Pollutant anomaly detection method based on multi-pollutant coordination and spatio-temporal feature fusion

Abstract

The application discloses a pollution abnormality detection method based on multi-pollutant cooperation and space-time feature fusion, which comprises the following steps: S1, constructing a multi-pollutant sensor network dataset; S2, cleaning and preprocessing the multi-pollutant sensor network dataset in a research area; S3, learning a directed graph adjacency matrix of a single-pollutant sensor network for the multi-pollutant sensor network dataset through Bayesian variational inference, and modeling an asymmetric causal relationship between sensors; S4, processing sensor nodes of a prediction target location by using a single-pollutant space feature extraction module, and performing space feature fusion between multi-pollutants based on a hierarchical attention mechanism; S5, using a space-time feature fusion module to fuse space features and a directed graph for the multi-pollutant sensor network data of the prediction target location, and performing multi-pollutant data prediction; and S6, generating an abnormal score and a dynamic update threshold value of an air pollutant monitoring value based on multi-pollutant cooperation, and performing abnormality detection.

Description

Technical Field

[0001] The present invention belongs to the technical field of urban pollutant data anomaly detection, and specifically relates to a pollutant anomaly detection method based on multi-pollutant collaboration and spatiotemporal feature fusion. Background Art

[0002] With the rapid development of the Internet of Things (IoT) technology, urban environmental monitoring systems continuously collect multidimensional spatiotemporal data (such as PM2.5, SO2, NO2, O3, and CO) through densely deployed sensor networks, providing a massive data foundation for environmental governance and public health management. However, the complex coupling of these data across time and space, as well as the nonlinear synergistic effects among multiple pollutants, present unprecedented challenges for anomaly detection. Existing methods are limited in modeling the synergistic mechanisms and dynamic spatiotemporal dependencies of multiple pollutants, resulting in low detection sensitivity, high false alarm rates, and poor interpretability in complex scenarios, making them unable to meet the needs of smart cities for high-precision environmental monitoring and management.

[0003] Current anomaly detection technologies face multiple bottlenecks in multi-pollutant collaborative analysis, dynamic spatiotemporal modeling, computational efficiency, and interpretability. Although some research has attempted to combine graph neural networks with time series models, their core flaws remain unresolved: First, multi-pollutant data is forcibly compressed into a single graph structure, leading to feature confusion and loss of key signals; second, graph structure learning relies on data-driven similarity metrics and ignores physical correlations (such as meteorological propagation paths), making the model sensitive to noise and difficult to generalize; third, anomaly scoring strategies rely on empirical thresholds and are unable to adapt to environmental changes. Therefore, a new detection framework is urgently needed that can simultaneously meet the following requirements: construct independent propagation path models for different pollutants to avoid feature confusion, while designing cross-pollutant collaborative mechanisms to capture complex pollution events; deeply integrate temporal evolution with spatial diffusion to achieve adaptive topological adjustment through a closed-loop feedback mechanism; reduce computational complexity while ensuring detection accuracy, and enhance decision interpretability through physically guided feature fusion. This technological gap severely restricts the real-time and reliability of smart environmental protection systems. The present invention aims to overcome these limitations and provide a high-precision and robust anomaly detection solution for urban environmental monitoring and management. Summary of the Invention

[0004] To solve the above problems, the present invention proposes a pollutant anomaly detection method based on the synergy of multiple pollutants and the fusion of spatiotemporal features. This method effectively utilizes the multi-pollutant synergy mechanism and dynamic spatiotemporal dependencies to improve the performance of urban pollutant anomaly detection and provide a high-precision and high-robustness anomaly detection solution for urban environmental monitoring and management.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] The pollutant anomaly detection method based on multi-pollutant collaboration and spatiotemporal feature fusion includes the following steps:

[0007] S1, collect data from multiple sensor nodes in the urban pollutant monitoring system and build a multi-pollutant sensor network dataset;

[0008] S2, cleaning and preprocessing of the multi-pollutant sensor network dataset in the study area;

[0009] S3. Using Bayesian variational inference to learn the directed graph adjacency matrix of the single-pollutant sensor network for the processed multi-pollutant sensor network dataset, modeling the asymmetric causal relationship between sensors;

[0010] The specific process of step S3 is:

[0011] S31. Modeling sensor networks as directed graphs ,in, is the set of sensor nodes that monitor pollutant c; is the adjacency matrix of pollutant c;

[0012] S32, through Schur decomposition, the adjacency matrix Decompose into an orthogonal matrix With the lower triangular matrix , combined with Gumbel-Sinkhorn distribution to optimize node sorting;

[0013] S33, introduce variational inference to maximize the lower bound of evidence and learn sparse directed adjacency matrix and eliminate self-loop paths through physical prior constraints;

[0014] S4. The sensor nodes at the predicted target location are processed using a single pollutant spatial feature extraction module. The directed graph-guided graph attention mechanism is used to extract the multi-hop neighborhood spatial features of the single pollutant. The spatial features of multiple pollutants are fused based on the hierarchical attention mechanism.

[0015] The specific process of step S4 is:

[0016] S41, based on multi-hop adjacency matrix Weighted fusion to generate aggregated adjacency matrix , where K is the number of hops; is the learnable attenuation coefficient of pollutant c;

[0017] S42. Computing spatial features of single pollutant sensor network data through a directed graph-guided graph attention mechanism , attention coefficient By adjacency weight Interaction score with features Jointly determined, the calculation formula is: , , ,in, is the spatial characteristics of pollutant c; To correct the linear unit, nonlinear activation is introduced; For pollutant c at node i The weight of attention on oneself; For pollutant c at node i For Node j The attention weight of is a shared linear transformation matrix used for feature mapping; For sensor nodes i In time t Input features of For sensor nodes j In time t Input features of Pollutant c node i The set of neighbors of For pollutant c at node i With node j The feature interaction score of is the activation function; is the transpose of the learnable parameter vector; For sensor nodes i The learnable embedding vector of ; e is the Hadamard product; For nodes i The shared linear transformation matrix of ; For sensor nodes j A learnable embedding vector for For nodes j The shared linear transformation matrix of ; is the adjacency weight of multiple pollutants on the subgraph; For pollutant c at node i With node m The feature interaction score of For sensor nodes m ; For nodes i On the subgraph of pollutant c k Hop neighbor collection; t For time;

[0018] S43. Calculate target pollutants based on historical data p With other pollutants c Pearson correlation coefficient , filter Top- K Strongly correlated pollutant subset ;in, is the covariance; is the observation sequence of target pollutant p; is the observation sequence of target pollutant c; is the standard deviation of the target pollutant p; is the standard deviation of the target pollutant c;

[0019] S44, in the subset Internal, dynamic attention energy , combined with real-time anomaly factors and statistical correlation and adjust the attention weights ,in, is the learnable parameter vector; is the spatial characteristics of the target pollutant p; For connection; is the spatial characteristics of pollutant c; is the balance coefficient; is the attention energy of pollutant c; is the attention energy of pollutant n;

[0020] S45. Generate normalized cross-pollutant attention weights through the Softmax function , weighted fusion of multi-source spatial features, the calculation formula is: ,in, is the fused global spatial feature;

[0021] S5. Use the spatiotemporal feature fusion module to fuse the spatial features and directed graph of the multi-pollutant sensor network data at the predicted target location, perform multi-pollutant data prediction while dynamically adjusting the directed graph structure, and perform dynamic attention adjustment between multiple pollutants based on the results of the multi-pollutant data prediction;

[0022] S6. Anomaly detection is performed based on the abnormal scores and dynamically updated thresholds of the air pollutant monitoring values ​​generated by the collaboration of multiple pollutants.

[0023] Preferably, the specific process of step S1 is as follows: by deploying 45 urban pollutant sensor nodes in the target area, environmental monitoring data for one consecutive year is collected at a frequency of minutes to construct a time series data set containing 8640 time points; each sensor node synchronously records the geographic location, timestamp and concentration values ​​of multiple pollutants to form a multidimensional space-time matrix.

[0024] Preferably, the specific process of step S2 is:

[0025] S21. Use minimum-maximum normalization to eliminate the dimensional differences of different pollutant sensors: ,in, is the normalized pollutant data, and the range is compressed to [0,1]; c is the pollutant type; n Number the sensor; For the n The raw data of pollutant c monitored by a sensor; For the training set n The minimum value of the pollutant c of each sensor; For the training set n The maximum value of the pollutant c of each sensor;

[0026] S22. Introducing an attenuation factor into sliding window time series data , used to suppress noise interference, the calculation formula is: ,in, For pollutant c at time t Sliding window time series data; , , , For the front Term attenuation factor; For the front Term-normalized multi-pollutant data matrix.

[0027] Preferably, the specific process of step S5 is:

[0028] S51, the fused global spatial features With sliding window time series data Splice into joint input ;

[0029] S52. Embedding multi-pollutant adjacency weights in gated recurrent units ,spatiotemporal feature fusion is achieved through topology-aware update gate and reset gate;

[0030] S53. Dynamically adjust cross-pollutant attention weights based on real-time prediction errors , optimize the feature fusion ratio;

[0031] S54, through the multi-task fully connected layer, synchronously generate the predicted values ​​of all pollutants, and use the weighted Huber loss function to optimize the model. The calculation formula is: ,in, For pollutants c In time t The predicted value of the sliding window time series data; is a fully connected layer; is the hidden state output by GRU; e represents the Hadamard product; the weighted Huber loss function is used to balance the optimization objectives of multi-task learning, and the calculation formula is: ,in, is the multi-task loss function; is the task weight coefficient; It is the single-task Huber loss;

[0032] Single-task Huber loss The calculation formula is: ,in, is the length of the training set; is the sliding window size; N c is the number of sensors; is the time series of pollutant c; For pollutants c The Huber loss threshold.

[0033] Preferably, the specific process of step S6 is:

[0034] S61. Calculate target pollutants p Local prediction bias , and generate normalized errors based on median-IQR standardization ;in, For the training set the median; For the training set interquartile range;

[0035] S62, Neighbor Consistency Error Quantization Node The difference in prediction deviation from neighboring nodes is calculated as: ,in, Target pollutants p Neighbor consistency error; Target pollutants p sensor nodes j local prediction bias; is the set of neighbor nodes of node i;

[0036] S63. Quantifying cross-pollutant synergistic bias , reflecting the abnormal consistency between the target pollutant and other pollutants, and the calculation formula is: ,in, For pollutants c sensor nodes i The local prediction deviation; C is the number of pollutant types;

[0037] S64: Weighted fusion of local deviation, neighbor consistency error, and cross-pollutant coordination error to generate the final anomaly score. The calculation formula is: ,in, For sensors i In time t Anomaly score of and are all balance coefficients, determined by grid search on the validation set;

[0038] S65. Calculate the dynamic threshold based on the sliding window quantile. The calculation formula is: ,in, is the dynamic threshold; is the third quartile of the anomaly score within the sliding window; is the dynamic threshold coefficient; is the interquartile range, IQR= - , is the first quartile;

[0039] determination For abnormal data.

[0040] After adopting the above technical solution, the present invention has the following beneficial effects:

[0041] 1. This invention can accurately model the differentiated spatiotemporal propagation characteristics of multiple pollutants: through Bayesian variational inference, an independent directed subgraph is constructed for each pollutant, explicitly characterizing the spatial causal dependencies within the pollutants (such as the difference between the diffusion path of PM2.5 and the chemical conversion path of SO2), overcoming the defect of the traditional single graph structure that confuses the characteristics of multiple pollutants, and significantly improving the ability to trace anomalies in complex pollution scenarios.

[0042] 2. The present invention can enhance the ability to perceive long-range key nodes: it proposes an attention mechanism guided by a multi-hop adjacency matrix, which dynamically fuses the neighborhood features of paths with different hop counts through a learnable attenuation coefficient, suppresses high-hop noise interference while retaining key causal relationship information, and effectively improves the sensitivity of abnormal data under complex topologies.

[0043] 3. The present invention adopts a hierarchical attention mechanism to fuse the collaborative features of multi-source pollutants: a hierarchical attention strategy of "coarse-grained screening → fine-grained allocation" is proposed, a subset of strongly correlated pollutants is screened based on historical statistical correlation, and the fusion weight is dynamically adjusted in combination with real-time anomaly probability to solve the problems of dimensionality explosion and noise interference in multi-pollutant scenarios.

[0044] 4. The present invention adopts multi-task spatiotemporal collaborative prediction and closed-loop weight optimization: a multi-task prediction model is designed to synchronously optimize the spatiotemporal characteristics of target pollutants and related pollutants, embeds adjacency weights through the GRU gating mechanism, and dynamically corrects cross-pollutant attention weights based on prediction errors, forming a "prediction-feedback-adjustment" closed-loop system.

[0045] 5. The present invention adopts multi-dimensional collaborative anomaly scoring and adaptive threshold calibration: it combines local prediction bias, neighbor consistency error and cross-pollutant collaborative error to construct a composite score, and dynamically sets the threshold based on the sliding window quantile to accurately distinguish between sensor failures and real pollution events. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flow chart of the present invention;

[0047] Figure 2 It is a flowchart of the present invention.

[0048] Figure 3 This is a module block diagram of the single pollutant subgraph structure learning algorithm of the present invention;

[0049] Figure 4 This is a block diagram of the directed graph-based cross-pollutant spatiotemporal feature fusion and multi-task prediction algorithm of the present invention;

[0050] Figure 5 This is a block diagram of the multi-pollutant collaborative anomaly scoring and threshold adaptive updating algorithm of the present invention;

[0051] Figure 6 This is a comparison chart of the abnormality detection effects of the present invention. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0053] like Figures 1 to 6 As shown in FIG, the pollutant anomaly detection method based on multi-pollutant collaboration and spatiotemporal feature fusion includes the following steps:

[0054] S1, collect data from multiple sensor nodes in the urban pollutant monitoring system and build a multi-pollutant sensor network dataset;

[0055] The specific process of step S1 is as follows: 45 urban pollutant sensor nodes deployed in the target area collect environmental monitoring data for one year at a frequency of one minute, constructing a time series dataset containing 8640 time points; each sensor node synchronously records the geographic location, timestamp, and concentration values ​​of multiple pollutants to form a multidimensional spatiotemporal matrix;

[0056] S2, cleaning and preprocessing of the multi-pollutant sensor network dataset in the study area;

[0057] The specific process of step S2 is:

[0058] S21, using min-max normalization to eliminate the dimension difference of different pollutant sensors: , wherein, is the normalized pollutant data, the range is compressed to [0, 1]; c is the pollutant type; n is the sensor number; is the original data of the pollutant c monitored by the n th sensor; is the minimum value of the pollutant c of the n th sensor in the training set; is the maximum value of the pollutant c of the n th sensor in the training set;

[0059] S22, introducing a decay factor to the sliding window time series data for suppressing noise interference, the calculation formula is: , wherein, is the sliding window time series data of the pollutant c at time t , , , , is the previous item decay factor; is the previous item normalized multi-pollutant data matrix;

[0060] S3, learning the directed graph adjacency matrix of the single pollutant sensor network through Bayesian variational inference on the processed multi-pollutant sensor network data set, modeling the asymmetric causal relationship between sensors;

[0061] The specific process of step S3 is:

[0062] S31, modeling the sensor network as a directed graph , wherein, is the set of sensor nodes monitoring the pollutant c; is the adjacency matrix of the pollutant c;

[0063] S32, decomposing the adjacency matrix into an orthogonal matrix and a lower triangular matrix , combining Gumbel-Sinkhorn distribution to optimize node sorting;

[0064] S33, introducing variational inference maximum evidence lower bound to learn sparse directed adjacency matrix , and eliminating self-loop paths through physical prior constraints;

[0065] S4. The sensor nodes at the predicted target location are processed using a single pollutant spatial feature extraction module. The directed graph-guided graph attention mechanism is used to extract the multi-hop neighborhood spatial features of the single pollutant. The spatial features of multiple pollutants are fused based on the hierarchical attention mechanism.

[0066] The specific process of step S4 is:

[0067] S41, based on multi-hop adjacency matrix Weighted fusion to generate aggregated adjacency matrix , where K is the number of hops; is the learnable attenuation coefficient of pollutant c;

[0068] S42. Computing spatial features of single pollutant sensor network data through a directed graph-guided graph attention mechanism , attention coefficient By adjacency weight Interaction score with features Jointly determined, the calculation formula is: , , ,in, is the spatial characteristics of pollutant c; To correct the linear unit, nonlinear activation is introduced; For pollutant c at node i The weight of attention on oneself; For pollutant c at node i For Node j The attention weight of is a shared linear transformation matrix used for feature mapping; For sensor nodes i In time t Input features of For sensor nodes j In time t Input features of Pollutant c node i The set of neighbors of For pollutant c at node i With node j The feature interaction score of is the activation function; is the transpose of the learnable parameter vector; For sensor nodes i The learnable embedding vector of ; e is the Hadamard product; For nodes i The shared linear transformation matrix of ; For sensor nodes j A learnable embedding vector for For nodesj shared linear transformation matrix; adjacency weight of the pollutant c on the subgraph G characteristic interaction score of the pollutant c on the node i characteristic interaction score of the pollutant c on the node m characteristic interaction score of the pollutant c on the node sensor node m ; sensor node i characteristic interaction score of the pollutant c on the node k hop neighbor set of the node t time

[0069] S43, calculating the target pollutant p Pearson correlation coefficient with other pollutants c based on historical data , screening Top- K strongly correlated pollutant subset ; wherein, covariance observation sequence of the target pollutant p observation sequence of the target pollutant c standard deviation of the target pollutant p standard deviation of the target pollutant c

[0070] S44, within the subset dynamic attention energy combined with real-time anomaly factor statistical correlation and adjusting attention weight , wherein, learnable parameter vector spatial feature of the target pollutant p connection spatial feature of the pollutant c balance coefficient attention energy of the pollutant c attention energy of the pollutant n

[0071] S45, generating normalized cross-pollutant attention weights by a Softmax function weighted fusion of multi-source spatial features, the calculation formula is: , wherein, global spatial feature after fusion

[0072] S5, using the spatio-temporal feature fusion module to fuse the spatial features of the multi-pollutant sensor network data of the predicted target location and the directed graph using spatio-temporal features, while dynamically adjusting the structure of the directed graph, performing multi-pollutant data prediction, and based on the results of the multi-pollutant data prediction, dynamically adjusting the attention between the multi-pollutants;

[0073] The specific process of step S5 is:

[0074] S51, the fused global spatial features and the sliding window time series data are spliced into joint inputs ;

[0075] S52, embedding multi-pollutant adjacency weights in the gated recurrent unit , realizing spatio-temporal feature fusion through topology-aware update gate and reset gate;

[0076] S53, dynamically adjusting the cross-pollutant attention weight based on real-time prediction error, optimizing the feature fusion ratio;

[0077] S54, synchronously generating the predicted values of all pollutants through the multi-task fully connected layer, adopting the weighted Huber loss function to optimize the model, and the calculation formula is: , wherein, is the predicted value of the pollutant c in the sliding window time series data at time t ; is the fully connected layer; is the hidden state of the GRU output; e represents the Hadamard product; the weighted Huber loss function is adopted to balance the optimization objective of multi-task learning, and the calculation formula is: , wherein, is the multi-task loss function; is the task weight coefficient; is the single-task Huber loss;

[0078] The calculation formula of the single-task Huber loss is: , wherein, is the training set time length; is the sliding window size; N c is the number of sensors; is the time series of pollutant c; is the Huber loss threshold of pollutant c ;

[0079] S6, based on the multi-pollutant collaborative generation of the abnormal score and the dynamic update threshold of the air pollutant monitoring value, performing anomaly detection;

[0080] The specific process of step S6 is:

[0081] S61, calculate the local prediction bias of the target pollutant p , and generate normalized error based on median-IQR standardization ; wherein, is the median of the training set ; is the interquartile range of the training set ;

[0082] S62, neighbor consistency error quantization node The difference in prediction bias from neighbor nodes is calculated as follows: , wherein, is the neighbor consistency error of the target pollutant p ; is the local prediction bias of the target pollutant p sensor node j ; is the neighbor node set of node i;

[0083] S63, quantify the cross-pollutant collaborative bias , reflecting the abnormal consistency of the target pollutant with other pollutants, calculated as: , wherein, is the local prediction bias of the pollutant c sensor node i ; C is the number of types of pollutants;

[0084] S64, weighted fusion of local bias, neighbor consistency error and cross-pollutant collaborative error to generate the final anomaly score, calculated as: , wherein, is the anomaly score of the sensor i at time t ; and are both balance coefficients determined by grid search on the validation set;

[0085] S65, calculate the dynamic threshold based on the sliding window quantile, calculated as: , wherein, is the dynamic threshold; is the third quartile of the anomaly score within the sliding window; is the dynamic threshold coefficient; is the interquartile range, IQR= - , is the first quartile;​

[0086] determining abnormal data.

[0087] The above description is merely that of the preferred embodiments of the application, but the protection scope of the application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the application, which should be covered by the protection scope of the application. Therefore, the protection scope of the application should be subject to the protection scope of the claims.

Claims

1. A pollutant anomaly detection method based on multi-pollutant collaboration and spatiotemporal feature fusion is characterized by: The following steps are involved: S1, collect data from multiple sensor nodes in the urban pollutant monitoring system and build a multi-pollutant sensor network dataset; S2, cleaning and preprocessing of the multi-pollutant sensor network dataset in the study area; S3. Using Bayesian variational inference to learn the directed graph adjacency matrix of the single-pollutant sensor network for the processed multi-pollutant sensor network dataset, modeling the asymmetric causal relationship between sensors; The specific process of step S3 is: S31. Modeling sensor networks as directed graphs ,in, is the set of sensor nodes that monitor pollutant c; is the adjacency matrix of pollutant c; S32, through Schur decomposition, the adjacency matrix Decompose into an orthogonal matrix With the lower triangular matrix , combined with Gumbel-Sinkhorn distribution to optimize node sorting; S33, introduce variational inference to maximize the lower bound of evidence and learn sparse directed adjacency matrix and eliminate self-loop paths through physical prior constraints; S4. The sensor nodes at the predicted target location are processed using a single pollutant spatial feature extraction module. The directed graph-guided graph attention mechanism is used to extract the multi-hop neighborhood spatial features of the single pollutant. The spatial features of multiple pollutants are fused based on the hierarchical attention mechanism. The specific process of step S4 is: S41, based on multi-hop adjacency matrix Weighted fusion to generate aggregated adjacency matrix , where K is the number of hops; is the learnable attenuation coefficient of pollutant c; S42, calculate the single pollutant sensor network data pollutant c at the sensor node through the graph attention mechanism guided by the directed graph i、 Spatial characteristics of time t , attention weight By adjacency weight Interaction score with features Jointly determined, the calculation formula is: , , ,in, To correct the linear unit, nonlinear activation is introduced; For pollutant c at node i The weight of attention on oneself; is a shared linear transformation matrix used for feature mapping; For sensor nodes i In time t Input features of For sensor nodes j In time t Input features of Pollutant c node i The set of neighbors of is the activation function; is the transpose of the learnable parameter vector; For sensor nodes i The learnable embedding vector of ; e is the Hadamard product; For nodes i The shared linear transformation matrix of ; For sensor nodes j A learnable embedding vector for For nodes j The shared linear transformation matrix of ; is the adjacency weight of multiple pollutants on the subgraph; For pollutant c at node i With node m The feature interaction score of For sensor nodes m ; For nodes i On the subgraph of pollutant c k Hop neighbor collection; t For time; S43. Calculate target pollutants based on historical data p With other pollutants c Pearson correlation coefficient , filter Top- K Strongly correlated pollutant subset ;in, is the covariance; is the observation sequence of target pollutant p; is the observation sequence of target pollutant c; is the standard deviation of the target pollutant p; is the standard deviation of the target pollutant c; S44, in the subset Internal, dynamic attention energy , combined with real-time anomaly factors and statistical correlation and adjust the attention weights ,in, is the learnable parameter vector; is the spatial characteristics of the target pollutant p; For connection; is the spatial characteristics of pollutant c; is the balance coefficient; is the attention energy of pollutant c; is the attention energy of pollutant n; S45. Generate normalized cross-pollutant attention weights through the Softmax function , weighted fusion of multi-source spatial features, the calculation formula is: ,in, is the fused global spatial feature; S5. Use the spatiotemporal feature fusion module to fuse the spatial features and directed graph of the multi-pollutant sensor network data at the predicted target location, perform multi-pollutant data prediction while dynamically adjusting the directed graph structure, and perform dynamic attention adjustment between multiple pollutants based on the results of the multi-pollutant data prediction; S6. Anomaly detection is performed based on the abnormal scores and dynamically updated thresholds of the air pollutant monitoring values ​​generated by the collaboration of multiple pollutants.

2. The pollutant anomaly detection method based on multi-pollutant collaboration and spatiotemporal feature fusion according to claim 1 is characterized in that: The specific process of step S1 is as follows: by deploying 45 urban pollutant sensor nodes in the target area, environmental monitoring data for one consecutive year is collected at a frequency of minutes to construct a time series dataset containing 8640 time points; each sensor node synchronously records the geographic location, timestamp and concentration values ​​of multiple pollutants to form a multidimensional spatiotemporal matrix.

3. The pollutant anomaly detection method based on multi-pollutant collaboration and spatiotemporal feature fusion according to claim 1 is characterized in that: The specific process of step S2 is: S21. Use minimum-maximum normalization to eliminate the dimensional differences of different pollutant sensors: ,in, is the normalized pollutant data, and the range is compressed to [0,1]; c is the pollutant type; n Number the sensor; For the n The raw data of pollutant c monitored by a sensor; For the training set n The minimum value of the pollutant c of each sensor; For the training set n The maximum value of the pollutant c of each sensor; S22. Introducing an attenuation factor into sliding window time series data , used to suppress noise interference, the calculation formula is: ,in, For pollutant c at time t Sliding window time series data; , , , For the front Term attenuation factor; For the front Term-normalized multi-pollutant data matrix.

4. The pollutant anomaly detection method based on multi-pollutant collaboration and spatiotemporal feature fusion according to claim 1 is characterized in that: The specific process of step S5 is: S51, the fused global spatial features With sliding window time series data Splice into joint input ; S52. Embedding multi-pollutant adjacency weights in gated recurrent units ,spatiotemporal feature fusion is achieved through topology-aware update gate and reset gate; S53. Dynamically adjust cross-pollutant attention weights based on real-time prediction errors , optimize the feature fusion ratio; S54, through the multi-task fully connected layer, synchronously generate the predicted values ​​of all pollutants, and use the weighted Huber loss function to optimize the model. The calculation formula is: ,in, For pollutants c In time t The predicted value of the sliding window time series data; is a fully connected layer; is the hidden state output by GRU; e represents the Hadamard product; the weighted Huber loss function is used to balance the optimization objectives of multi-task learning, and the calculation formula is: ,in, is the multi-task loss function; is the task weight coefficient; It is the single-task Huber loss; Single-task Huber loss The calculation formula is: ,in, is the length of the training set; is the sliding window size; N c is the number of sensors; is the time series of pollutant c; For pollutants c The Huber loss threshold.

5. The pollutant anomaly detection method based on multi-pollutant collaboration and spatiotemporal feature fusion according to claim 1 is characterized in that: The specific process of step S6 is: S61. Calculate target pollutants p Local prediction bias , and generate normalized errors based on median-IQR standardization ;in, For the training set the median; For the training set interquartile range; S62, Neighbor Consistency Error Quantization Node The difference in prediction deviation from neighboring nodes is calculated as: ,in, Target pollutants p Neighbor consistency error; Target pollutants p sensor nodes j local prediction bias; is the set of neighbor nodes of node i; S63. Quantifying cross-pollutant synergistic bias , reflecting the abnormal consistency between the target pollutant and other pollutants, and the calculation formula is: ,in, For pollutants c sensor nodes i The local prediction deviation; C is the number of pollutant types; S64: Weighted fusion of local deviation, neighbor consistency error, and cross-pollutant coordination error to generate the final anomaly score. The calculation formula is: ,in, For sensors i In time t The abnormal score of and are all balance coefficients, determined by grid search on the validation set; S65. Calculate the dynamic threshold based on the sliding window quantile. The calculation formula is: ,in, is the dynamic threshold; is the third quartile of the anomaly score within the sliding window; is the dynamic threshold coefficient; is the interquartile range, IQR= - , is the first quartile; determination For abnormal data.

Images