] In step S130, the original gas concentration sequence is input into the cross-scale time modeling unit for scale processing. In the cross-scale time modeling unit, the original gas concentration sequence is used to generate a series of sequences with different time resolutions using an average pooling layer. The shared embedding layer is used to map each series of sequences into a depth representation. Each depth representation is then split into seasonal components and trend components. After fusion using a differentiated fusion strategy, seasonal features and trend features are obtained. After merging the seasonal features and trend features, a time representation feature that comprehensively encodes multi-scale seasonal fluctuations and macro trends is generated.

[0024] Step S140: Input the fused similarity matrix and time representation features into the spatiotemporal prediction unit to predict gas concentration and obtain the gas concentration prediction result.

[0025] In this method, a neural network architecture is proposed, which specifically includes a dynamic multi-graph fusion unit (step S120), a cross-scale temporal modeling unit (step S130), and a spatiotemporal prediction unit (step S140), as follows: Figure 2 As shown.

[0026] In step S100, the multimodal real-time data includes data related to the environment of each sensor and gas concentration data. For example, the environment-related data includes temperature and humidity data of the environment where each sensor is located, and historical gas concentration data. The gas categories mainly include hydrogen sulfide, oxygen, sulfur dioxide, etc. At the same time, the location data of each sensor receiving the multimodal real-time data and the distance between each sensor node are also acquired.

[0027] In step S110, when constructing each graph structure based on the multimodal real-time data: based on the position data of each sensor, the adjacency matrix is ​​calculated using a thresholded Gaussian kernel to construct a distance graph structure representing the spatial correlation between sensor nodes based on physical distance. By utilizing the cosine similarity measure of environmental data related to each sensor, environmental association is quantified, and an environmental similarity graph structure is constructed to capture the similarity of environmental conditions between nodes. The data related to the environment of each sensor includes temperature and humidity data. Furthermore, Pearson correlation of gas concentration data is used to construct a concentration correlation graph structure representing the similarity of gas concentrations between capture nodes. In fact, in all the graph structures mentioned above, the nodes all refer to sensor nodes.

[0028] In this embodiment, during the training process, the dynamic spatial correlation between learnable nodes is used to construct an adaptive graph structure.

[0029] Specifically, distance structure diagram This represents the spatial correlation between sensor nodes based on physical distance. The adjacency matrix is ​​calculated using a thresholded Gaussian kernel, and the distance structure graph... The weights of the edges between nodes are represented as follows: ; In the above formula, It is a node and The Euclidean distance between them To control the distribution, and Define sparsity.

[0030] Specifically, the environmental similarity structure diagram Capture the similarity of environmental conditions between nodes, such as temperature and humidity. The edge weights are calculated as follows: ; In the above formula, Represents a node i The environmental feature vector.

[0031] Specifically, the structure of the concentration correlation plot To capture the similarity of gas concentration at nodes, the edge weights are calculated as follows: ; In the above formula, Represents a node and Pearson correlation between concentration sequences.

[0032] Specifically, adaptive graph The dynamic spatial correlations between nodes are learned during training. Adaptive weights are learned in the following way: ; In the above formula, It is a trainable node embedding, and α is a scaling factor.

[0033] In step S120, within the dynamic multi-graph fusion unit, the dynamic multi-graph attention mechanism consists of two parts: spatial attention, used to capture node-level correlations within each graph, and graph attention, used to model the interactions between different graph structures, such as... Figure 3 As shown in the above structural diagram. Each node in Generate a graph-aware representation .

[0034] In this embodiment, constructing a global graph-level embedding for each graph structure and a graph-level embedding for each node includes: aggregating the node features in each graph structure using two fully connected layers to obtain the global graph-level embedding for each graph structure. By using Node2Vec to encode the local neighborhood structure of each node in each graph structure, the graph-level embedding corresponding to each node is obtained. Meanwhile, in the k-th attention block, the input representation is denoted as... , among which the figure Middle node The hidden state is denoted as .

[0035] In this embodiment, the input to the dynamic multi-graph attention mechanism is a graph-aware representation. as well as Based on these two inputs, first calculate and These two attention scores are then processed by Softmax to obtain the corresponding attention score. and Then based on The attention scores P and Q are obtained, and finally, the two are fused through a gating attention mechanism. The fusion result is then combined with... This is then used as input for the next attention block. The output of the final layer is transformed by a dense projection layer and fed into the subsequent prediction module. Specifically, this is done to learn the spatial correlations within each structure graph, i.e., the similarity between all nodes in each structure graph. In the k-th attention block, the structure graph... Middle node The in-graph attention output is denoted as And use an attention mechanism for computation: ; In the above formula, Except The set of all nodes except those mentioned above. The attention weights are represented by... The result obtained using SoftMax reflects the nodes in the same graph. right The attention scores are derived based on their hidden states and multi-graph embeddings. .

[0036] Furthermore, to better capture the correlations between different graphs, a spatial attention layer is used to target the graphs. Each node in Calculate the attention-weighted sum across all construction graphs: ; In the above formula, It is attention weight, which is determined by... The quantization image obtained through SoftMax. Relative to the diagram Middle node The relative importance of [the factors]. The correlation score is calculated as follows: .

[0037] In this embodiment, in the dynamic multi-graph fusion unit, after the intra-graph attention weights and cross-graph attention weights are fused using the gating fusion mechanism, the fused data is converted into a compact representation using a dense projection layer. Then, the fused data in the compact representation is multiplied by its own transpose matrix to obtain the fusion similarity matrix.

[0038] Furthermore, the dynamic multi-graph fusion unit is composed of multiple stacked attention blocks. Each attention block contains a graph attention block and a spatial attention block. Therefore, the result of the gated fusion is used as the input to the next attention block. Figure 3 The output shown is the result of the last layer, so it is directly input into the final prediction module.

[0039] Specifically, a gating fusion mechanism is applied to integrate in-graph attention weights. Cross-graph attention weights And as the next attention module The input is then processed. Finally, a dense projection layer is used to process the output of the final dynamic multi-graph attention module. Convert to compact representation Then by... Multiply by its transpose to calculate the fusion similarity matrix ,Right now This matrix encodes the comprehensive spatial correlation between all nodes by aggregating structural knowledge from multiple graphs.

[0040] In step S130, in order to effectively capture the dynamic time pattern of the gas concentration sequence (time series) in the underground integrated pipe gallery, a multi-scale time representation is first generated from the original input (i.e., the historical gas concentration sequence) in the cross-scale time modeling unit.

[0041] In this embodiment, given past gas concentration observations That is, historical data, among which, L It is the sequence length. C This refers to the type of gas or the number of sensor variables; applying average pooling will... x downsampling S Different time resolutions: ; Here, the lowest level sequence It preserves the finest temporal details, while higher-level sequences (e.g.) This focuses on long-term or coarse-grained variations. Then, each of these multi-scale sequences is mapped to a latent space through a shared embedding layer to obtain a set of deep representations: ; Next, to further extract interpretable temporal structure from multi-scale representations in the underground utility tunnel scenario, seasonal trend decomposition is performed on each time series at different scales within the cross-scale time modeling unit. Specifically, for the first... Time, multi-scale input Using an Autoformer-inspired decomposition function, it is decomposed into seasonal components. and trend components : ;

[0042] Furthermore, considering the different characteristics of seasonal and trend components, different fusion strategies are employed to handle them in cross-scale time modeling units. For example... Figure 4 As shown, the seasonal component is blended in a bottom-up manner, aggregating fine-grained changes to guide coarser-grained seasonality, while the trend component is blended in a top-down manner, utilizing higher-level stable macro trends to guide local time structures. The update rule for the multi-scale representation is as follows: ; In the above formula, This represents a two-layer perceptron with an intermediate GELU activation function, used for cross-channel transformation of fused temporal features. and These represent seasonal and trend-based combined operations, respectively.

[0043] Specifically, for the seasonal component, a bottom-up residual fusion strategy is adopted to capture the periodic hierarchical structure. This strategy is based on... The definition is as follows: ; In the above formula, This is implemented using two time-linear layers with a GELU activation function in between, which directs the input from... Dimension Alignment Dimension.

[0044] Specifically, regarding trend components, for trend components A top-down residual fusion strategy is adopted, which allows coarse-scale trends to guide finer-scale refinement, as shown below:

[0045] In the two formulas above, Trend-Mix ( ) and Season-Mix The identical structure ensures consistency of trends across scales.

[0046] like Figure 2 As shown, in the cross-scale time modeling unit, this dual-stream hybrid structure gradually integrates fine-grained seasonal fluctuations and macro-level trend patterns, enabling accurate time modeling under the complex dynamic conditions of underground utility tunnels. After the cross-scale seasonal and trend components are processed by the CSMix and TXMix modules respectively, the resulting time features are merged and passed through a multilayer perceptron (MLP). This MLP is used to transform the fused features into the final time representation, expressed as... This output As the time input for the subsequent spatiotemporal joint prediction module.

[0047] In step S140, the spatiotemporal prediction unit includes multiple stacked spatiotemporal blocks and a multilayer perceptron. Each spatiotemporal block consists of a graph convolutional network (GCN) and a temporal convolutional network (TCN), which expands preset spatial and temporal representation features into spatiotemporal representation features. Within the spatiotemporal block, the GCN processes the spatiotemporal representation features and the fused similarity matrix at each time step to obtain spatially enhanced representation features. The TCN processes the spatially enhanced representation features along the time axis for each node to obtain the output data of the spatiotemporal block. The final output data of the multiple stacked spatiotemporal blocks is flattened and input into the multilayer perceptron to obtain the gas concentration prediction result.

[0048] In this embodiment, to jointly simulate the temporal dynamics and spatial dependencies in underground public tunnels, a stacked structure of multiple spatiotemporal blocks is adopted in the spatiotemporal joint prediction unit. Each spatiotemporal block consists of a graph convolutional network (GCN) and a temporal convolutional network (TCN).

[0049] Specifically, given spatial input and time input First, they are extended into a unified spatiotemporal representation. For each spacetime block The spacetime block first occurs at the corresponding time step. t Apply GCN operations and use similarity graphs Capture spatial dependencies between nodes. This is in each t Time-based spatial augmentation representation Then, for each node m TCN is applied along the time axis, using stacked dilated convolutional layers (DILs) to capture short-term and long-term temporal dependencies. The formula for each block can be expressed as: ;

[0050] In this embodiment, in each Graph Convolutional Network (GCN), a Hybrid Jump Propagation Layer (MPL) is used to selectively aggregate information from spatial neighbors with different jump distances. The Temporal Convolutional Network (TCN) module is constructed from Dilated Convolutional Initial Layers (DILs), which apply multiple convolutional kernels with different dilation rates and kernel sizes to efficiently extract multi-cycle temporal patterns. These structures are as follows... Figure 5 As shown, where, Figure 5 (a) shows a schematic diagram of the multi-scale dilated convolutional layer structure in a temporal convolutional network. Figure 5 (b) shows a schematic diagram of the hybrid jump propagation layer structure in a graph convolutional network.

[0051] Furthermore, in the spatiotemporal joint prediction unit, after all K After each spacetime block, the final output is... It is flattened and fed into a multilayer perceptron (MLP) to generate the final prediction. This spatiotemporal architecture enables the present invention to comprehensively capture the complex dynamic patterns of gas concentration evolution in underground utility tunnels in both space and time.

[0052] This paper also demonstrates the effectiveness of the above method through experiments. Two different datasets were used in the experiments: the H2S dataset and the PUB dataset. The H2S dataset contains hydrogen sulfide gas data collected by 28 sensors located in different compartments of an underground utility tunnel in a city, along with related environmental data such as temperature and humidity. The PUB dataset is a large-scale air quality dataset containing data from 92 air quality monitoring stations, used to assess hourly PM2.5 concentrations in a province. The H2S dataset covers the period from January 1st of a given year to December 31st of the following year, while the PUB dataset covers the period from January 1st to May 31st of a given year. Data collection was performed hourly for both datasets. The H2S dataset contains hydrogen sulfide gas concentrations collected by sensors in different compartments of an underground utility tunnel in a city. The PUB dataset contains PM2.5 concentrations collected at various stations in a city over a given year. The training, validation, and test sets for both datasets were divided in a 7 / 1 / 2 ratio. For missing values ​​in the dataset, the mean method is used for imputation to ensure that outliers detected by the underground tunnel sensors do not interfere with model performance.

[0053] In the experiment, two commonly used metrics were used to evaluate model performance: mean absolute error (MAE) and root mean square error (RMSE). This set of metrics covers a range of characteristics related to different aspects of gas concentration prediction. Their definitions are as follows: ; ; In the above formula, This represents the observed value of environmental factors at the i-th time point in the integrated utility tunnel, while Represents the predicted value at the same point in time. Variable This represents the total number of points in time.

[0054] Meanwhile, in the experiments, our method was compared with several commonly used methods and state-of-the-art time series prediction models, including: statistical methods: using the statistical characteristics of historical data to predict future concentrations, including the Autoregressive Integral Moving Average (ARIMA), Vector Autoregressive (VAR), and Exponential Smoothing (ES); heuristic methods: simple but effective time series prediction methods, including DLinear-I and DLineaer-S, iTransformer, PatchTST, and TimeMixer; and graph neural network (GNN) based methods: predictors that utilize the powerful spatiotemporal modeling capabilities of spatiotemporal graph neural networks, including the Spatiotemporal Information Diffusion Model (STID), the Spatiotemporal Attention Enhanced Transformer (STAEformer), the Spatiotemporal Prediction Graph Neural Network (STPGNN), and AirPhyNet.

[0055] For the gas concentration prediction task, as shown in Table 1, our method (DMGFL) achieves state-of-the-art performance across all evaluation metrics and prediction durations. In particular, compared to traditional statistical methods such as ARIMA, VAR, and ES, DMGFL significantly reduces the mean absolute error (MAE) and root mean square error (RMSE), indicating that deep learning-based methods are best suited for learning nonlinear and complex spatiotemporal patterns in underground utility tunnel environments. On the H2S dataset, compared to the recent deep learning-based baseline method AirPhyNet, DMGFL improves MAE by approximately 6.33% and RMSE by 5.41% over a 12-hour prediction duration; over a longer 24-hour prediction duration, MAE and RMSE are significantly improved by 7.95% and 6.44%, respectively. This suggests that the DMGFL method best reflects long-term time dependencies and is most useful for predictions on longer timescales. Similarly, on the PUB dataset, DMGFL consistently outperforms all baseline methods, including state-of-the-art deep learning competitors such as iTransformer, DLinear, PatchTST, and TimeMixer. Compared to AirPhyNet, the previously reported best-performing method, DMGFL reduces MAE and RMSE by 7.41% and 9.77%, respectively, for 12-hour prediction durations; and improves MAE and RMSE by 3.71% and 4.59%, respectively, for 24-hour prediction durations. These results further highlight DMGFL's advantages in capturing gas concentration pattern modeling, particularly when dealing with rapid short-term fluctuations and long-term trends.

[0056] Table 1

[0057] Furthermore, while Transformer-based methods such as PatchTST, iTransformer, and STAEformer are competitive with classical statistical methods, they still lag significantly behind DMGFL. For example, in predicting 12-hour durations on the PUB dataset, the DMGFL model outperforms the popular Transformer-based PatchTST method by approximately 9.8% on MAE and 10.6% on RMSE. This clearly demonstrates that, compared to traditional Transformer-based architectures, our method's dynamic multi-graph fusion and cross-scale learning framework possess stronger representational capabilities for modeling complex spatiotemporal correlations. Moreover, among baseline deep learning methods, models based on simple linear layers, such as DLinear-I and DLinear-S, outperform their more complex Transformer-based counterparts.

[0058] However, even compared to these simpler yet effective methods, DMGFL still achieves significant improvements, further demonstrating the robustness and effectiveness of the proposed framework. Finally, our method is compared with several state-of-the-art spatiotemporal models, including STID, STAEformer, STPGNN, and AirPhyNet. AirPhyNet exhibits relatively strong performance, particularly on the PUB dataset (e.g., a MAE of 7.46531 and an RMSE of 9.48051 for 12-hour predictions), highlighting its ability to leverage graph-based inductive bias. However, DMGFL further surpasses AirPhyNet, demonstrating its superior ability to integrate multiple spatial dependencies and capture dynamic temporal patterns. Similarly, while STID and STAEformer improve upon traditional models by incorporating spatial location encoding and attention mechanisms, they still fall short compared to DMGFL. STPGNN achieves promising results by modeling spatiotemporal correlations through graph convolutions, but lacks explicit multi-graph fusion and cross-scale feature modeling. Overall, these comprehensive experimental results demonstrate that our method consistently performs well across diverse datasets and prediction durations. The significant reduction in MAE and RMSE demonstrates its superior ability to handle the complex spatiotemporal prediction tasks inherent in predicting gas concentrations in urban underground utility tunnels.

[0059] Furthermore, to demonstrate the effectiveness of the present invention (DMGFL) in capturing dynamic temporal patterns, two representative sensors, namely sensor 2 and sensor 17, were selected from the two datasets mentioned above to visualize and analyze the prediction results compared with the state-of-the-art model AirPhyNet.

[0060] like Figure 6 As shown in (a), sensor 2 is used as an example to illustrate DMGFL's ability to accurately capture short-term abrupt changes. Within the shadow prediction region, gas concentrations exhibit significant fluctuations, reflecting sudden changes in environmental conditions. DMGFL successfully tracked these abrupt changes, providing predictions that closely match the actual situation. In contrast, AirPhyNet tends to smooth these fluctuations, resulting in significant prediction delays and lower accuracy in short-term responses. DMGFL's superior performance in capturing short-term abrupt changes can be attributed to its integrated cross-scale temporal modeling, which effectively extracts and models transient temporal dynamics.

[0061] Figure 6(b) highlights sensor 17 to demonstrate DMGFL's ability to accurately predict long-term trends. Within the shaded area, gas concentration values ​​exhibit a clear and consistent trend, reflecting gradual environmental changes. Throughout the prediction range, DMGFL consistently tracks the actual trend, maintaining high prediction accuracy even in long-term prediction scenarios. In contrast, AirPhyNet shows deviations in its predictions of actual trends, particularly during periods of prolonged trend duration. DMGFL's robust trend-following capability is primarily attributed to its dynamic multi-map fusion module, which dynamically integrates spatial relationships to enhance the prediction of duration patterns. These comparative analyses highlight DMGFL's robust performance and validate its comprehensive ability to effectively capture both short-term changes and long-term trends.

[0062] The aforementioned method for predicting the state of complex urban environments based on dynamic multi-graph fusion first constructs multiple prior graphs based on domain knowledge of urban underground integrated pipe corridors. These include a spatial proximity graph derived from physical distance, an environmental similarity graph based on shared environmental attributes, and a concentration correlation graph capturing the similarity of gas concentration patterns. Furthermore, during training, a learned adaptive graph is introduced to capture the dynamic correlations between nodes. To effectively integrate information from these different graphs, a dual-attention mechanism combining spatial and graph attention is employed to dynamically fuse multi-graph representations. Secondly, to address the dynamic temporal patterns of gas concentration data, a cross-scale temporal feature extraction and fusion mechanism is designed. This involves seasonal trend decomposition of time-series data, followed by customized feature alignment and fusion strategies, enabling the model to more accurately capture short-term pattern changes and long-term trends at different time scales. Finally, graph convolutional networks (GCNs) and temporal convolutional networks (TCNs) are combined to construct a unified spatiotemporal prediction unit. This unit effectively fuses spatial and temporal representations to improve the accuracy of future gas concentration predictions.

[0063] Compared to existing technologies, this method is the first to study a spatiotemporal prediction model based on dynamic multi-graph fusion for predicting gas concentration in urban underground utility tunnels, thus advancing the research frontier of this emerging field. By introducing a model based on dynamic multi-graph fusion to capture the complex spatial relationships between sensor nodes and incorporating cross-scale correlation learning, the dynamic temporal patterns of gas concentration sequences are modeled, comprehensively capturing the complex dynamic patterns of gas concentration evolution in underground utility tunnels in both space and time.

[0064] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0065] In one embodiment, such as Figure 7 As shown, a device for predicting the state of complex urban environments based on dynamic multi-graph fusion is provided, comprising: a real-time data acquisition module 200, a graph structure construction module 210, a dynamic multi-graph fusion module 220, a cross-scale time modeling module 230, and a gas concentration prediction module 240, wherein: The real-time data acquisition module 200 is used to acquire multimodal real-time data, which is time-series data received by multiple sensor nodes set under the urban underground integrated pipe gallery; Graph structure construction module 210 is used to construct distance graph structure, environmental similarity graph structure and concentration correlation graph structure respectively based on multimodal real-time data; The dynamic multi-graph fusion module 220 is used to input all the constructed graph structures into the dynamic multi-graph fusion unit, construct a global graph-level embedding for each graph structure, and construct a graph-level embedding for each node. Based on the graph-level embedding and the global graph-level embedding, a graph-aware representation of each node is constructed. Through the spatial attention layer and the graph attention layer, intra-graph attention weights and cross-graph attention weights are generated based on the graph-aware representation and the global graph-level embedding, respectively. Then, the intra-graph attention weights and cross-graph attention weights are fused using a gating fusion mechanism to generate a fusion similarity matrix. The cross-scale time modeling module 230 is used to input the original gas concentration sequence into the cross-scale time modeling unit for scale processing. In the cross-scale time modeling unit, the original gas concentration sequence is used to generate a set of sequences with different time resolutions by using an average pooling layer. Each set of sequences is mapped to a depth representation by using a shared embedding layer. Each depth representation is then split into seasonal components and trend components, and seasonal features and trend features are obtained by fusing them using a differentiated fusion strategy. After merging the seasonal features and trend features, a time representation feature that comprehensively encodes multi-scale seasonal fluctuations and macro trends is generated. The gas concentration prediction module 240 is used to input the fused similarity matrix and time representation features into the spatiotemporal prediction unit to predict the gas concentration and obtain the gas concentration prediction result.

[0066] Specific limitations regarding the urban complex environment state prediction device based on dynamic multi-map fusion can be found in the limitations of the urban complex environment state prediction method based on dynamic multi-map fusion mentioned above, and will not be repeated here. Each module in the aforementioned urban complex environment state prediction device based on dynamic multi-map fusion can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0067] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 8 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a method for predicting gas concentration in integrated utility tunnels based on dynamic multi-image fusion. The display screen can be an LCD screen or an e-ink display screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0068] Those skilled in the art will understand that Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0069] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps: Acquire multimodal real-time data, which is time-series data received by multiple sensor nodes installed under the urban underground integrated pipe gallery; Based on real-time multimodal data, distance graph structure, environmental similarity graph structure, and concentration correlation graph structure are constructed respectively. All constructed graph structures are input into the dynamic multi-graph fusion unit. A global graph-level embedding is constructed for each graph structure, and a graph-level embedding is constructed for each node. A graph-aware representation of each node is constructed based on the graph-level embedding and the global graph-level embedding. Intra-graph attention weights and cross-graph attention weights are generated through the spatial attention layer and the graph attention layer based on the graph-aware representation and the global graph-level embedding, respectively. Then, the intra-graph attention weights and cross-graph attention weights are fused using a gating fusion mechanism to generate a fusion similarity matrix. The original gas concentration sequence is input into a cross-scale time modeling unit for scale processing. In the cross-scale time modeling unit, an average pooling layer is used to generate multiple sequence sets with different time resolutions from the original gas concentration sequence. A shared embedding layer is used to map each sequence set into a depth representation. Each depth representation is then split into seasonal and trend components, and a differentiated fusion strategy is used to fuse them to obtain seasonal and trend features. After merging the seasonal and trend features, a time representation feature that comprehensively encodes multi-scale seasonal fluctuations and macro trends is generated. The fused similarity matrix and time representation features are input into the spatiotemporal prediction unit to predict gas concentration, and the gas concentration prediction result is obtained.

[0070] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Acquire multimodal real-time data, which is time-series data received by multiple sensor nodes installed under the urban underground integrated pipe gallery; Based on real-time multimodal data, distance graph structure, environmental similarity graph structure, and concentration correlation graph structure are constructed respectively. All constructed graph structures are input into the dynamic multi-graph fusion unit. A global graph-level embedding is constructed for each graph structure, and a graph-level embedding is constructed for each node. A graph-aware representation of each node is constructed based on the graph-level embedding and the global graph-level embedding. Intra-graph attention weights and cross-graph attention weights are generated through the spatial attention layer and the graph attention layer based on the graph-aware representation and the global graph-level embedding, respectively. Then, the intra-graph attention weights and cross-graph attention weights are fused using a gating fusion mechanism to generate a fusion similarity matrix. The original gas concentration sequence is input into a cross-scale time modeling unit for scale processing. In the cross-scale time modeling unit, an average pooling layer is used to generate multiple sequence sets with different time resolutions from the original gas concentration sequence. A shared embedding layer is used to map each sequence set into a depth representation. Each depth representation is then split into seasonal and trend components, and a differentiated fusion strategy is used to fuse them to obtain seasonal and trend features. After merging the seasonal and trend features, a time representation feature that comprehensively encodes multi-scale seasonal fluctuations and macro trends is generated. The fused similarity matrix and time representation features are input into the spatiotemporal prediction unit to predict gas concentration, and the gas concentration prediction result is obtained.

[0071] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0072] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0073] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for predicting the state of complex urban environments based on dynamic multi-graph fusion, characterized in that: The method includes: Acquire multimodal real-time data, which is time-series data received by multiple sensor nodes installed in the urban underground integrated pipe gallery. The multimodal real-time data includes data related to the environment of each sensor and gas concentration data. The data related to the environment of each sensor includes temperature data and humidity data. Based on multimodal real-time data, a distance graph structure, an environment similarity graph structure, a concentration correlation graph structure, and an adaptive graph structure are constructed respectively. Specifically, the distance graph structure, representing the spatial correlation between sensor nodes based on physical distance, is constructed by calculating the adjacency matrix using a thresholded Gaussian kernel based on the location data of each sensor. The environment similarity graph structure, which captures the similarity of environmental conditions between nodes, is constructed by quantifying the environmental association using the cosine similarity of the data related to the environment of each sensor. The concentration correlation graph structure, which captures the similarity of gas concentration between nodes, is constructed by using the Pearson correlation of the gas concentration data. The adaptive graph structure is constructed by assigning trainable node embeddings to each sensor node and calculating adaptive edge weights between nodes based on the similarity between the node embeddings to learn the dynamic spatial correlation between nodes. All constructed graph structures are input into the dynamic multi-graph fusion unit. A global graph-level embedding is constructed for each graph structure, and a graph-level embedding is constructed for each node. A graph-aware representation of each node is constructed based on the graph-level embedding and the global graph-level embedding. Intra-graph attention weights and cross-graph attention weights are generated through the spatial attention layer and the graph attention layer based on the graph-aware representation and the global graph-level embedding, respectively. Then, the intra-graph attention weights and cross-graph attention weights are fused using a gating fusion mechanism to generate a fusion similarity matrix. The original gas concentration sequence is input into a cross-scale time modeling unit for scale processing. In this unit, an average pooling layer is used to generate multiple sequence sets with different time resolutions. A shared embedding layer maps each sequence set to a depth representation. Each depth representation is then decomposed into seasonal and trend components, which are fused using a differentiated fusion strategy to obtain seasonal and trend features. These features are then merged to generate a time representation feature that comprehensively encodes multi-scale seasonal fluctuations and macroeconomic trends. Specifically, the cross-scale time modeling unit employs a differentiated fusion strategy to fuse the seasonal and trend components, including: For the seasonal component, a bottom-up residual fusion strategy is adopted to satisfy: In the above formula, The seasonal component is represented by Season-Mix, which is a seasonal fusion module implemented using two time-linear layers with a GELU activation function set between them. For the trend component, a top-down residual fusion strategy is adopted to satisfy: In the above formula, The trend component is represented by Trend-Mix, which is a trend fusion module. The trend fusion module has the same structure as the seasonality fusion module. The fused similarity matrix and time representation features are input into the spatiotemporal prediction unit to predict gas concentration, and the gas concentration prediction result is obtained.

2. The method for predicting the state of complex urban environments based on dynamic multi-graph fusion according to claim 1, characterized in that, Before constructing the graph structure, the position data of each sensor that receives the multimodal real-time data are also acquired.

3. The method for predicting the state of complex urban environments based on dynamic multi-graph fusion according to claim 2, characterized in that, The construction of a global graph-level embedding for each graph structure and a graph-level embedding for each node includes: Two fully connected layers are used to aggregate the node features in each graph structure to obtain the global graph-level embedding for each graph structure. The local neighborhood structure of each node in each graph structure is encoded using Node2Vec to obtain the graph-level embedding corresponding to each node.

4. The method for predicting the state of complex urban environments based on dynamic multi-graph fusion according to claim 3, characterized in that, In the dynamic multi-graph fusion unit, after fusing the intra-graph attention weights and cross-graph attention weights using a gating fusion mechanism, the fused data is converted into a compact representation using a dense projection layer. Then, the fused data in the compact representation is multiplied by its own transpose matrix to obtain the fusion similarity matrix.

5. The method for predicting the state of complex urban environments based on dynamic multi-graph fusion according to any one of claims 1-4, characterized in that, The spatiotemporal prediction unit includes multiple spatiotemporal blocks stacked sequentially and a multilayer perceptron, wherein each spatiotemporal block consists of a graph convolutional network and a temporal convolutional network. The preset spatial representation features and the temporal representation features are extended into spatiotemporal representation features. In the spatiotemporal block, the spatiotemporal representation features and the fused similarity matrix are processed at each time step using a graph convolutional network to obtain spatially enhanced representation features. The spatial augmented representation features of each node are processed along the time axis using the temporal convolutional network to obtain the output data of the spatiotemporal block; After flattening the final output data of the spatiotemporal blocks of multiple stacked structures, the data is input into the multilayer sensor to obtain the gas concentration prediction result.

6. The method for predicting the state of complex urban environments based on dynamic multi-graph fusion according to claim 5, characterized in that, In the spatiotemporal prediction unit: The graph convolutional network is constructed from hybrid jump propagation layers; The temporal convolutional network is constructed from a dilated convolutional initial layer.

7. A device for predicting the state of complex urban environments based on dynamic multi-graph fusion, characterized in that, The device implements the urban complex environment state prediction method based on dynamic multi-graph fusion as described in any one of claims 1-6, including: The real-time data acquisition module is used to acquire multimodal real-time data, which is time-series data received by multiple sensor nodes set under the urban underground integrated pipe gallery; The graph structure construction module is used to construct distance graph structure, environmental similarity graph structure, and concentration correlation graph structure based on multimodal real-time data. The dynamic multi-graph fusion module is used to input all the constructed graph structures into the dynamic multi-graph fusion unit. It constructs a global graph-level embedding for each graph structure and a graph-level embedding for each node. It constructs a graph-aware representation for each node based on the graph-level embedding and the global graph-level embedding. It generates intra-graph attention weights and cross-graph attention weights based on the graph-aware representation and the global graph-level embedding through the spatial attention layer and the graph attention layer, respectively. Then, it uses a gating fusion mechanism to fuse the intra-graph attention weights and cross-graph attention weights to generate a fusion similarity matrix. The cross-scale time modeling module is used to input the original gas concentration sequence into the cross-scale time modeling unit for scale processing. In the cross-scale time modeling unit, the original gas concentration sequence is used to generate a series of sequences with different time resolutions using an average pooling layer. Each series of sequences is mapped to a depth representation using a shared embedding layer. Each depth representation is then split into seasonal and trend components, and then fused using a differentiated fusion strategy to obtain seasonal and trend features. After merging the seasonal and trend features, a time representation feature that comprehensively encodes multi-scale seasonal fluctuations and macro trends is generated. The gas concentration prediction module is used to input the fused similarity matrix and time representation features into the spatiotemporal prediction unit to predict the gas concentration and obtain the gas concentration prediction result.

Images