Emergency monitoring state estimation method based on graph-enhanced spatiotemporal kriging model

Abstract

The application discloses an emergency monitoring state estimation method based on a graph-enhanced spatiotemporal Kriging model, which comprises the following steps: obtaining the pollutant concentration of each sensor at a plurality of sample observation positions within a sample time window; adding noise to the pollutant concentration of the masked sample observation position within the sample time window through a to-be-trained diffusion model to obtain a noisy pollutant concentration; predicting the noise in the noisy pollutant concentration of the masked sample observation position through the to-be-trained diffusion model and denoising to reconstruct the pollutant concentration corresponding to the masked sample observation position; and training the to-be-trained diffusion model with the objective of minimizing the difference between the added noise and the predicted noise of the to-be-trained diffusion model. In the training process of the to-be-trained diffusion model, the masked sample observation position and the masked sample observation position are dynamically determined. The trained diffusion model is used to predict the pollutant concentration of an unobserved position where no sensor is deployed.

Description

Technical Field

[0001] This application belongs to the field of model training technology, specifically involving an emergency monitoring state estimation method based on a graph-augmented spatiotemporal kriging model. Background Technology

[0002] With the acceleration of urbanization and the popularization of IoT sensing technology, solar energy systems, location-based social networks for air quality monitoring, and various monitoring devices in transportation networks have been widely deployed in key areas, forming observation networks covering key regions. However, due to equipment deployment costs, geographical limitations, and communication constraints, monitoring systems cannot achieve full-area, all-time coverage, leaving many areas without directly installed sensors. How to accurately assess and provide risk warnings for conditions in these unobserved locations based on real-time data from limited observation locations has become a problem that needs to be solved.

[0003] Solar energy systems, location-based social networks for air quality monitoring, and observational data from transportation networks exhibit significant spatiotemporal coupling characteristics. In the temporal dimension, a specific observational indicator (such as pollutant concentration, population density, or temperature change) shows continuous dynamic fluctuations as events evolve. In the spatial dimension, adjacent regions experience spatial interaction effects due to physical proximity, environmental connectivity, or propagation path dependence.

[0004] This spatiotemporal correlation means that the observed data exhibit autocorrelation and cross-correlation in both the spatiotemporal dimensions, making it difficult for traditional independent modeling methods to effectively capture its complex evolutionary patterns. Meanwhile, existing dynamic graph modeling methods often rely on the similarity of sequence data for dynamic correlation capture. This paradigm contradicts the assumption that node data in character scenes is unobserved, thus creating application challenges.

[0005] Currently, methods for spatiotemporal interpolation and prediction based on historical observation data are mainly applied to emergency observation tasks such as pollution diffusion simulation, infectious disease transmission prediction, and traffic flow inference. Existing technologies typically rely on common graph convolutional network models, using static matrices or fixed adjacency relationships to characterize spatial dependencies. However, in real emergency scenarios, the strength of spatial correlations often evolves dynamically over time—for example, pollutants undergo asymmetric diffusion due to changes in wind direction, and population flow paths change as the event progresses. Simplifying dynamic spatial relationships to static topology ignores key spatiotemporal interaction mechanisms, leading to biased estimations of the state of unobserved locations and consequently, an inability to accurately obtain data for unobserved locations. Summary of the Invention

[0006] To address the aforementioned technical problems, this application provides an emergency monitoring status estimation method based on a graph-enhanced spatiotemporal kriging model.

[0007] In a first aspect, embodiments of this application provide an emergency monitoring state estimation method based on a graph-enhanced spatiotemporal kriging model, the method comprising: The pollutant concentrations collected by sensors at multiple sample observation locations within the sample time window were obtained. Based on the semantic associations between the multiple sample observation locations, construct the sample adjacency matrix corresponding to the multiple sample observation locations; From the multiple sample observation positions, a portion of the sample observation positions are randomly selected as masked sample observation positions, and the remaining sample observation positions are used as unmasked sample observation positions. By using the diffusion model to be trained, noise is added to the pollutant concentration at the observation location of the masked sample within the sample time window to obtain the noisy pollutant concentration at the observation location of the masked sample within the sample time window; Given the pollutant concentration at the observation location of the unmasked sample, the noisy pollutant concentration at the observation location of the masked sample, and the sample adjacency matrix, the noise in the noisy pollutant concentration at the observation location of the masked sample is predicted by the diffusion model to be trained, and the noisy pollutant concentration at the observation location of the masked sample is denoised to reconstruct the pollutant concentration at the observation location of the masked sample. With the objective of minimizing the difference between the added noise and the noise predicted by the diffusion model to be trained, the diffusion model to be trained is trained to obtain a trained diffusion model. The trained diffusion model is used to: predict the pollutant concentration at unobserved locations without deployed sensors within the target time window based on the pollutant concentration collected by sensors at multiple target observation locations within the target time window.

[0008] Secondly, embodiments of this application provide an emergency monitoring state estimation device based on a graph-enhanced spatiotemporal kriging model, the device comprising: The sample contaminant concentration acquisition module is used to obtain the contaminant concentration collected by sensors at multiple sample observation locations within the sample time window. The sample adjacency matrix construction module is used to construct the sample adjacency matrix corresponding to the multiple sample observation locations based on the semantic association between the multiple sample observation locations; The sample partitioning module is used to randomly select a portion of the sample observation positions from the multiple sample observation positions as masked sample observation positions, and to use the remaining sample observation positions as unmasked sample observation positions. The noise addition module is used to add noise to the pollutant concentration at the observation location of the masked sample within the sample time window using the diffusion model to be trained, so as to obtain the noisy pollutant concentration at the observation location of the masked sample within the sample time window; The training execution module is used to predict the noise in the noisy pollutant concentration at the observation position of the masked sample using the diffusion model to be trained, under the conditions of the pollutant concentration at the observation position of the unmasked sample, the noisy pollutant concentration at the observation position of the masked sample, and the sample adjacency matrix, and to denoise the noisy pollutant concentration at the observation position of the masked sample in order to reconstruct the pollutant concentration at the observation position of the masked sample. An iterative training module is used to train the diffusion model to be trained with the goal of minimizing the difference between the added noise and the noise predicted by the diffusion model to be trained, so as to obtain a trained diffusion model. The trained diffusion model is used to predict the pollutant concentration at unobserved locations without deployed sensors within the target time window, based on the pollutant concentration collected by sensors at multiple target observation locations within the target time window.

[0009] Thirdly, embodiments of this application provide an electronic device, which includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the steps in the emergency monitoring status estimation method based on a graph-enhanced spatiotemporal kriging model as described in the first aspect.

[0010] Fourthly, embodiments of this application provide a readable storage medium storing a program or instructions that, when executed by a processor, implement the steps in the emergency monitoring state estimation method based on a graph-enhanced spatiotemporal kriging model as described in the first aspect.

[0011] This application provides an emergency monitoring state estimation method, apparatus, device, and medium based on a graph-enhanced spatiotemporal kriging model. The method includes: obtaining pollutant concentrations collected by sensors at multiple sample observation locations within a sample time window; constructing a sample adjacency matrix corresponding to the multiple sample observation locations based on semantic associations between them; randomly selecting a portion of the sample observation locations as masked sample observation locations, and designating the remaining locations as unmasked sample observation locations; adding noise to the pollutant concentrations at the masked sample observation locations within the sample time window using a diffusion model to be trained, thereby obtaining the noisy pollutant concentrations at the masked sample observation locations within the sample time window; and then, in the unmasked sample observation locations... Given the pollutant concentration corresponding to the location, the noisy pollutant concentration corresponding to the masked sample observation location, and the sample adjacency matrix, the noise in the noisy pollutant concentration corresponding to the masked sample observation location is predicted by the diffusion model to be trained, and the noisy pollutant concentration corresponding to the masked sample observation location is denoised to reconstruct the pollutant concentration corresponding to the masked sample observation location. The diffusion model to be trained is trained with the objective of minimizing the difference between the added noise and the noise predicted by the diffusion model to be trained, resulting in a trained diffusion model. The trained diffusion model is used to: predict the pollutant concentration at unobserved locations without deployed sensors within the target time window, based on the pollutant concentrations collected by sensors at multiple target observation locations within the target time window.

[0012] The technical solutions described above enable the training of diffusion models under the premise of training difficulties caused by the lack of historical information. By utilizing the spatial correlation strength between observation locations and known pollutant concentration data of the trained diffusion model set, the pollutant concentration at observation locations without deployed sensors can be accurately predicted within the same time window. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating an emergency monitoring status estimation method based on a graph-enhanced spatiotemporal kriging model provided in an embodiment of this application. Figure 2 This is an overall framework diagram of an emergency monitoring state estimation method based on a graph-enhanced spatiotemporal kriging model provided in an embodiment of this application; Figure 3 This is a schematic diagram illustrating the prediction of air pollutant index quality in a certain region using a diffusion model, as provided in this application. Figure 4 This is a schematic diagram of the framework of an emergency monitoring state estimation device based on a graph-enhanced spatiotemporal kriging model provided in one embodiment of this application; Figure 5A schematic diagram of the hardware structure of an electronic device to implement an embodiment of this application. Detailed Implementation

[0014] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0015] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0016] This application proposes a diffusion model for emergency observation, addressing the needs for state assessment and alarm for unobserved locations without deployed sensors. It delves into the dynamic evolution of spatiotemporal dependencies to estimate the state of unobserved locations. Taking regional pollution concentration observation and early warning as an example, the correlation between adjacent observation locations is not static but dynamically changes over time due to time-varying environmental factors such as prevailing wind direction and diffusion conditions. Traditional static graph models struggle to capture such dynamic spatial correlations. Therefore, this application uses a dynamic graph to capture these complex correlations. To ensure the feasibility of the dynamic graph approach, this application introduces a diffusion model-based framework to adaptively construct the graph structure. During inference, this framework alternately performs conditional probability estimation of the unobserved location state and selection of observation locations for graph-enhanced dynamic neighborhoods: the diffusion model provides more reasonable conditional probability estimates based on data from relevant observation locations, while the attention mechanism uses the graph structure and existing estimates to select more accurate dynamic neighborhoods. These two processes iterate alternately, actively learning the spatiotemporal state of relevant neighboring observation locations. The diffusion model's structured noise scheduling mechanism effectively guides the model to measure uncertainty during iteration, avoiding errors in dynamic neighborhood selection caused by uncertainty, and achieving efficient integration of wide-area information. Furthermore, by employing a rescaling strategy based on observation location weights, the model enhances its generalization ability from training to inference in real-world complex scenarios. Through learning dynamic spatial correlations, this application can more stably reconstruct the global risk field, improving the accuracy of inferences about the risk status of hidden areas and the timeliness of alarms, thus providing a reliable decision-making basis for emergency resource scheduling and risk prevention and control.

[0017] The following description, in conjunction with the accompanying drawings, details an emergency monitoring status estimation method based on a graph-enhanced spatiotemporal kriging model provided in this application, through specific embodiments and application scenarios.

[0018] Figure 1 This is a flowchart illustrating an emergency monitoring state estimation method based on a graph-enhanced spatiotemporal kriging model, according to an embodiment of this application. Figure 2 This is an overall framework diagram of an emergency monitoring status estimation method based on a graph-enhanced spatiotemporal kriging model provided in one embodiment of this application.

[0019] refer to Figure 1 This application provides an emergency monitoring state estimation method based on a graph-enhanced spatiotemporal kriging model, the method comprising steps S11 to S16: Step S11: Obtain the pollutant concentrations collected by the sensors at multiple sample observation locations within the sample time window.

[0020] In this embodiment, during the training phase of the diffusion model, multiple sample observation locations are pre-set in multiple different geographical locations. The sample observation location refers to the location where pollutant concentration needs to be observed. Each sample observation location is equipped with a sensor. Through the pre-set sensor, the pollutant concentration at the sample observation location where the sensor is located is collected within each sample time window (e.g., hourly, daily), thereby obtaining the pollutant concentration at each sample observation location within each sample time window.

[0021] The pollutant concentrations at multiple sample observation locations collected within each sample time window are used as a set of training data for the diffusion model to be trained.

[0022] Step S12: Based on the semantic association between the multiple sample observation locations, construct the sample adjacency matrix corresponding to the multiple sample observation locations.

[0023] In this embodiment, the sample adjacency matrix is ​​used to represent the semantic association between multiple sample observation locations. Each element in the sample adjacency matrix represents the spatial association strength between the two sample observation locations corresponding to that element. Through the sample adjacency matrix, the semantic association between each sample observation location can be obtained. The closer the semantic association between two sample observation locations, the stronger the spatial association between the two sample observation locations. Conversely, the farther the semantic association between two sample observation locations, the weaker the spatial association between the two sample observation locations.

[0024] For example, during the training phase, it includes If there are 100 sample observation locations, then this The sample adjacency matrix corresponding to each sample observation location is A1. Wherein, for the sample adjacency matrix A1 (corresponding to Figure 2 The A in this context is distinct from the elements in the target adjacency matrix A2 during the reasoning phase. This indicates the spatial correlation strength between the observation location of the i-th sample and the observation location of the j-th sample.

[0025] Step S13: Randomly select a portion of the sample observation positions from the multiple sample observation positions as masked sample observation positions, and use the remaining sample observation positions as unmasked sample observation positions.

[0026] In this embodiment, in order to simulate unobserved locations where no sensors are deployed and pollutant data cannot be collected during actual applications, multiple sample observation locations can be randomly divided into masked sample observation locations and unmasked sample observation locations during the training phase. Masked sample observation locations refer to sample observation locations where the collected sensor data needs to be masked to simulate observation locations where no sensors are deployed (corresponding to unobserved locations in the inference phase). Unmasked sample observation locations refer to sample observation locations where no masking of the collected sensor data is required to simulate observation locations where sensors are deployed (corresponding to target observation locations in the inference phase).

[0027] It should be noted that the division between unmasked sample observation positions and masked sample observation positions among multiple sample observation positions is dynamic in different training rounds during the training phase.

[0028] Specifically, in the emergency monitoring network for pollutant concentrations used to train the diffusion model during the training phase, the spatial dependencies between multiple sample observation locations are represented by a graph structure. The depiction includes, Indicates all Each sample observation location (i.e., the graph structure corresponding to each sample observation location) V1 is a set of nodes (in a given set), V1 includes Observation locations of unmasked samples and Observation location of masked sample and satisfy and A1 This is the sample adjacency matrix.

[0029] Step S14: Using the diffusion model to be trained, noise is added to the pollutant concentration at the observation location of the masked sample within the sample time window to obtain the noisy pollutant concentration at the observation location of the masked sample within the sample time window.

[0030] In this embodiment, after dividing multiple sample observation locations and generating a sample adjacency matrix, it is necessary to add noise to the pollutant concentration of the masked sample observation locations within the sample time window through the diffusion model to be trained, so that the actual pollutant concentration of the masked sample observation locations in each training data is masked within the sample time window and represented as a noisy pollutant concentration.

[0031] Step S15: Under the conditions of the pollutant concentration corresponding to the observation position of the unmasked sample, the noisy pollutant concentration corresponding to the observation position of the masked sample, and the sample adjacency matrix, the noise in the noisy pollutant concentration corresponding to the observation position of the masked sample is predicted by the diffusion model to be trained, and the noisy pollutant concentration corresponding to the observation position of the masked sample is denoised to reconstruct the pollutant concentration corresponding to the observation position of the masked sample.

[0032] In this embodiment, the pollutant concentration corresponding to the unmasked sample observation position, the noisy pollutant concentration corresponding to the masked sample observation position, and the sample adjacency matrix are input into the diffusion model to be trained. The diffusion model can learn the spatial correlation strength between multiple sample observation positions between the sample adjacency matrices corresponding to the graph structure and gradually optimize the noise prediction result of the noisy pollutant concentration at the masked sample observation position. In this process, the most associated sample observation position with the masked sample observation position is dynamically selected by combining the attention mechanism to predict the noise of the noisy pollutant concentration and perform noise reduction to reconstruct the pollutant concentration corresponding to the masked sample observation position.

[0033] This simulates the process by which the diffusion model, in actual inference, uses the spatial correlation strength between the actual pollutant concentration collected within a time window and the observation location where the actual pollutant concentration is located, and the unobserved location where pollutant concentration prediction is needed, to predict the pollutant concentration at the same time window where no sensors are deployed.

[0034] Step S16: With the goal of minimizing the difference between the added noise and the noise predicted by the diffusion model to be trained, the diffusion model to be trained is trained to obtain a trained diffusion model. The trained diffusion model is used to: predict the pollutant concentration at unobserved locations without deployed sensors within the target time window based on the pollutant concentration collected by sensors at multiple target observation locations within the target time window.

[0035] In this embodiment, the diffusion model to be trained is trained with the goal of minimizing the difference between the noise added by the diffusion model to the pollutant concentration at the mask sample observation position within the sample time window in step S14 and the noise predicted by the diffusion model to be trained. When the training objective is met, the diffusion model to be trained is determined to be completed, and the trained diffusion model is obtained. This model is then used to predict the pollutant concentration in unobserved areas where no sensors are deployed in a real-world application scenario.

[0036] The diffusion model to be trained is trained by iteratively training it with multiple sets of data and various construction methods. Starting from step S13, each input data is randomly masked and noise is added, and multiple rounds of training are performed to improve the model's ability to model the distribution of the most likely clean data among the noisy data under any given noise.

[0037] Specifically, such as Figure 2 As shown in the overall framework diagram, the framework mainly includes the following key components: (a) represents the noise addition and training process of the diffusion model to be trained during the training phase; (b) Represents the inference phase of the trained diffusion model and the optimization process after adjustment using a scaling strategy based on the change in the number of observation locations (see subsequent embodiments for the inference phase). (c) represents the processing structure within the diffusion model, which is based on... As input, This represents the known conditions used in model prediction. This indicates that the object that the diffusion model needs to predict, i.e. the target of noise reduction, is unknown and needs to be predicted by the model.

[0038] During the training phase, Corresponding observation position of unmasked sample .

[0039] During the reasoning stage, Pollutant concentration at the target observation location where sensors were deployed The diffusion model uses noise prediction As output Indicates in Under these conditions, the predicted noise can be predicted by noise reduction. The corresponding pollutant concentration.

[0040] Specifically, during the training phase, the process of adding noise to the pollutant concentration at the observation location of the masked sample within the sample time window is as follows: Figure 2As shown in part (a) of the figure, “①”, “②”, and “③” are nodes, i.e., sample observation positions. Among them, “①” is the masked sample observation position, and “②” and “③” are the unmasked sample observation positions. From left to right, they are sequentially connected by “+”. Noise is added k times in the manner described above.

[0041] After adding noise, the noise is then processed from right to left using the denoising function of the diffusion model. Noise removal is performed on the concentration of noisy pollutants using k diffusion steps.

[0042] The process of training a diffusion model can be formalized as follows: And through a conditional denoising function To achieve prediction during the training phase, where, Represents the identity matrix. This represents the predicted noise variance (generally, the fixed variance from the noise addition process is directly applied). ), , This represents the noise coefficient used in the process of constructing noisy data for the model.

[0043] The training objective of the diffusion model to be trained is to minimize the difference between the added noise and the noise predicted by the diffusion model to be trained, as shown in the following formula: ; in, Represents the loss function. Represents the mathematical expectation, for any pair of pairs that follow a standard normal distribution ( Noise with a mean of 0 and a unit covariance matrix And the expected number of diffusion steps k, This refers to adding noise to the pollutant concentration at the observation location of the masked sample within the sample time window during the k-th diffusion step. The process This represents the raw pollutant concentration at the masked sample observation location within the sample time window. This is the attenuation factor in the k-th diffusion step, used to control the initial value. The retention ratio, In accordance with The added noise term increases with the number of diffusion steps k. The less information is retained.

[0044] The process of the final trained diffusion model during the inference phase can be formalized as follows: And through a conditional denoising function To achieve prediction during the reasoning stage, It emphasizes the adjustment of scaling strategies based on changes in the number of observation locations during the inference phase.

[0045] The technical solutions described above enable the training of diffusion models under the premise of training difficulties caused by the lack of historical information. By utilizing the spatial correlation strength between observation locations and known pollutant concentration data of the trained diffusion model set, the pollutant concentration at observation locations without deployed sensors can be accurately predicted within the same time window.

[0046] In conjunction with the technical solutions of the above embodiments, an embodiment of this application also provides another emergency monitoring state estimation method based on a graph-enhanced spatiotemporal kriging model. In this method, step S15, "predicting the noise in the noisy pollutant concentration corresponding to the masked sample observation location under the conditions of the pollutant concentration at the unmasked sample observation location, the noisy pollutant concentration at the masked sample observation location, and the sample adjacency matrix," specifically includes steps S15-1 to S15-6: Step S15-1: The pollutant concentrations corresponding to the observation positions of the unmasked samples and the noisy pollutant concentrations corresponding to the observation positions of the masked samples are processed by the diffusion model to be trained to obtain the pollutant concentration representation of the first sample.

[0047] In this embodiment, the diffusion model to be trained takes the pollutant concentration corresponding to the observation position of the unmasked sample and the noisy pollutant concentration corresponding to the observation position of the masked sample as input and combines them with a self-attention mechanism to obtain the first sample pollutant concentration representation. In the case of n nodes, the sample pollutant concentration is represented as an n*n matrix.

[0048] Step S15-2: Determine the scaling factor based on the diffusion step number k; the diffusion step number k is the number of noise addition operations performed on the pollutant concentration corresponding to the observation position of the mask sample.

[0049] In this embodiment, in order for the diffusion model to be trained to adaptively learn how to apply the sample adjacency matrix in the predefined graph structure G1... The spatial correlation strength is fused into the attention matrix corresponding to the pollutant concentration of the first sample, and a scaling factor is introduced. , Represented by diffusion step embedding Scaling factors dynamically generated by linear layers , This indicates a linear projection.

[0050] Step S15-3: According to the scaling factor, the sample adjacency matrix is ​​fused with the first sample pollutant concentration representation through the diffusion model to be trained to obtain the second sample pollutant concentration representation.

[0051] In this embodiment, the diffusion model to be trained is combined with a scaling factor. This bounded matrix The Q-embedding representation of sample time t and the K-embedding representation of sample time t are fused to obtain the second sample pollutant concentration representation at sample time t. The process is shown in the following formula:

[0052] By introducing a scaling factor The diffusion model adaptively controls how the sample adjacency matrix A1 in the graph structure G1 is integrated into the attention matrix corresponding to the first sample pollutant concentration representation, and modulates the noise level and prediction refinement by the diffusion step number k, thereby obtaining... The spatial dependencies between the masked sample observation location and other sample observation locations are captured, enabling the graph-enhanced self-attention mechanism in the diffusion model to aggregate pollutant concentration information from sample observation locations related to the unmasked sample observation location during the prediction of noise in the noisy pollutant concentration at the masked sample observation location. In this embodiment, the obtained first sample pollutant concentration representation is fused with the sample adjacency matrix, so that the graph structure information corresponding to the sample adjacency matrix is ​​incorporated into the attention matrix corresponding to the first sample pollutant concentration representation, resulting in a second sample pollutant concentration representation. The second sample pollutant concentration representation aggregates the spatial correlation strength between the masked sample observation location and other sample observation locations from the graph structure information.

[0053] Specifically, in the process of predicting the second sample pollutant concentration representation at the masked sample observation location using the diffusion model, it is necessary to combine the pollutant concentrations of other sample observation locations related to the masked sample observation location. Among them, the pollutant concentration corresponding to the sample observation location with a stronger spatial correlation with the masked sample observation location plays a greater role in generating the second sample pollutant concentration, and vice versa. This allows the spatial dependence between various sample observation locations to be captured, enabling the graph-enhanced self-attention mechanism in the diffusion model to aggregate information on the pollutant concentrations of sample observation locations related to the unmasked sample observation location to be predicted during the prediction of noise in the noisy pollutant concentration at the masked sample observation location.

[0054] Step S15-4: Determine the diffusion step embedding representation based on the diffusion step number k. In this embodiment, the diffusion step number k quantifies the scale by which the diffusion model predicts noise, and the diffusion step embedding representation is as follows: .

[0055] Step S15-5: Determine the embedded representation of the sample time based on any sample time within the sample time window.

[0056] In this embodiment, any sample time t within the sample time window is selected, and a sample time embedding representation is generated based on that sample time t. .

[0057] Step S15-6: Determine the input embedding representation for the sample time based on the pollutant concentration at the observation position of the unmasked sample at the sample time and the noisy pollutant concentration at the observation position of the masked sample at the sample time.

[0058] In this embodiment, the diffusion model to be trained combines the pollutant concentration at the observation location of the unmasked sample at sample time t with the noisy pollutant concentration at the observation location of the masked sample at sample time t to obtain the pollutant concentration at sample time t. And based on the pollutant concentration at sample time t Through function The pollutant concentration at sample time t Mapped to the input embedding representation at sample time t .

[0059] in, For linear projection, Number of diffusion steps The diffusion step embedding representation, The embedding representation is used for the sample time step.

[0060] Step S15-7: Based on the input embedding representation of the sample time, determine the Q embedding representation, the K embedding representation, and the V embedding representation of the sample time; In this embodiment, the input embedding representation for sample time t The Q-embedding representation of sample time t is computed in graph-enhanced self-attention. K-embedding representation of sample time t and the V embedding representation of sample time t The specific calculation method is as follows: ; in, , , These are the weight matrices for QKV, respectively.

[0061] Step S15-8: The transformed sample adjacency matrix is ​​fused with the Q-embedded representation and the K-embedded representation of the sample time to obtain the second sample pollutant concentration representation at the sample time.

[0062] In this embodiment, the diffusion model to be trained uses the transform function to transform the distance-based sample adjacency matrix in the graph structure G1. (A1 corresponds to) Figure 2 In the training phase, A) is transformed into a bounded matrix representing the correlation between nodes in the graph structure (each node corresponds to a sample observation position). Then, for this bounded matrix The Q-embedding representation of sample time t and the K-embedding representation of sample time t are fused to obtain the second sample pollutant concentration representation at sample time t. .

[0063] The process is shown in the following formula:

[0064] in, The first sample pollutant concentration at time t is represented by a bounded matrix. The addition of this feature allows the information from the graph structure G1 to be incorporated during the self-attention computation process. It is an association matrix, which represents the pollutant concentration of the second sample.

[0065] In conjunction with the technical solutions of the above embodiments, an embodiment of this application also provides another pollutant concentration prediction model training method based on a graph-enhanced spatiotemporal kriging model for emergency monitoring status estimation. In this method, the method further includes steps S41 to S42: Step S41: The second pollutant concentration representation is fused with the V embedding representation of the sample time to obtain the third sample pollutant concentration representation of the sample time.

[0066] In this embodiment, the V embedding representation at sample time t is based on the input embedding representation at sample time t. Calculated .

[0067] Step S42: The third sample pollutant concentration representation at the sample time is fused with the input embedding representation at the sample time to obtain the fourth sample pollutant concentration representation at the sample time; In this embodiment, to enhance the stability of the diffusion model to be trained during the training process, a gating fusion mechanism is also incorporated, as shown in the following equation:

[0068] in, For linear transformation, For the Sigmoid function, This represents the pollutant concentration of the third sample at the sample time. For the residual connections of this layer, This represents the pollutant concentration of the fourth sample at the final output time of this layer.

[0069] "Based on the second sample pollutant concentration representation, the noise in the noisy pollutant concentration corresponding to the observation position of the mask sample is predicted by the diffusion model to be trained" specifically includes: using the fourth sample pollutant concentration representation of the sample time output by the (l-1)th layer as the input embedding representation of the sample time of the lth layer, until the fourth sample pollutant concentration representation of the sample time output by the last layer of the diffusion model to be trained is obtained; l is any integer from 1 to L, and L is the number of layers of the diffusion model to be trained; Based on the fourth sample pollutant concentration representation at the sample time point output by the last layer of the diffusion model to be trained, the noise in the noisy pollutant concentration corresponding to the masked sample observation position predicted by the diffusion model to be trained is obtained.

[0070] In this embodiment, as Figure 2 As shown, the diffusion model to be trained is a model containing L layers. Except for the last layer of the model, the pollutant concentration representation of the fourth sample at each sample time of the output of each layer will be used as the input embedding representation of the sample time of the next layer. The diffusion model obtained after training is similar in the application stage, and will not be elaborated further.

[0071] In conjunction with the technical solutions of the above embodiments, an embodiment of this application also provides another emergency monitoring status estimation method based on a graph-enhanced spatiotemporal kriging model. In this method, the trained diffusion model is used to accurately predict the pollutant concentration at unobserved locations without deployed sensors within the target time window based on the pollutant concentration data of multiple target observation locations with deployed sensors within the target time window, as well as the spatial relationship between the observation locations.

[0072] The process specifically includes steps S51 to S54: Step S51: Obtain the pollutant concentrations collected by sensors at multiple target observation locations within the target time window.

[0073] In this embodiment, during the inference phase of the diffusion model, multiple observation locations are pre-set in multiple different geographical locations. The observation location refers to the location where pollutant concentration needs to be observed. The observation locations are divided into target observation locations and unobserved locations. The target observation location refers to the observation location where a sensor is set up, and the unobserved location refers to the observation location where no sensor is set up.

[0074] Sensors are installed at each target observation location. The pollutant concentration at each target observation location within each target time window (e.g., hourly or daily) is collected using the pre-set sensors. This allows us to obtain the pollutant concentration at each target observation location within each target time window.

[0075] Step S52: Construct a target adjacency matrix based on the semantic association between every two locations of the unobserved locations where no sensors are deployed and the multiple target observation locations.

[0076] In this embodiment, the target adjacency matrix is ​​used to represent the semantic association between multiple observation locations. Each element in the target adjacency matrix represents the spatial association strength between the two observation locations corresponding to that element. Through the target adjacency matrix, the semantic association between each observation location can be obtained. The closer the semantic association between two observation locations, the stronger the spatial association between the two observation locations. Conversely, the farther the semantic association between two observation locations, the weaker the spatial association between the two observation locations.

[0077] For example, in the emergency monitoring network for pollutant concentrations during the reasoning phase, it includes... If there are several observation locations, then this The target adjacency matrix corresponding to each observation location is A2. Wherein, for the target adjacency matrix A2 (corresponding to Figure 2 The A in this context is distinct from the elements in the sample adjacency matrix A1 during the inference phase. This indicates the spatial correlation strength between the i-th observation location and the j-th observation location.

[0078] Specifically, The spatial dependencies between observation locations are represented by a graph structure. To depict, among which Indicates all The set of observation locations (each observation location corresponds to a node in graph structure G2), including Target observation location with deployed sensors and Unobserved locations without deployed sensors ,satisfy and .

[0079] Step S53: Initialize the pollutant concentration at unobserved locations without deployed sensors within the target time window to a random pollutant concentration using the trained diffusion model.

[0080] In this embodiment, in order to predict the pollutant concentration at unobserved locations where no sensors are deployed within the target time window, the pollutant concentration at these unobserved locations needs to be set to a random value, i.e., a random pollutant concentration.

[0081] in, For data observed at target observation locations where sensors have been deployed, including data distributed across... The target observation location is at Observation records within a continuous time step, each Represented as related to pollutant concentration A multidimensional feature vector (such as pollutant type, population density, etc.). All observation data can be represented as... L refers to the target time window, while data from unobserved locations without deployed sensors can be represented as... .because The data lacks historical and known information. Therefore, it is necessary to set the data from these unobserved locations as random values, i.e., random pollutant concentrations, so that the trained diffusion model can gradually adjust the random values ​​to obtain the desired results. .

[0082] Step S54: Under the conditions of the pollutant concentrations corresponding to the multiple target observation locations, the random pollutant concentrations corresponding to the unobserved locations, and the target adjacency matrix, the noise in the random pollutant concentrations corresponding to the unobserved locations is predicted using the trained diffusion model, and the noise in the random pollutant concentrations corresponding to the unobserved locations is denoised to predict the pollutant concentrations at the unobserved locations within the target time window.

[0083] In this embodiment, during the inference phase of the trained diffusion model, the random pollutant concentration p( at unobserved locations) is... Starting from this point, based on the pollutant concentrations at multiple target observation locations and the target adjacency matrix, and through K iterations of noise reduction, the pollutant concentrations at unobserved locations within the target time window are finally predicted. .

[0084] That is, using from The target observation location within the target time window Known data within Predicted in At each unobserved location within the target time window Value inside Therefore, this process can be formally expressed as a mapping function. , so that:

[0085] In conjunction with the technical solutions of the above embodiments, an embodiment of this application also provides another emergency monitoring state estimation method based on a graph-enhanced spatiotemporal kriging model. In this method, considering that the attention mechanism has a certain robustness to graph structure changes, the introduction of unobserved positions in the inference process of the diffusion model will lead to an increase in the size of the attention matrix, which may cause the performance degradation problem of the attention distribution tending to be uniform.

[0086] Based on this, this embodiment proposes a scaling strategy based on the change in the number of observation positions. The core idea of ​​this strategy is to scale the attention matrix during the inference phase of the diffusion model to maintain the alignment of the distribution characteristics of the attention matrix during the inference phase with those during the training phase. Since the number of observation positions during the inference phase is greater than that during the training phase, the edge weights may tend to become more uniform. Therefore, an adjustment coefficient is introduced into the attention matrix before calculating the softmax. This can reduce the impact of having more observation locations in the inference phase than in the training phase.

[0087] The step S54, "predicting the noise in the random pollutant concentration corresponding to the unobserved location using the trained diffusion model under the conditions of the pollutant concentration corresponding to the multiple target observation locations, the random pollutant concentration corresponding to the unobserved location, and the target adjacency matrix," specifically includes steps S54-1 to S54-5: Step S54-1: The pollutant concentrations corresponding to the multiple target observation locations and the random pollutant concentrations corresponding to the unobserved locations are processed by the diffusion model to be trained to obtain a representation of the first target pollutant concentration.

[0088] In this embodiment, the trained diffusion model takes the pollutant concentration at the target observation location within the target time window t and the random pollutant concentration at the unobserved location within the target time window t as inputs. By combining the self-attention mechanism, the concentration of the first target pollutant within the target time window t is obtained. .

[0089] Similar to the training process, this process can be represented as: ;

[0090] Where t refers to any moment in the target time window, and T refers to the total time of the target time window.

[0091] Step S54-2: Determine the target scaling factor based on the target diffusion step number k; the target diffusion step number k is: the kth prediction performed on the noise in the random pollutant concentration corresponding to the unobserved location; Step S54-3: Transform the target adjacency matrix according to the target scaling factor, and fuse the transformed target adjacency matrix with the first target pollutant concentration representation to obtain the second target pollutant concentration representation; Step S54-3 specifically includes: determining the target time embedding representation based on any target time within the target time window; determining the input embedding representation of the target time based on the pollutant concentration at the target observation location at the target time and the random pollutant concentration at the unobserved location at the target time; determining the Q embedding representation, K embedding representation, and V embedding representation of the target time based on the input embedding representation of the target time; fusing the transformed target adjacency matrix with the Q embedding representation and K embedding representation of the target time to obtain the second target pollutant concentration representation of the target time.

[0092] Steps S54-2 and S54-3 are similar to the training process.

[0093] Step S54-4: The target adjacency matrix is ​​fused with the first target pollutant concentration representation using the trained diffusion model to obtain the second target pollutant concentration representation; Step S54-5, the total number of the number of unobserved locations and the total number of the plurality of target observed locations. Greater than the total number of the multiple sample observation locations In this case, determine the adjustment coefficient. .

[0094] In this embodiment, when the total number N of multiple observation locations in the inference phase (including the sum of the number of multiple unobserved locations and the number of multiple target observation locations) is greater than the total number of multiple sample observation locations in the training phase... In this case, an adjustment coefficient needs to be introduced. .

[0095] Step S54-4, according to the adjustment coefficient The expression of the second target pollutant concentration is adjusted to obtain the adjusted expression of the second target pollutant concentration.

[0096] In this embodiment, after obtaining the adjustment coefficient, the representation of the second target pollutant concentration is adjusted according to the adjustment coefficient.

[0097] Steps S54-5: The adjusted second target pollutant concentration representation is fused with the V embedding representation of the target time to obtain the third target pollutant concentration representation of the target time; the third target pollutant concentration representation of the target time is fused with the input embedding representation of the target time to obtain the fourth target pollutant concentration representation of the target time; the fourth sample pollutant concentration representation of the target time output by the (l-1)th layer of the trained diffusion model is used as the input embedding representation of the target time by the lth layer, until the fourth target pollutant concentration representation of the target time output by the last layer of the trained diffusion model is obtained; l is any integer from 1 to L, and L is the number of layers in the trained diffusion model; based on the fourth sample pollutant concentration representation of the target time output by the last layer of the trained diffusion model, the noise in the random pollutant concentration corresponding to the unobserved location predicted by the trained diffusion model is obtained.

[0098] Specifically, the adjustment coefficient during the training phase. During the inference phase, the number of all observation locations (including the sum of target observation locations with deployed sensors and unobserved locations without deployed sensors) If the number of sample observation locations (including masked and unmasked sample observation locations) exceeds the number of observation locations during the training phase, then the adjustment coefficient needs to be adjusted. Adjustments are made based on the total number of unobserved locations and multiple observed target locations. and the total number of observation locations for multiple samples Obtained. And according to The adjusted second target pollutant concentration was determined; among which, Given the target adjacency matrix, the trained diffusion model will... After transformation, a bounded matrix is ​​obtained. .

[0099] This avoids the performance degradation problem caused by the uniformization of attention distribution due to the expansion of the attention matrix size when a large number of observation nodes are introduced. Specifically, by introducing an adjustment coefficient λ, the attention distribution is sharpened as the number of observation positions increases, effectively offsetting the uniformization trend caused by the expansion of the graph structure G2. Thus, based on the graph weight distribution (rather than simply node degree), effective alignment of the graph structure for training and inference is achieved, enhancing the generalization ability of the diffusion model in real and complex emergency scenarios.

[0100] To verify the effectiveness of the technical solution presented in this application, this application provides effect visualization and interpretability analysis, specifically... Figure 3 This is a schematic diagram illustrating how the diffusion model trained in this application predicts the air pollutant index quality of a certain region. (Reference) Figure 3 ,right Figure 3 In the prediction process of observation position 16 in (b), the predefined graph only observed the proximity of red and green nodes. However, the diffusion model of this scheme captured multiple related nodes (blue nodes) that were not in the predefined graph structure through the interaction of dynamic graph learning and evaluation. Figure 3 (c) in the figure demonstrates that the nodes captured by this scheme are more similar in sequence features and peak patterns by visualizing the values ​​of adjacent observation positions in the two figures. In the end, this scheme estimates more accurate results compared with the schemes in related technologies.

[0101] This application conducted comprehensive experiments on eight different datasets. These datasets cover multiple application domains: four for transportation systems (METRLA, PEMS-BAY, SEA-LOOP, PEMS07), two for air quality monitoring networks (AQI-36, AQI), and two for solar energy systems (NREL-MD, NREL-AL). The overall performance of the proposed model (DKM-GA) is compared with all baselines (GLTL [Bahadori et al.(2014)], MPGRU [Cini et al.(2022)] and GRIN [Cini et al.(2022)]; KCN [Appleby et al.(2020)], IGNNK [Wu et al.(2022)], LSJSTN [Hu et al.(2024)], INCREASE [Zheng et al.(2023)] and KITS [Xu et al.(2025)]), as shown in Table 1. Three standard evaluation metrics were used for fair comparison: mean absolute error (MAE), mean absolute percentage error (MAPE), and mean relative error (MRE).

[0102] Table 1 demonstrates that the model proposed in this application achieves state-of-the-art performance, with an improvement of up to 12.66% (on AQI-36). We attribute this superior performance to our diffusion-based framework and graph-enhanced attention with a rescaling strategy. The former effectively enhances graph learning by refining values ​​to make them more accurate. The latter captures global and local information to form accurate graphs, enabling fine-grained refinement of Kriging results.

[0103] Table 1

[0104] Figure 4 This is a schematic diagram of the framework of an emergency monitoring state estimation device based on a graph-enhanced spatiotemporal kriging model, provided in one embodiment of this application. (Refer to...) Figure 4 One embodiment of this application provides an emergency monitoring state estimation device based on a graph-enhanced spatiotemporal kriging model, the device comprising: The sample pollutant concentration acquisition module 11 is used to obtain the pollutant concentration collected by the sensors at multiple sample observation locations within the sample time window. The sample adjacency matrix construction module 12 is used to construct the sample adjacency matrix corresponding to the multiple sample observation locations based on the semantic association between the multiple sample observation locations. The sample segmentation module 13 is used to randomly select a portion of the sample observation positions from the plurality of sample observation positions as masked sample observation positions, and to use the remaining sample observation positions as unmasked sample observation positions. The noise addition module 14 is used to add noise to the pollutant concentration at the observation position of the masked sample within the sample time window through the diffusion model to be trained, so as to obtain the noisy pollutant concentration at the observation position of the masked sample within the sample time window. The training execution module 15 is used to predict the noise in the noisy pollutant concentration at the observation position of the masked sample through the diffusion model to be trained, under the conditions of the pollutant concentration at the observation position of the unmasked sample, the noisy pollutant concentration at the observation position of the masked sample, and the sample adjacency matrix, and to denoise the noisy pollutant concentration at the observation position of the masked sample in order to reconstruct the pollutant concentration at the observation position of the masked sample. The iterative training module 16 is used to train the diffusion model to be trained with the goal of minimizing the difference between the added noise and the noise predicted by the diffusion model to be trained, so as to obtain a trained diffusion model. The trained diffusion model is used to predict the pollutant concentration at unobserved locations without deployed sensors within the target time window based on the pollutant concentration collected by sensors at multiple target observation locations within the target time window.

[0105] It should be noted that the emergency monitoring status estimation method based on a graph-augmented spatiotemporal Kriging model provided in this application embodiment can be executed by an emergency monitoring status estimation device based on a graph-augmented spatiotemporal Kriging model, or by a control module within that device for executing the method. This application embodiment uses the execution of the method by the device to load the algorithm as an example to illustrate the emergency monitoring status estimation method based on a graph-augmented spatiotemporal Kriging model provided in this application embodiment.

[0106] The emergency monitoring status estimation device based on the graph-enhanced spatiotemporal kriging model in this application embodiment can be a device, or it can be a component, integrated circuit, or chip in a terminal. The device can be a mobile electronic device or a non-mobile electronic device. For example, mobile electronic devices can be mobile phones, tablets, laptops, PDAs, in-vehicle electronic devices, wearable devices, ultra-mobile personal computers (UMPCs), netbooks, or personal digital assistants (PDAs), etc., while non-mobile electronic devices can be servers, network-attached storage (NAS), personal computers (PCs), televisions (TVs), ATMs, or self-service machines, etc. This application embodiment does not impose specific limitations.

[0107] The emergency monitoring status estimation device based on the graph-enhanced spatiotemporal kriging model in this application embodiment can be a device with an operating system. This operating system can be Android, iOS, or other possible operating systems; this application embodiment does not specifically limit it.

[0108] The emergency monitoring state estimation device based on the graph-enhanced spatiotemporal kriging model provided in this application embodiment can achieve... Figures 1 to 3 The various processes implemented by the emergency monitoring state estimation device based on the graph-enhanced spatiotemporal kriging model in the method embodiment will not be described again here to avoid repetition.

[0109] Optionally, Figure 5 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. This application also provides an electronic device; it should be noted that the electronic device in this application includes the mobile electronic device and non-mobile electronic device described above.

[0110] The electronic device includes, but is not limited to, components such as: radio frequency unit, network module, audio output unit, input unit, sensor, display unit, user input unit, interface unit, memory, and processor.

[0111] Those skilled in the art will understand that electronic devices may also include power supplies (such as batteries) that supply power to various components. The power supply may be connected to the processor logic through a power management system, thereby enabling functions such as managing charging, discharging, and power consumption through the power management system. Figure 5The electronic device structure shown does not constitute a limitation on the electronic device. The electronic device may include more or fewer components than shown, or combine certain components, or have different component arrangements, which will not be elaborated here.

[0112] As an example, such as Figure 5 As shown, the electronic device 600 includes a memory 610 and a processor 620. The memory 610 and the processor 620 are connected via a bus for communication. The memory 610 stores a computer program that can run on the processor 620 to implement the steps in the emergency monitoring status estimation method based on the graph-enhanced spatiotemporal kriging model disclosed in the above embodiments of this application.

[0113] As the apparatus is basically similar to the method embodiment, it is described in a relatively simple way. For relevant details, please refer to the description of the method embodiment.

[0114] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0115] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, embodiments of this application can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects.

[0116] Furthermore, this application embodiment also provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the various processes of the above-described emergency monitoring state estimation method embodiment based on graph-enhanced spatiotemporal kriging model and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0117] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0118] This application embodiment also provides a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor. The processor is used to run programs or instructions to implement the various processes of the above-described emergency monitoring state estimation method embodiment based on graph-enhanced spatiotemporal Kriging model, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0119] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0120] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0121] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0122] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. An emergency monitoring state estimation method based on a graph-enhanced spatiotemporal kriging model, characterized in that, The method includes: The pollutant concentrations collected by sensors at multiple sample observation locations within the sample time window were obtained. Based on the semantic associations between the multiple sample observation locations, construct the sample adjacency matrix corresponding to the multiple sample observation locations; From the multiple sample observation positions, a portion of the sample observation positions are randomly selected as masked sample observation positions, and the remaining sample observation positions are used as unmasked sample observation positions. By using the diffusion model to be trained, noise is added to the pollutant concentration at the observation location of the masked sample within the sample time window to obtain the noisy pollutant concentration at the observation location of the masked sample within the sample time window; Given the pollutant concentration at the observation location of the unmasked sample, the noisy pollutant concentration at the observation location of the masked sample, and the sample adjacency matrix, the noise in the noisy pollutant concentration at the observation location of the masked sample is predicted by the diffusion model to be trained, and the noisy pollutant concentration at the observation location of the masked sample is denoised to reconstruct the pollutant concentration at the observation location of the masked sample. With the objective of minimizing the difference between the added noise and the noise predicted by the diffusion model to be trained, the diffusion model to be trained is trained to obtain a trained diffusion model. The trained diffusion model is used to: predict the pollutant concentration at unobserved locations without deployed sensors within the target time window based on the pollutant concentration collected by sensors at multiple target observation locations within the target time window.

2. The emergency monitoring state estimation method based on a graph-enhanced spatiotemporal kriging model according to claim 1, characterized in that, Given the pollutant concentration at the observation location of the unmasked sample, the noisy pollutant concentration at the observation location of the masked sample, and the sample adjacency matrix, the noise in the noisy pollutant concentration at the observation location of the masked sample is predicted by the diffusion model to be trained, including at least: The pollutant concentrations corresponding to the observation locations of the unmasked samples and the noisy pollutant concentrations corresponding to the observation locations of the masked samples are processed by the diffusion model to be trained to obtain a representation of the pollutant concentration of the first sample. The scaling factor is determined based on the diffusion step number k; the diffusion step number k is the number of noise addition operations performed on the pollutant concentration corresponding to the observation position of the mask sample. The sample adjacency matrix is ​​transformed according to the scaling factor, and the transformed sample adjacency matrix is ​​merged with the first sample pollutant concentration representation to obtain the second sample pollutant concentration representation.

3. The emergency monitoring state estimation method based on a graph-enhanced spatiotemporal kriging model according to claim 2, characterized in that, Given the pollutant concentration at the observation location of the unmasked sample, the noisy pollutant concentration at the observation location of the masked sample, and the sample adjacency matrix, the method of predicting the noise in the noisy pollutant concentration at the observation location of the masked sample using the diffusion model to be trained further includes: Determine the diffusion step embedding representation based on the diffusion step number k; Determine the embedded representation of the sample time based on any sample time within the sample time window; Based on the contaminant concentration at the observation position of the unmasked sample at the sample time and the noisy contaminant concentration at the observation position of the masked sample at the sample time, the input embedding representation at the sample time is determined; Based on the input embedding representation of the sample time, determine the Q embedding representation, the K embedding representation, and the V embedding representation of the sample time; The transformed sample adjacency matrix is ​​fused with the Q-embedding representation and the K-embedding representation of the sample time to obtain the second sample pollutant concentration representation at the sample time.

4. The pollutant concentration prediction model training method according to claim 3, and the emergency monitoring status estimation method based on the graph-enhanced spatiotemporal kriging model, are characterized in that... Given the pollutant concentration at the observation location of the unmasked sample, the noisy pollutant concentration at the observation location of the masked sample, and the sample adjacency matrix, the method of predicting the noise in the noisy pollutant concentration at the observation location of the masked sample using the diffusion model to be trained further includes: The second pollutant concentration representation is fused with the V-embedding representation of the sample time to obtain the third sample pollutant concentration representation at the sample time; The third sample pollutant concentration representation at the sample time is fused with the input embedding representation at the sample time to obtain the fourth sample pollutant concentration representation at the sample time; The fourth sample pollutant concentration at the time of the sample output from the (l-1)th layer is used as the input embedding representation of the sample time from the lth layer until the fourth sample pollutant concentration at the time of the sample output from the last layer of the diffusion model to be trained is obtained; l is any integer from 1 to L, and L is the number of layers of the diffusion model to be trained. Based on the fourth sample pollutant concentration representation at the sample time point output by the last layer of the diffusion model to be trained, the noise in the noisy pollutant concentration corresponding to the masked sample observation position predicted by the diffusion model to be trained is obtained.

5. The emergency monitoring state estimation method based on a graph-enhanced spatiotemporal kriging model according to claim 1, characterized in that, The method further includes: Obtain the pollutant concentrations collected by sensors at multiple target observation locations within the target time window; Based on the unobserved locations where no sensors are deployed and the semantic associations between every two locations among the multiple target observation locations, a target adjacency matrix is ​​constructed; The trained diffusion model initializes the pollutant concentration at unobserved locations without deployed sensors within the target time window as a random pollutant concentration. Given the pollutant concentrations at the multiple target observation locations, the random pollutant concentrations at the unobserved locations, and the target adjacency matrix, the noise in the random pollutant concentrations at the unobserved locations is predicted using the trained diffusion model, and the noise in the random pollutant concentrations at the unobserved locations is denoised to predict the pollutant concentrations at the unobserved locations within the target time window.

6. The emergency monitoring state estimation method based on a graph-enhanced spatiotemporal kriging model according to claim 5, characterized in that, Given the pollutant concentrations at the multiple target observation locations, the random pollutant concentrations at the unobserved locations, and the target adjacency matrix, the noise in the random pollutant concentrations at the unobserved locations is predicted using the trained diffusion model, including at least: The pollutant concentrations corresponding to the multiple target observation locations and the random pollutant concentrations corresponding to the unobserved locations are processed by the diffusion model to be trained to obtain a representation of the first target pollutant concentration. The target scaling factor is determined based on the target diffusion step number k; the target diffusion step number k is the kth prediction performed on the noise in the random pollutant concentration corresponding to the unobserved location. The target adjacency matrix is ​​transformed according to the target scaling factor, and the transformed target adjacency matrix is ​​fused with the first target pollutant concentration representation to obtain the second target pollutant concentration representation. The total number of the unobserved locations and the total number of the plurality of target observed locations. Greater than the total number of the multiple sample observation locations In this case, determine the adjustment coefficient. ; According to the adjustment coefficient The expression of the second target pollutant concentration is adjusted to obtain the adjusted expression of the second target pollutant concentration.

7. The method according to claim 6, characterized in that, Given the pollutant concentrations at the multiple target observation locations, the random pollutant concentrations at the unobserved locations, and the target adjacency matrix, the method of predicting noise in the random pollutant concentrations at the unobserved locations using the trained diffusion model further includes: Determine the target time embedding representation based on any target time within the target time window; The input embedding representation for the target time is determined based on the pollutant concentration at the target observation location at the target time and the random pollutant concentration at the unobserved location at the target time. Based on the input embedding representation of the target time, determine the Q embedding representation, the K embedding representation, and the V embedding representation of the target time; The transformed target adjacency matrix is ​​fused with the Q-embedded representation and the K-embedded representation at the target time to obtain the second target pollutant concentration representation at the target time.

8. The method according to claim 7, characterized in that, Given the pollutant concentrations at the multiple target observation locations, the random pollutant concentrations at the unobserved locations, and the target adjacency matrix, the method of predicting noise in the random pollutant concentrations at the unobserved locations using the trained diffusion model further includes: The adjusted second target pollutant concentration representation is fused with the V-embedding representation of the target time to obtain the third target pollutant concentration representation of the target time; The third target pollutant concentration representation at the target time is fused with the input embedding representation at the target time to obtain the fourth target pollutant concentration representation at the target time; The fourth sample pollutant concentration at the target time, output by the (l-1)th layer of the trained diffusion model, is used as the input embedding representation of the target time at the lth layer, until the fourth target pollutant concentration at the target time is obtained from the output of the last layer of the trained diffusion model; l is any integer from 1 to L, and L is the number of layers in the trained diffusion model. Based on the fourth sample pollutant concentration representation at the target time output by the last layer of the trained diffusion model, the noise in the random pollutant concentration corresponding to the unobserved location predicted by the trained diffusion model is obtained.

9. An emergency monitoring state estimation device based on a graph-enhanced spatiotemporal kriging model, characterized in that, The device includes: The sample contaminant concentration acquisition module is used to obtain the contaminant concentration collected by sensors at multiple sample observation locations within the sample time window. The sample adjacency matrix construction module is used to construct the sample adjacency matrix corresponding to the multiple sample observation locations based on the semantic association between the multiple sample observation locations; The sample partitioning module is used to randomly select a portion of the sample observation positions from the multiple sample observation positions as masked sample observation positions, and to use the remaining sample observation positions as unmasked sample observation positions. The noise addition module is used to add noise to the pollutant concentration at the observation location of the masked sample within the sample time window using the diffusion model to be trained, so as to obtain the noisy pollutant concentration at the observation location of the masked sample within the sample time window; The training execution module is used to predict the noise in the noisy pollutant concentration at the observation position of the masked sample using the diffusion model to be trained, under the conditions of the pollutant concentration at the observation position of the unmasked sample, the noisy pollutant concentration at the observation position of the masked sample, and the sample adjacency matrix, and to denoise the noisy pollutant concentration at the observation position of the masked sample in order to reconstruct the pollutant concentration at the observation position of the masked sample. An iterative training module is used to train the diffusion model to be trained with the goal of minimizing the difference between the added noise and the noise predicted by the diffusion model to be trained, so as to obtain a trained diffusion model. The trained diffusion model is used to predict the pollutant concentration at unobserved locations without deployed sensors within the target time window, based on the pollutant concentration collected by sensors at multiple target observation locations within the target time window.

10. An electronic device, characterized in that, The electronic device includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the steps in the emergency monitoring status estimation method based on a graph-enhanced spatiotemporal kriging model as described in any one of claims 1-8.

Images