A magnetic field signal space-time decoding intelligent detection method based on a graph neural network

Abstract

The application discloses a kind of based on graph neural network's magnetic field signal space-time decoding intelligent detection method, comprising the following steps: collecting magnetic field signal, pre-processing generates magnetic field signal time sequence;Extract magnetic field amplitude, direction and rate of change characteristic component, generate window observation state vector sequence;Generation space-time propagation constraint parameter set;Build dynamic heterogeneous space-time correlation graph;Calculate path level message, generate node path message set;Perform multi-channel aggregation, input improved MPNN to perform constrained message passing and node state update, generate space-time decoding state sequence;Calculate structure consistency index and constraint violation degree, generate space-time decoding discriminant;Generate counterexample trajectory and obtain minimum contradictory subgraph, generate magnetic field signal space-time decoding intelligent detection result.The application adopts the space-time decoding method of graph neural network, realizes the high robust detection and explainable positioning of magnetic field structure anomaly, with the advantages of high precision, strong adaptability.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection of spatiotemporal decoding of magnetic field signals, and in particular to an intelligent detection method for spatiotemporal decoding of magnetic field signals based on graph neural networks. Background Technology

[0002] Magnetic field signal detection technology is widely used in industrial equipment monitoring, structural safety assessment and complex electromagnetic environment perception. Existing technologies typically acquire magnetic field signals through multi-axis magnetic field sensors and use time-domain or frequency-domain feature extraction methods combined with statistical analysis or traditional machine learning models to achieve anomaly identification. In recent years, some methods have introduced deep learning or graph models to model multi-point magnetic field signals, but most of them are based on static graph structures or simple time series modeling, which makes it difficult to simultaneously characterize the coupling relationship of magnetic field signals in the spatial dimension and the propagation characteristics in the temporal dimension.

[0003] Existing technologies, when faced with time-varying coupling, time lag, and complex noise interference in multi-axis and multi-point magnetic field signals, typically rely on synchronization assumptions or simple alignment strategies, making it difficult to construct a computable dynamic spatiotemporal correlation structure. This results in insufficient robustness in identifying structural anomalies. Furthermore, existing methods often output a single anomaly score or label, lacking interpretable analysis of anomaly formation paths and key contradictory structures, making it difficult to accurately locate the source and propagation mechanism of anomalies. Summary of the Invention

[0004] One objective of this invention is to propose an intelligent detection method for spatiotemporal decoding of magnetic field signals based on graph neural networks. This invention employs a spatiotemporal decoding method using graph neural networks to achieve highly robust detection and interpretable localization of structural anomalies in magnetic fields, possessing the advantages of high accuracy and strong adaptability.

[0005] According to an embodiment of the present invention, a spatiotemporal decoding intelligent detection method for magnetic field signals based on graph neural networks includes the following steps: Acquire magnetic field signals from a multi-axis, multi-point magnetic field sensor and preprocess them to generate a magnetic field signal time sequence; The magnetic field signal time sequence is segmented according to the sliding time window, and the characteristic components of magnetic field amplitude, direction and rate of change in each window are extracted to generate a window observation state vector sequence. Based on the spatial layout topology of the magnetic field sensor and the temporal sequence and allowable lag relationship corresponding to the sliding time window, a set of spatiotemporal propagation constraint parameters is generated. A dynamic heterogeneous spatiotemporal correlation graph is constructed based on the window observation state vector sequence and the spatiotemporal propagation constraint parameter set. The dynamic heterogeneous spatiotemporal correlation graph includes time edges, spatial edges, coupling edges, and lag edges, and corresponding constraint parameters are configured for each type of edge. In a dynamic heterogeneous spatiotemporal correlation graph, a set of spatiotemporal paths that satisfy the set of spatiotemporal propagation constraints is generated. Path-level messages are calculated along each spatiotemporal path, and a set of node path messages is generated. Perform multi-channel aggregation on the node path message set according to edge type, input it into the improved MPNN for constrained message passing and node state update, and generate a spatiotemporal decoded state sequence; Calculate the structural consistency index and constraint violation degree based on the spatiotemporal decoding state sequence, and generate the spatiotemporal decoding discriminant; Based on the spatiotemporal decoding discriminant, it is determined whether there is a structural anomaly in the magnetic field signal. When a structural anomaly is determined to exist, a counterexample trajectory is generated and the minimum contradiction subgraph is obtained by inversion, thus generating the spatiotemporal decoding intelligent detection result of the magnetic field signal.

[0006] Optionally, the preprocessing specifically includes: noise suppression, baseline correction, time alignment, amplitude normalization, anomaly removal, and time sequence reconstruction.

[0007] Optionally, the generation of the window observation state vector sequence specifically includes: Set the window length and sliding step size of the sliding time window, and perform sliding segmentation on the magnetic field signal time sequence according to the time index to obtain multiple time windows arranged in chronological order; For each time window, magnetic field signals within the corresponding time index range are selected to form a set of magnetic field signals corresponding to that time window. Amplitude calculation processing is performed on the magnetic field signal set to obtain the amplitude characteristic components corresponding to the time window; Vector synthesis processing is performed on the sampled values ​​of the multi-axis magnetic field signals in each axis of the magnetic field signal set to calculate the magnetic field direction vector of the corresponding time window, which is used as the direction feature component; Perform differential calculation on the magnetic field signals at adjacent sampling times in the magnetic field signal set to obtain the rate of change characteristic component corresponding to the time window; The amplitude feature components, direction feature components, and rate of change feature components corresponding to the same time window are combined in a preset order to generate a window observation state vector corresponding to that time window. Arrange the window observation state vectors corresponding to each time window according to the time index order of the time window to generate a window observation state vector sequence.

[0008] Optionally, the generation of the spatiotemporal propagation constraint parameter set specifically includes: Acquire spatial layout information of multi-axis multi-point magnetic field sensors, determine the spatial adjacency relationship between magnetic field sensors based on the relative positional relationship and topological connection relationship between magnetic field sensors, and generate spatial layout topological relationship of magnetic field sensors; Based on the time index relationship of the sliding time window corresponding to the window observation state vector sequence, the time sequence relationship of the sliding time window is generated; Under the constraint of the time sequence, based on the preset maximum allowable time offset range, the time index difference between sliding time windows is compared and filtered to determine the time window pairs that meet the allowable lag conditions, and the allowable lag relationship between sliding time windows is generated. Based on the spatial layout topology of magnetic field sensors, the spatial propagation paths of magnetic field signals between different magnetic field sensors are filtered, and propagation relationships that satisfy the spatial adjacency condition are retained. Corresponding spatial propagation constraint parameters are assigned to the propagation relationships. Based on the temporal sequence of the sliding time window, the propagation direction of the magnetic field signal in the time dimension is constrained, the propagation relationship that meets the temporal sequence requirement is selected, and the propagation relationship is assigned a time propagation constraint parameter; Based on the allowable hysteresis relationship, the propagation range of the magnetic field signal across the sliding time window is limited, the propagation relationship that satisfies the hysteresis condition is selected, and the propagation relationship is assigned a hysteresis propagation constraint parameter; Spatial propagation constraint parameters, temporal propagation constraint parameters, and lag propagation constraint parameters are collected and organized to generate a set of spatiotemporal propagation constraint parameters.

[0009] Optionally, the generation of the dynamic heterogeneous spatiotemporal correlation graph specifically includes: Each window observation state vector in the window observation state vector sequence is converted into a graph node, and a corresponding sliding time window time index is attached to each graph node to generate a graph node set. Under the same sliding time window time index, based on the spatial layout topology of the magnetic field sensors in the spatiotemporal propagation constraint parameter set, connections reflecting spatial adjacency relationships are established between graph nodes in the graph node set to form a spatial edge set. Around the graph nodes with time indices of different sliding time windows, according to the time sequence rules determined by the set of spatiotemporal propagation constraint parameters, the node connection relationship that reflects the direction of time evolution is constructed to obtain the set of time edges; For graph node pairs with offset time indices, the allowable lag rules defined in the spatiotemporal propagation constraint parameter set are combined to introduce cross-time window connections between graph nodes that satisfy the lag constraints, generating a set of lag edges. For graph nodes originating from different magnetic field sensors, by combining the topological relationship of the spatial layout of the magnetic field sensors with the corresponding information of the window observation state vector, a connection structure reflecting the magnetic field coupling relationship is constructed between the graph nodes to form a set of coupling edges. Assign edge type identifiers to the spatial edge set, temporal edge set, lag edge set, and coupled edge set respectively, and load the corresponding constraint parameters to different edge types according to the spatiotemporal propagation constraint parameter set; By combining the graph node set, spatial edge set, temporal edge set, lag edge set, coupling edge set, and their corresponding constraint parameters, a dynamic heterogeneous spatiotemporal correlation graph is generated.

[0010] Optionally, the generation of the node path message set specifically includes: Based on the dynamic heterogeneous spatiotemporal association graph, combined with the spatiotemporal propagation constraint parameter set, the propagation legality of the connection relationship between graph nodes is determined, and the edges allowed to participate in propagation under the constraints of spatial propagation, temporal propagation and delayed propagation are determined, generating a set of propagable edges; Based on the set of propagable edges, a node sequence from the starting node to the subsequent nodes is constructed in the dynamic heterogeneous spatiotemporal association graph according to the connection order between nodes. The node sequence that satisfies the spatiotemporal propagation constraint parameter set is determined as the spatiotemporal path, and a set of spatiotemporal paths is generated. For each spatiotemporal path in the set of spatiotemporal paths, the corresponding window observation state vector is read sequentially according to the order of the graph nodes in the spatiotemporal path to form the path node state sequence corresponding to the spatiotemporal path. Along each spatiotemporal path, perform aggregation calculations on the path node state sequence to obtain the path-level message corresponding to that spatiotemporal path; For the same graph node, all path-level messages corresponding to the spatiotemporal paths that pass through the graph node are collected to generate the node path message corresponding to the graph node. The node path messages corresponding to all graph nodes in the dynamic heterogeneous spatiotemporal correlation graph are collected to generate a node path message set.

[0011] Optionally, the generation of the spatiotemporal decoding state sequence specifically includes: For each graph node in a dynamic heterogeneous spatiotemporal relational graph, the node path messages corresponding to the graph node are split according to the edge types traversed during the generation process, forming a set of multiple types of path messages corresponding to spatial edges, temporal edges, lag edges, and coupling edges respectively. While splitting the message sets of multiple paths, corresponding constraint parameters from the spatiotemporal propagation constraint parameter set are introduced to apply message propagation constraints consistent with their edge types to each type of path message. Based on the set of multiple path messages after the constraints are applied, independent aggregation processing is performed on the path messages corresponding to spatial edges, temporal edges, lag edges, and coupling edges to generate channel-level aggregated messages corresponding to spatial channels, temporal channels, lag channels, and coupling channels, respectively. The aggregated messages of each channel corresponding to the same graph node are combined in a preset order to generate a multi-channel node feature set corresponding to that graph node. The multi-channel node feature set is input into the improved MPNN. Under the continuous constraint of the spatiotemporal propagation constraint parameter set, constrained message passing processing is performed between graph nodes in the dynamic heterogeneous spatiotemporal association graph to generate the updated intermediate state of each graph node in the current sliding time window. The improvements to the MPNN specifically include: Using a multi-channel node feature set as input to the graph node, the graph node can simultaneously represent the information corresponding to spatial propagation, temporal propagation, lag propagation, and magnetic field coupling propagation during message passing. During the message passing process, a spatiotemporal propagation constraint parameter set is introduced to continuously apply propagation constraints to the message passing process between graph nodes in the dynamic heterogeneous spatiotemporal association graph. During message passing and node state update, the message passing input is the node path message after multi-channel aggregation and the application of propagation constraints. Based on updating intermediate states, perform node state update operations on each graph node to generate a spatiotemporal decoding state corresponding to the corresponding sliding time window; Arrange the spatiotemporal decoding states generated under each sliding time window according to the time index order of the sliding time window to generate a spatiotemporal decoding state sequence.

[0012] Optionally, the generation of the spatiotemporal decoding discriminant specifically includes: Under the constraints of dynamic heterogeneous spatiotemporal correlation graph, based on the spatiotemporal correlation relationships formed between graph nodes through time edges, spatial edges, coupling edges and lag edges, consistency calculation is performed on the spatiotemporal decoding states corresponding to the interconnected graph nodes in the spatiotemporal decoding state sequence to generate structural consistency index. Under different time index conditions corresponding to the same graph node, the spatiotemporal decoding state corresponding to the graph node in the spatiotemporal decoding state sequence changes with time, and a time series consistency index is generated. Based on structural consistency index and temporal consistency index, the node association relationships that violate the spatiotemporal propagation constraints in the dynamic heterogeneous spatiotemporal association graph are identified, and the constraint violation degree is generated. The structural consistency index and constraint violation degree are jointly aggregated and processed to generate a spatiotemporal decoding discriminant.

[0013] Optionally, the generation of the spatiotemporal decoding intelligent detection result of the magnetic field signal specifically includes: Based on the spatiotemporal decoding discrimination value, the nodes and their relationships in the dynamic heterogeneous spatiotemporal correlation graph are judged, the node association regions that meet the anomaly judgment conditions of the spatiotemporal decoding discrimination value are identified, and the structural anomalies of the magnetic field signal are determined. In cases where a structural anomaly is determined to exist in the magnetic field signal, the correlation paths between the anomaly nodes are traced in time index order along the spatiotemporal propagation relationship in the dynamic heterogeneous spatiotemporal correlation graph to generate counterexample trajectories. Based on the counterexample trajectory, the nodes and edges related to the counterexample trajectory in the dynamic heterogeneous spatiotemporal correlation graph are filtered to construct a candidate anomaly subgraph containing the nodes and relationships related to the counterexample trajectory; The candidate abnormal subgraph is inverted. Based on the constraints of minimizing the number of nodes and simplifying the relationships, nodes and relationships in the candidate abnormal subgraph that do not constitute the minimum contradiction structure are gradually pruned to generate the minimum contradiction subgraph. Based on the minimum contradiction subgraph, the abnormal nodes, corresponding time index intervals and edge type information contained therein are collected to generate the spatiotemporal decoding intelligent detection results of magnetic field signals.

[0014] The beneficial effects of this invention are: This invention constructs an intelligent detection method for spatiotemporal decoding of magnetic field signals based on graph neural networks. It achieves unified modeling and joint analysis of multi-axis, multi-point magnetic field signals under spatial propagation, temporal evolution, and hysteresis effects. Compared with existing technologies that rely solely on time series features or static spatial relationships, this invention introduces a dynamic heterogeneous spatiotemporal correlation graph. It explicitly incorporates the spatial layout topology of magnetic field sensors, the temporal sequence of sliding time windows, and the allowable hysteresis relationship into the calculation process. This allows the propagation law of magnetic field signals in complex physical environments to be expressed in a structured form, thereby significantly improving the modeling capability for time-varying coupled magnetic field signals.

[0015] This invention further combines local changes in magnetic field signals with global propagation patterns by generating a set of spatiotemporal paths that satisfy spatiotemporal propagation constraints and calculating path-level messages along the paths. Then, through multi-channel aggregation and an improved MPNN, constrained message passing and node state updates are executed to form a stable and distinguishable spatiotemporal decoding state sequence. This processing method effectively suppresses the impact of non-structural interference such as noise and drift on the detection results, improves the robustness and temporal stability of structural anomaly detection, and, by calculating structural consistency indices and constraint violation degrees, enables anomaly judgment to be based on the objective basis of spatiotemporal structural instability, avoiding the risk of misjudgment caused by relying solely on thresholds or single scores.

[0016] In the output stage of the detection results, this invention can intuitively reveal the key path, related nodes and propagation relationship of anomaly formation through counterexample trajectory tracing and minimum contradiction subgraph inversion, providing clear and interpretable evidence support for anomaly source location and mechanism analysis. In summary, this invention not only achieves high-precision intelligent detection of complex magnetic field signals, but also takes into account the interpretability and engineering applicability of the results, and is suitable for magnetic field anomaly monitoring and analysis scenarios in a variety of complex electromagnetic environments. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an intelligent detection method for spatiotemporal decoding of magnetic field signals based on graph neural networks proposed in this invention; Figure 2 This is a schematic diagram of the construction of a dynamic heterogeneous spatiotemporal correlation graph for a magnetic field signal spatiotemporal decoding intelligent detection method based on graph neural networks proposed in this invention; Figure 3 This is a schematic diagram illustrating the constrained spatiotemporal path generation and path-level message calculation of a spatiotemporal decoding intelligent detection method for magnetic field signals based on graph neural networks proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figures 1-3 A spatiotemporal decoding intelligent detection method for magnetic field signals based on graph neural networks includes the following steps: Acquire magnetic field signals from a multi-axis, multi-point magnetic field sensor and preprocess them to generate a magnetic field signal time sequence; The magnetic field signal time sequence is segmented according to the sliding time window, and the characteristic components of magnetic field amplitude, direction and rate of change in each window are extracted to generate a window observation state vector sequence. Based on the spatial layout topology of the magnetic field sensor and the temporal sequence and allowable lag relationship corresponding to the sliding time window, a set of spatiotemporal propagation constraint parameters is generated. A dynamic heterogeneous spatiotemporal correlation graph is constructed based on the window observation state vector sequence and the spatiotemporal propagation constraint parameter set. The dynamic heterogeneous spatiotemporal correlation graph includes time edges, spatial edges, coupling edges, and lag edges, and corresponding constraint parameters are configured for each type of edge. In a dynamic heterogeneous spatiotemporal correlation graph, a set of spatiotemporal paths that satisfy the set of spatiotemporal propagation constraints is generated. Path-level messages are calculated along each spatiotemporal path, and a set of node path messages is generated. Perform multi-channel aggregation on the node path message set according to edge type, input it into the improved MPNN for constrained message passing and node state update, and generate a spatiotemporal decoded state sequence; Calculate the structural consistency index and constraint violation degree based on the spatiotemporal decoding state sequence, and generate the spatiotemporal decoding discriminant; Based on the spatiotemporal decoding discriminant, it is determined whether there is a structural anomaly in the magnetic field signal. When a structural anomaly is determined to exist, a counterexample trajectory is generated and the minimum contradiction subgraph is obtained by inversion, thus generating the spatiotemporal decoding intelligent detection result of the magnetic field signal.

[0020] In this embodiment, the preprocessing specifically includes: noise suppression, baseline correction, time alignment, amplitude normalization, anomaly removal, and time sequence reconstruction.

[0021] In this embodiment, the generation of the window observation state vector sequence specifically includes: Set the window length and sliding step size of the sliding time window, and perform sliding segmentation on the magnetic field signal time sequence according to the time index to obtain multiple time windows arranged in chronological order; For each time window, magnetic field signals within the corresponding time index range are selected to form a set of magnetic field signals corresponding to that time window. Amplitude calculation processing is performed on the magnetic field signal set to obtain the amplitude characteristic components corresponding to the time window; Vector synthesis processing is performed on the sampled values ​​of the multi-axis magnetic field signals in each axis of the magnetic field signal set to calculate the magnetic field direction vector of the corresponding time window, which is used as the direction feature component; Perform differential calculation on the magnetic field signals at adjacent sampling times in the magnetic field signal set to obtain the rate of change characteristic component corresponding to the time window; The amplitude feature components, direction feature components, and rate of change feature components corresponding to the same time window are combined in a preset order to generate a window observation state vector corresponding to that time window. Arrange the window observation state vectors corresponding to each time window according to the time index order of the time window to generate a window observation state vector sequence.

[0022] In this embodiment, the generation of the spatiotemporal propagation constraint parameter set specifically includes: Acquire spatial layout information of multi-axis multi-point magnetic field sensors, determine the spatial adjacency relationship between magnetic field sensors based on the relative positional relationship and topological connection relationship between magnetic field sensors, and generate spatial layout topological relationship of magnetic field sensors; Based on the time index relationship of the sliding time window corresponding to the window observation state vector sequence, the time sequence relationship of the sliding time window is generated; Under the constraint of the time sequence, based on the preset maximum allowable time offset range, the time index difference between sliding time windows is compared and filtered to determine the time window pairs that meet the allowable lag conditions, and the allowable lag relationship between sliding time windows is generated. The generation of the allowed hysteresis relationship specifically includes: Based on the temporal sequence of sliding time windows, the propagation direction of the magnetic field signal in the time dimension is limited, determining time window pairs that are only allowed to propagate from the earlier time-indexed sliding time window to the later time-indexed sliding time window. According to a preset maximum allowable time offset range, the time index difference between time window pairs is compared and filtered to determine whether each time window pair meets the allowable lag condition. When the time index difference of a time window pair falls within the maximum allowable time offset range, the time window pair is determined to meet the allowable lag condition, and an allowable lag relationship is established between the corresponding sliding time windows. When the time index difference of a time window pair exceeds the maximum allowable time offset range, the time window pair is determined not to meet the allowable lag condition, and an allowable lag relationship is not established between the corresponding sliding time windows. By sequentially performing this determination process on each time window pair, the allowable lag relationship between sliding time windows is generated. Based on the spatial layout topology of magnetic field sensors, the spatial propagation paths of magnetic field signals between different magnetic field sensors are filtered, and propagation relationships that satisfy the spatial adjacency condition are retained. Corresponding spatial propagation constraint parameters are assigned to the propagation relationships. The specific spatial propagation constraint parameters assigned include: Based on the spatial layout topology of magnetic field sensors, the spatial connection relationships between the magnetic field signals corresponding to different magnetic field sensors are traversed to identify whether each magnetic field sensor satisfies the spatial adjacency condition. For magnetic field sensor pairs that satisfy the spatial adjacency condition, it is determined that the magnetic field signal can propagate spatially between the corresponding magnetic field sensors, and the propagation relationship between the magnetic field sensor pairs is determined as a valid spatial propagation relationship. Spatial propagation constraint parameters are assigned to the valid spatial propagation relationship to limit the legality of the spatial propagation of the magnetic field signal between different magnetic field sensors. This judgment and parameter assignment process is executed sequentially for each magnetic field sensor pair to complete the assignment of spatial propagation constraint parameters. Based on the temporal sequence of the sliding time window, the propagation direction of the magnetic field signal in the time dimension is constrained, the propagation relationship that meets the temporal sequence requirement is selected, and the propagation relationship is assigned a time propagation constraint parameter; The specific parameters for assigning time propagation constraints include: Based on the temporal sequence of sliding time windows, the temporal connection relationship between magnetic field signals corresponding to different sliding time windows is traversed to identify whether the temporal sequence requirement is met between each pair of sliding time windows. For sliding time window pairs that meet the time sequence requirements, it is determined that the magnetic field signal is allowed to propagate along the time index from the earlier sliding time window to the later sliding time window, and the corresponding propagation relationship is determined as the effective time propagation relationship; a time propagation constraint parameter is assigned to the effective time propagation relationship to limit the propagation direction of the magnetic field signal in the time dimension; by performing this determination and parameter assignment process on each sliding time window pair in sequence, the time propagation constraint parameter assignment is completed. Based on the allowable hysteresis relationship, the propagation range of the magnetic field signal across the sliding time window is limited, the propagation relationship that satisfies the hysteresis condition is selected, and the propagation relationship is assigned a hysteresis propagation constraint parameter; The parameters for assigning the hysteresis propagation constraint specifically include: Based on the permissible hysteresis relationship between sliding time windows, the propagation span across the window for magnetic field signals corresponding to different sliding time windows is traversed and determined to identify whether each sliding time window pair is within the permissible hysteresis time offset range. For sliding time window pairs within the permissible hysteresis relationship range, the propagation span of the magnetic field signal across the sliding time window is limited according to the permissible hysteresis relationship. Cross-window propagation relationships that meet the hysteresis condition are selected and determined as effective hysteresis propagation relationships. Hysteresis propagation constraint parameters are assigned to the effective hysteresis propagation relationships to limit the time offset range of the magnetic field signal during the propagation process across the sliding time window. By sequentially performing the propagation span determination and parameter assignment process on each sliding time window pair, the assignment of hysteresis propagation constraint parameters is completed. Spatial propagation constraint parameters, temporal propagation constraint parameters, and lag propagation constraint parameters are collected and organized to generate a set of spatiotemporal propagation constraint parameters.

[0023] In this embodiment, the generation of the dynamic heterogeneous spatiotemporal correlation graph specifically includes: Each window observation state vector in the window observation state vector sequence is converted into a graph node, and a corresponding sliding time window time index is attached to each graph node to generate a graph node set. Under the same sliding time window time index, based on the spatial layout topology of the magnetic field sensors in the spatiotemporal propagation constraint parameter set, connections reflecting spatial adjacency relationships are established between graph nodes in the graph node set to form a spatial edge set. Around the graph nodes with time indices of different sliding time windows, according to the time sequence rules determined by the set of spatiotemporal propagation constraint parameters, the node connection relationship that reflects the direction of time evolution is constructed to obtain the set of time edges; For graph node pairs with offset time indices, the allowable lag rules defined in the spatiotemporal propagation constraint parameter set are combined to introduce cross-time window connections between graph nodes that satisfy the lag constraints, generating a set of lag edges. The generation of the lagging edge set specifically includes: For graph node pairs with offset time indices in dynamic heterogeneous spatiotemporal graphs, the time index offset relationship between graph node pairs is traversed and determined based on the sliding time window time index corresponding to each graph node, and graph node pairs with cross-sliding time window relationships are selected. Combining the allowable lag rules defined in the spatiotemporal propagation constraint parameter set, lag condition determination is performed on graph node pairs with cross-sliding time window relationships to identify graph node pairs that satisfy the lag constraints limited by the allowable lag rules. Cross-time window connection relationships are introduced between graph node pairs that satisfy the lag constraints, and the connection relationships are determined as lag edges. By performing a cross-time window connection process on graph node pairs that satisfy the lag constraints, a set of lag edges is generated. For graph nodes originating from different magnetic field sensors, by combining the topological relationship of the spatial layout of the magnetic field sensors with the corresponding information of the window observation state vector, a connection structure reflecting the magnetic field coupling relationship is constructed between the graph nodes to form a set of coupling edges. The generation of the coupling edge set specifically includes: For graph nodes originating from different magnetic field sensors in a dynamic heterogeneous spatiotemporal correlation graph, the sensor origin of the graph nodes is determined based on the correspondence between the window observation state vector and the magnetic field sensor, and graph node pairs originating from different magnetic field sensors are selected. Combining the spatial layout topology of the magnetic field sensors, it is determined whether there is a magnetic field coupling relationship between graph node pairs originating from different magnetic field sensors, and graph node pairs that satisfy the magnetic field coupling relationship are identified. A connection structure reflecting the magnetic field coupling relationship is constructed between graph node pairs that satisfy the magnetic field coupling relationship, and the connection structure is determined as the coupling edge. By performing this connection process on graph node pairs that satisfy the magnetic field coupling relationship, a set of coupling edges is generated. Assign edge type identifiers to the spatial edge set, temporal edge set, lag edge set, and coupled edge set respectively, and load the corresponding constraint parameters to different edge types according to the spatiotemporal propagation constraint parameter set; The constraint parameters are specifically loaded onto different edge types, including: Edge types are identified for the spatial edge set, temporal edge set, lag edge set, and coupled edge set in the dynamic heterogeneous spatiotemporal relation graph to distinguish the propagation relationship types corresponding to different edge sets. Based on the spatiotemporal propagation constraint parameter set, the various edge sets of identified edge types are traversed to determine the constraint parameter types corresponding to each edge set. Spatial propagation constraint parameters are loaded into the spatial edges of the spatial edge set; temporal propagation constraint parameters are loaded into the temporal edges of the temporal edge set; lag propagation constraint parameters are loaded into the lag edges of the lag edge set; and coupled propagation constraint parameters are loaded into the coupled edges of the coupled edge set. By performing this process separately for different edge types, the constraint parameter loading for different edge types is completed. By combining the graph node set, spatial edge set, temporal edge set, lag edge set, coupling edge set, and their corresponding constraint parameters, a dynamic heterogeneous spatiotemporal correlation graph is generated.

[0024] In this embodiment, the generation of the node path message set specifically includes: Based on the dynamic heterogeneous spatiotemporal association graph, combined with the spatiotemporal propagation constraint parameter set, the propagation legality of the connection relationship between graph nodes is determined, and the edges allowed to participate in propagation under the constraints of spatial propagation, temporal propagation and delayed propagation are determined, generating a set of propagable edges; The generation of the propagable edge set specifically includes: Based on a dynamic heterogeneous spatiotemporal relational graph, the connections between graph nodes formed by spatial edges, temporal edges, lag edges, and coupling edges are traversed to obtain candidate propagation connections between graph nodes. Combining the spatiotemporal propagation constraint parameter set, the propagation legality of the candidate propagation connections is determined under spatial propagation, temporal propagation, and lag propagation conditions, respectively, to identify connections that satisfy the corresponding propagation constraints. Connections that satisfy the spatial, temporal, and lag propagation constraints are determined as edges allowed to participate in propagation. By aggregating the connections that satisfy the propagation legality conditions, a set of propagable edges is generated. Based on the set of propagable edges, a node sequence from the starting node to the subsequent nodes is constructed in the dynamic heterogeneous spatiotemporal association graph according to the connection order between nodes. The node sequence that satisfies the spatiotemporal propagation constraint parameter set is determined as the spatiotemporal path, and a set of spatiotemporal paths is generated. The generation of the spatiotemporal path set specifically includes: Based on the set of propagable edges, the connection relationships between graph nodes in a dynamic heterogeneous spatiotemporal graph are traversed, and a node sequence pointing from the starting graph node to subsequent graph nodes is constructed according to the connection order between the graph nodes. Combining the spatiotemporal propagation constraint parameter set, the propagation constraint consistency judgment is performed on the node sequence to identify the node sequence that satisfies the spatial propagation, temporal propagation, and lag propagation constraints. The node sequence that satisfies the spatiotemporal propagation constraint parameter set is determined as the spatiotemporal path. By aggregating the node sequences that satisfy the spatiotemporal propagation constraint parameter set, a spatiotemporal path set is generated. For each spatiotemporal path in the set of spatiotemporal paths, the corresponding window observation state vector is read sequentially according to the order of the graph nodes in the spatiotemporal path to form the path node state sequence corresponding to the spatiotemporal path. Along each spatiotemporal path, perform aggregation calculations on the path node state sequence to obtain the path-level message corresponding to that spatiotemporal path; For the same graph node, all path-level messages corresponding to the spatiotemporal paths that pass through the graph node are collected to generate the node path message corresponding to the graph node. The node path messages corresponding to all graph nodes in the dynamic heterogeneous spatiotemporal correlation graph are collected to generate a node path message set.

[0025] In this embodiment, the generation of the spatiotemporal decoding state sequence specifically includes: For each graph node in a dynamic heterogeneous spatiotemporal relational graph, the node path messages corresponding to the graph node are split according to the edge types traversed during the generation process, forming a set of multiple types of path messages corresponding to spatial edges, temporal edges, lag edges, and coupling edges respectively. The generation of the multipath message set specifically includes: For each graph node in a dynamic heterogeneous spatiotemporal relational graph, the node path message corresponding to that graph node is obtained. The node path message originates from each spatiotemporal path that passes through the graph node. Based on the edge type traversed during the generation process, the node path message is split into categories, namely spatial edges, temporal edges, lag edges, and coupling edges. The node path messages that traverse spatial edges, temporal edges, lag edges, and coupling edges are then aggregated to form path message sets corresponding to spatial edges, temporal edges, lag edges, and coupling edges, respectively. While splitting the message sets of multiple paths, corresponding constraint parameters from the spatiotemporal propagation constraint parameter set are introduced to apply message propagation constraints consistent with their edge types to each type of path message. The application of the propagation constraint conditions specifically includes: During the process of splitting node path messages according to edge type, a spatiotemporal propagation constraint parameter set is introduced for the resulting multi-type path message set; based on the edge type corresponding to the multi-type path message set, propagation constraint parameters consistent with the edge type are selected from the spatiotemporal propagation constraint parameter set; the selected propagation constraint parameters are loaded into the path message of the corresponding edge type, and message propagation constraint conditions consistent with its edge type are applied to the path message of that type. Based on the set of multiple path messages after the constraints are applied, independent aggregation processing is performed on the path messages corresponding to spatial edges, temporal edges, lag edges, and coupling edges to generate channel-level aggregated messages corresponding to spatial channels, temporal channels, lag channels, and coupling channels, respectively. The aggregated messages of each channel corresponding to the same graph node are combined in a preset order to generate a multi-channel node feature set corresponding to that graph node. The multi-channel node feature set is input into the improved MPNN. Under the continuous constraint of the spatiotemporal propagation constraint parameter set, constrained message passing processing is performed between graph nodes in the dynamic heterogeneous spatiotemporal association graph to generate the updated intermediate state of each graph node in the current sliding time window. The improvements to the MPNN specifically include: Using a multi-channel node feature set as input to graph nodes, the graph nodes can simultaneously represent information corresponding to spatial propagation, temporal propagation, delayed propagation, and magnetic field coupling propagation during message passing. During message passing, a spatiotemporal propagation constraint parameter set is introduced to continuously apply propagation constraints to the message passing process between graph nodes in a dynamic heterogeneous spatiotemporal association graph, ensuring that messages are passed only under connection relationships that satisfy spatial, temporal, and delayed propagation constraints. During message passing and node state updates, the message passing input is the node path message aggregated through multiple channels and subject to propagation constraints, allowing the node state update result to simultaneously reflect the combined influence of multiple propagation paths and their propagation constraints. Through the combined effect of the multi-channel node feature set and the spatiotemporal propagation constraint parameter set, constrained message passing and node state updates are achieved for graph nodes in a dynamic heterogeneous spatiotemporal association graph. Based on updating intermediate states, perform node state update operations on each graph node to generate a spatiotemporal decoding state corresponding to the corresponding sliding time window; Arrange the spatiotemporal decoding states generated under each sliding time window according to the time index order of the sliding time window to generate a spatiotemporal decoding state sequence.

[0026] In this embodiment, the generation of the spatiotemporal decoding discriminant specifically includes: Under the constraints of dynamic heterogeneous spatiotemporal correlation graph, based on the spatiotemporal correlation relationships formed between graph nodes through time edges, spatial edges, coupling edges and lag edges, consistency calculation is performed on the spatiotemporal decoding states corresponding to the interconnected graph nodes in the spatiotemporal decoding state sequence to generate structural consistency index. The generation of the structural consistency index specifically includes: Under the constraints of a dynamic heterogeneous spatiotemporal correlation graph, the spatiotemporal correlation relationships formed between graph nodes through time edges, spatial edges, lag edges, and coupling edges are used to locate the correlation relationships of graph nodes in the spatiotemporal decoding state sequence, thus identifying the graph node pairs that are mutually related under the spatiotemporal correlation constraints. For each identified mutually related graph node pair, the spatiotemporal decoding state of the graph node under the corresponding sliding time window is obtained. Based on the correlation relationship type of the mutually related graph node pairs in the dynamic heterogeneous spatiotemporal correlation graph, consistency calculation is performed on the spatiotemporal decoding state corresponding to the graph node to obtain a consistency result reflecting the degree of state consistency of the graph node under the spatiotemporal correlation constraints. By aggregating the spatiotemporal decoding state consistency results of all mutually related graph node pairs, a structural consistency index is generated. Under different time index conditions corresponding to the same graph node, the spatiotemporal decoding state corresponding to the graph node in the spatiotemporal decoding state sequence changes with time, and a time series consistency index is generated. Based on structural consistency index and temporal consistency index, the node association relationships that violate the spatiotemporal propagation constraints in the dynamic heterogeneous spatiotemporal association graph are identified, and the constraint violation degree is generated. The generation of the constraint violation specifically includes: Based on the spatiotemporal relationships formed by time edges, spatial edges, lag edges, and coupling edges in the dynamic heterogeneous spatiotemporal correlation graph, the node correlation relationships between graph nodes are located to determine the set of node correlation relationships that need to participate in constraint consistency analysis. For the set of node correlation relationships, the structural consistency index and temporal consistency index corresponding to the node correlation relationship are read to form an index combination for constraint consistency analysis. Based on the spatiotemporal propagation constraints, a constraint consistency judgment is performed on the index combination to identify node associations that are inconsistent with the spatiotemporal propagation constraints under the representation of structural consistency indicators and temporal consistency indicators. By aggregating and quantifying the node associations that are inconsistent with the spatiotemporal propagation constraints, a constraint violation degree reflecting the degree to which node associations violate the spatiotemporal propagation constraints is generated. The structural consistency index and constraint violation degree are jointly aggregated and processed to generate a spatiotemporal decoding discriminant.

[0027] In this embodiment, the generation of the spatiotemporal decoding intelligent detection result of the magnetic field signal specifically includes: Based on the spatiotemporal decoding discrimination value, the nodes and their relationships in the dynamic heterogeneous spatiotemporal correlation graph are judged, the node association regions that meet the anomaly judgment conditions of the spatiotemporal decoding discrimination value are identified, and the structural anomalies of the magnetic field signal are determined. In cases where a structural anomaly is determined to exist in the magnetic field signal, the correlation paths between the anomaly nodes are traced in time index order along the spatiotemporal propagation relationship in the dynamic heterogeneous spatiotemporal correlation graph to generate counterexample trajectories. The generation of the counterexample trajectory specifically includes: In a dynamic heterogeneous spatiotemporal correlation graph, based on the abnormal nodes corresponding to the spatiotemporal decoding discriminant, the starting abnormal node for participating in the tracing of counterexample trajectories is determined. According to the spatiotemporal propagation relationship formed by time edges, spatial edges, lag edges, and coupling edges in the dynamic heterogeneous spatiotemporal correlation graph, starting from the starting abnormal node, the correlation path between abnormal nodes is reversed according to the time index order of the sliding time window to construct the propagation path sequence corresponding to the abnormal node. The node sequence obtained in the reverse tracing process that conforms to the spatiotemporal propagation relationship constraints is determined as the counterexample trajectory. Based on the counterexample trajectory, the nodes and edges related to the counterexample trajectory in the dynamic heterogeneous spatiotemporal correlation graph are filtered to construct a candidate anomaly subgraph containing the nodes and relationships related to the counterexample trajectory; The generation of the candidate anomaly subgraph specifically includes: Based on the counterexample trajectory, the graph node corresponding to the counterexample trajectory is located in the dynamic heterogeneous spatiotemporal correlation graph, and the trajectory-related nodes participating in the construction of the candidate anomaly subgraph are determined; based on the trajectory-related nodes, the edges connecting the trajectory-related nodes in the dynamic heterogeneous spatiotemporal correlation graph are filtered to determine the association relationship corresponding to the counterexample trajectory; the trajectory-related nodes and their corresponding association relationships are combined to construct a candidate anomaly subgraph containing nodes and association relationships related to the counterexample trajectory. The candidate abnormal subgraph is inverted. Based on the constraints of minimizing the number of nodes and simplifying the relationships, nodes and relationships in the candidate abnormal subgraph that do not constitute the minimum contradiction structure are gradually pruned to generate the minimum contradiction subgraph. The generation of the minimum contradictory subgraph specifically includes: Using candidate anomaly subgraphs as the inversion processing object, an initial subgraph for contradiction structure analysis is constructed based on the nodes and their relationships contained in the candidate anomaly subgraphs. Under the constraint of the counterexample trajectory, the nodes and their relationships in the initial subgraph are progressively pruned according to the constraints of minimizing the number of nodes and simplifying the relationships. In each round of pruning, the consistency of whether the pruned subgraph still maintains the contradiction structure under the counterexample trajectory constraint is verified, and the nodes and relationships that can maintain the contradiction structure are retained. By continuously executing the pruning and verification process under the premise of satisfying the counterexample trajectory constraint, the minimum contradiction subgraph is generated. Based on the minimum contradiction subgraph, the abnormal nodes, corresponding time index intervals and edge type information contained therein are collected to generate the spatiotemporal decoding intelligent detection results of magnetic field signals.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to the operating environment of a large industrial equipment cluster. Continuous monitoring and intelligent analysis of multi-axis, multi-point magnetic field signals were performed. This environment contains multiple high-power motors, frequency converters, and metal structural components. Complex magnetic field disturbances are generated during equipment start-up, shutdown, load switching, and changes in operating conditions. These disturbances propagate spatially through the metal structure and exhibit significant temporal delays. Furthermore, environmental noise and long-term drift factors are also superimposed. Traditional magnetic field monitoring methods in this environment typically only obtain overall intensity changes or simple anomaly alarms, making it difficult to distinguish between local structural anomalies and random interference, and also difficult to locate the formation and propagation path of anomalies, posing significant challenges to equipment operation and maintenance and safety analysis.

[0029] In the above scenario, multi-axis, multi-point magnetic field sensors are deployed according to the spatial structure and key parts of the equipment, enabling each sensor to synchronously acquire magnetic field signals from different positions and directions. The acquired magnetic field signals are first processed through a preprocessing procedure to eliminate environmental noise, correct baseline drift, and perform time alignment and normalization to form a stable magnetic field signal time sequence. The magnetic field signal time sequence is then segmented according to a sliding time window method, and feature components such as magnetic field amplitude, direction, and rate of change are extracted within each time window to construct a window observation state vector sequence, so that the evolution characteristics of the magnetic field signal in the time dimension can be clearly expressed.

[0030] Based on this, a spatiotemporal propagation constraint parameter set is generated by combining the spatial layout of magnetic field sensors with the temporal sequence and allowable lag between sliding time windows. This constraint parameter set incorporates the spatial adjacency of the equipment structure, the temporal directionality of magnetic field propagation, and the allowable time delay during physical propagation into the constraint system. Subsequently, a dynamic heterogeneous spatiotemporal correlation graph is constructed based on the window observation state vector sequence and the spatiotemporal propagation constraint parameter set. The magnetic field state under different time windows is mapped as graph nodes, and time edges, spatial edges, coupling edges, and lag edges are established between nodes, so that the correlation of magnetic field signals under complex spatial structures and temporal evolution conditions is explicitly presented in the form of a graph structure.

[0031] After the dynamic heterogeneous spatiotemporal correlation graph is constructed, the system generates a set of spatiotemporal paths that satisfy the spatiotemporal propagation constraints in the graph structure. Each spatiotemporal path corresponds to a possible propagation trajectory of magnetic field disturbance in the device structure. The path contains multiple graph nodes and their relationships. Path-level messages are calculated along each spatiotemporal path, and the magnetic field state information of multiple time windows and multiple sensor locations on the path is integrated to form a set of node path messages. This processing method enables local magnetic field changes to be analyzed in the context of the global structure, effectively avoiding misjudgments caused by insufficient information at a single point or at a single moment.

[0032] Multi-channel aggregation is performed on the node path message set according to edge type, and the information corresponding to spatial propagation, temporal propagation, lag propagation and magnetic field coupling propagation are collected respectively and input into the improved message passing neural network. In this network, the spatiotemporal propagation constraint parameter set continuously acts on the message passing process, so that the message can only propagate along the correlation that satisfies the physical and temporal constraints, thereby generating a stable and discriminative spatiotemporal decoding state sequence. Through this constrained message passing mechanism, the system can effectively suppress abnormal responses caused by random noise or short-term fluctuations, and make the detection results more focused on the structural changes of the magnetic field.

[0033] After the spatiotemporal decoding state sequence is generated, the system further calculates the structural consistency index and constraint violation degree, and performs a comprehensive analysis of each node and its relationship in the graph structure. When the equipment is operating normally, the propagation relationship of the magnetic field signal in space and time can well meet the given constraints, and the spatiotemporal decoding state shows high consistency. When structural anomalies or potential faults occur inside the equipment, the magnetic field disturbance will deviate significantly in the propagation path and coupling relationship, causing the relationship of some nodes to violate the spatiotemporal propagation constraints, which is then accurately reflected in the spatiotemporal decoding discrimination quantity.

[0034] After an anomaly is identified, the system further traces the anomaly nodes along the propagation relationship in the dynamic heterogeneous spatiotemporal correlation graph, generating counterexample trajectories to describe the evolution process of the anomaly from its initial location to the diffusion area. Based on this, by inverting and pruning the subgraph, a minimal contradictory subgraph containing key nodes and key correlations is obtained. This subgraph can clearly present the minimum structural conditions on which the anomaly formation depends, allowing operation and maintenance personnel to intuitively understand the source of the anomaly, its scope of impact, and its propagation mechanism without relying on experience for guesswork.

[0035] Through the above implementation process, it can be verified that the present invention has good feasibility and stability in complex industrial magnetic field environments. Compared with existing information technology methods that only output anomaly scores or simple alarms, the present invention can accurately identify structural anomalies in magnetic fields under complex noise and time-varying coupling conditions, and at the same time provide clear structured interpretation results. This capability helps to improve the reliability of equipment status perception and provides more sufficient and intuitive technical support for subsequent operation and maintenance decisions, risk assessment and safety assurance.

[0036] Table 1. Overall Performance Comparison of Different Magnetic Field Anomaly Detection Methods

[0037] As can be seen from Table 1, the method of the present invention outperforms the comparative methods in several key indicators such as anomaly detection accuracy, anomaly localization capability, time-lag anomaly recognition capability, and noise resistance stability. However, the improvement is kept within a reasonable range, which is in line with the objective law of gradual improvement of algorithm performance in complex engineering scenarios.

[0038] Regarding anomaly detection accuracy, traditional time-series statistical methods rely primarily on single-time-dimensional features and fixed thresholds, lacking the ability to model spatial coupling relationships in multi-axis, multi-point magnetic field signals, resulting in relatively low accuracy. While deep time-series network-based methods can characterize certain time dependencies, they still treat each sensor signal as a relatively independent input, making it difficult to reflect the propagation relationship of magnetic field signals in spatial structures. Static graph neural network methods introduce spatial topological information, but their graph structure is fixed and cannot adapt to the correlation changes that occur as magnetic field signals evolve over time. This invention further improves anomaly detection accuracy by dynamically constructing heterogeneous spatiotemporal correlation graphs and decoding them under spatiotemporal propagation constraints, while avoiding the instability caused by overfitting.

[0039] In terms of anomaly false alarm rate and false positive rate, the method of this invention shows a more balanced improvement effect. Traditional methods are prone to misjudging local disturbances as anomalies or missing anomalies caused by structural propagation when there is noise and drift interference. Although deep temporal networks and static graph neural networks can alleviate this problem to a certain extent, they still lack explicit constraints on time-delayed propagation. This invention introduces spatiotemporal propagation constraint parameters in the path generation and message passing stages, so that propagation relationships that do not conform to the spatial topology or temporal evolution law are naturally filtered out, thereby reducing false positive rate and effectively reducing false negatives.

[0040] In terms of anomaly localization accuracy and time-lag anomaly recognition rate, this invention has a more significant advantage over other methods. Traditional methods and deep temporal networks mainly rely on overall scoring, making it difficult to accurately locate the source of anomalies. Although static graph neural networks can provide some spatial localization capabilities, the localization results still have biases due to the neglect of time lag factors. This invention constructs a set of spatiotemporal paths that satisfy propagation constraints and calculates path-level messages along the paths, so that the propagation process of anomalies in the spatial and temporal dimensions can be fully characterized, thereby significantly improving the accuracy of anomaly localization and time-lag recognition.

[0041] In terms of noise resistance and result interpretability, the method of this invention also shows a stable advantage. The multi-channel path message aggregation mechanism reduces the impact of single sensor noise on the overall judgment, while the improved MPNN performs message passing under constrained conditions, effectively suppressing the spread of random disturbances in the graph structure. At the same time, through the generation of counterexample trajectories and minimum contradiction subgraphs, the detection results can be presented in a structured form, providing a clear basis for anomaly analysis. This is a capability that existing methods cannot achieve simultaneously.

[0042] Overall, this invention, while maintaining engineering feasibility and numerical rationality, achieves stable detection and interpretable analysis of structural anomalies in complex magnetic field signals. Its performance improvement has a clear source and a complete technical path, and it has good practical application value and prospects for promotion.

[0043] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A spatiotemporal decoding intelligent detection method for magnetic field signals based on graph neural networks, characterized in that, Includes the following steps: Acquire magnetic field signals from a multi-axis, multi-point magnetic field sensor and preprocess them to generate a magnetic field signal time sequence; The magnetic field signal time sequence is segmented according to the sliding time window, and the characteristic components of magnetic field amplitude, direction and rate of change in each window are extracted to generate a window observation state vector sequence. Based on the spatial layout topology of the magnetic field sensor and the temporal sequence and allowable lag relationship corresponding to the sliding time window, a set of spatiotemporal propagation constraint parameters is generated. A dynamic heterogeneous spatiotemporal correlation graph is constructed based on the window observation state vector sequence and the spatiotemporal propagation constraint parameter set. The dynamic heterogeneous spatiotemporal correlation graph includes time edges, spatial edges, coupling edges, and lag edges, and corresponding constraint parameters are configured for each type of edge. In a dynamic heterogeneous spatiotemporal correlation graph, a set of spatiotemporal paths that satisfy the set of spatiotemporal propagation constraints is generated. Path-level messages are calculated along each spatiotemporal path, and a set of node path messages is generated. Perform multi-channel aggregation on the node path message set according to edge type, input it into the improved MPNN for constrained message passing and node state update, and generate a spatiotemporal decoded state sequence; Calculate the structural consistency index and constraint violation degree based on the spatiotemporal decoding state sequence, and generate the spatiotemporal decoding discriminant; Based on the spatiotemporal decoding discriminant, it is determined whether there is a structural anomaly in the magnetic field signal. When a structural anomaly is determined to exist, a counterexample trajectory is generated and the minimum contradiction subgraph is obtained by inversion, thus generating the spatiotemporal decoding intelligent detection result of the magnetic field signal.

2. The intelligent detection method for spatiotemporal decoding of magnetic field signals based on graph neural networks according to claim 1, characterized in that, The preprocessing specifically includes: noise suppression, baseline correction, time alignment, amplitude normalization, anomaly removal, and time sequence reconstruction.

3. The intelligent detection method for spatiotemporal decoding of magnetic field signals based on graph neural networks according to claim 1, characterized in that, The generation of the window observation state vector sequence specifically includes: Set the window length and sliding step size of the sliding time window, and perform sliding segmentation on the magnetic field signal time sequence according to the time index to obtain multiple time windows arranged in chronological order; For each time window, magnetic field signals within the corresponding time index range are selected to form a set of magnetic field signals corresponding to that time window. Amplitude calculation processing is performed on the magnetic field signal set to obtain the amplitude characteristic components corresponding to the time window; Vector synthesis processing is performed on the sampled values ​​of the multi-axis magnetic field signals in each axis of the magnetic field signal set to calculate the magnetic field direction vector of the corresponding time window, which is used as the direction feature component; Perform differential calculation on the magnetic field signals at adjacent sampling times in the magnetic field signal set to obtain the rate of change characteristic component corresponding to the time window; The amplitude feature components, direction feature components, and rate of change feature components corresponding to the same time window are combined in a preset order to generate a window observation state vector corresponding to that time window. Arrange the window observation state vectors corresponding to each time window according to the time index order of the time window to generate a window observation state vector sequence.

4. The intelligent detection method for spatiotemporal decoding of magnetic field signals based on graph neural networks according to claim 1, characterized in that, The generation of the spatiotemporal propagation constraint parameter set specifically includes: Acquire spatial layout information of multi-axis multi-point magnetic field sensors, determine the spatial adjacency relationship between magnetic field sensors based on the relative positional relationship and topological connection relationship between magnetic field sensors, and generate spatial layout topological relationship of magnetic field sensors; Based on the time index relationship of the sliding time window corresponding to the window observation state vector sequence, the time sequence relationship of the sliding time window is generated; Under the constraint of the time sequence, based on the preset maximum allowable time offset range, the time index difference between sliding time windows is compared and filtered to determine the time window pairs that meet the allowable lag conditions, and the allowable lag relationship between sliding time windows is generated. Based on the spatial layout topology of magnetic field sensors, the spatial propagation paths of magnetic field signals between different magnetic field sensors are filtered, and propagation relationships that satisfy the spatial adjacency condition are retained. Corresponding spatial propagation constraint parameters are assigned to the propagation relationships. Based on the temporal sequence of the sliding time window, the propagation direction of the magnetic field signal in the time dimension is constrained, the propagation relationship that meets the temporal sequence requirement is selected, and the propagation relationship is assigned a time propagation constraint parameter; Based on the allowable hysteresis relationship, the propagation range of the magnetic field signal across the sliding time window is limited, the propagation relationship that satisfies the hysteresis condition is selected, and the propagation relationship is assigned a hysteresis propagation constraint parameter; Spatial propagation constraint parameters, temporal propagation constraint parameters, and lag propagation constraint parameters are collected and organized to generate a set of spatiotemporal propagation constraint parameters.

5. The intelligent detection method for spatiotemporal decoding of magnetic field signals based on graph neural networks according to claim 1, characterized in that, The generation of the dynamic heterogeneous spatiotemporal correlation graph specifically includes: Each window observation state vector in the window observation state vector sequence is converted into a graph node, and a corresponding sliding time window time index is attached to each graph node to generate a graph node set. Under the same sliding time window time index, based on the spatial layout topology of the magnetic field sensors in the spatiotemporal propagation constraint parameter set, connections reflecting spatial adjacency relationships are established between graph nodes in the graph node set to form a spatial edge set. Around the graph nodes with time indices of different sliding time windows, according to the time sequence rules determined by the set of spatiotemporal propagation constraint parameters, the node connection relationship that reflects the direction of time evolution is constructed to obtain the set of time edges; For graph node pairs with offset time indices, the allowable lag rules defined in the spatiotemporal propagation constraint parameter set are combined to introduce cross-time window connections between graph nodes that satisfy the lag constraints, generating a set of lag edges. For graph nodes originating from different magnetic field sensors, by combining the topological relationship of the spatial layout of the magnetic field sensors with the corresponding information of the window observation state vector, a connection structure reflecting the magnetic field coupling relationship is constructed between the graph nodes to form a set of coupling edges. Assign edge type identifiers to the spatial edge set, temporal edge set, lag edge set, and coupled edge set respectively, and load the corresponding constraint parameters to different edge types according to the spatiotemporal propagation constraint parameter set; By combining the graph node set, spatial edge set, temporal edge set, lag edge set, coupling edge set, and their corresponding constraint parameters, a dynamic heterogeneous spatiotemporal correlation graph is generated.

6. The intelligent detection method for spatiotemporal decoding of magnetic field signals based on graph neural networks according to claim 1, characterized in that, The generation of the node path message set specifically includes: Based on the dynamic heterogeneous spatiotemporal association graph, combined with the spatiotemporal propagation constraint parameter set, the propagation legality of the connection relationship between graph nodes is determined, and the edges allowed to participate in propagation under the constraints of spatial propagation, temporal propagation and delayed propagation are determined, generating a set of propagable edges; Based on the set of propagable edges, a node sequence from the starting node to the subsequent nodes is constructed in the dynamic heterogeneous spatiotemporal association graph according to the connection order between nodes. The node sequence that satisfies the spatiotemporal propagation constraint parameter set is determined as the spatiotemporal path, and a set of spatiotemporal paths is generated. For each spatiotemporal path in the set of spatiotemporal paths, the corresponding window observation state vector is read sequentially according to the order of the graph nodes in the spatiotemporal path to form the path node state sequence corresponding to the spatiotemporal path. Along each spatiotemporal path, perform aggregation calculations on the path node state sequence to obtain the path-level message corresponding to that spatiotemporal path; For the same graph node, all path-level messages corresponding to the spatiotemporal paths that pass through the graph node are collected to generate the node path message corresponding to the graph node. The node path messages corresponding to all graph nodes in the dynamic heterogeneous spatiotemporal correlation graph are collected to generate a node path message set.

7. The intelligent detection method for spatiotemporal decoding of magnetic field signals based on graph neural networks according to claim 1, characterized in that, The generation of the spatiotemporal decoding state sequence specifically includes: For each graph node in a dynamic heterogeneous spatiotemporal relational graph, the node path messages corresponding to the graph node are split according to the edge types traversed during the generation process, forming a set of multiple types of path messages corresponding to spatial edges, temporal edges, lag edges, and coupling edges respectively. While splitting the message sets of multiple paths, corresponding constraint parameters from the spatiotemporal propagation constraint parameter set are introduced to apply message propagation constraints consistent with their edge types to each type of path message. Based on the set of multiple path messages after the constraints are applied, independent aggregation processing is performed on the path messages corresponding to spatial edges, temporal edges, lag edges, and coupling edges to generate channel-level aggregated messages corresponding to spatial channels, temporal channels, lag channels, and coupling channels, respectively. The aggregated messages of each channel corresponding to the same graph node are combined in a preset order to generate a multi-channel node feature set corresponding to that graph node. The multi-channel node feature set is input into the improved MPNN. Under the continuous constraint of the spatiotemporal propagation constraint parameter set, constrained message passing processing is performed between graph nodes in the dynamic heterogeneous spatiotemporal association graph to generate the updated intermediate state of each graph node in the current sliding time window. The improvements to the MPNN specifically include: Using a multi-channel node feature set as input to the graph node, the graph node can simultaneously represent the information corresponding to spatial propagation, temporal propagation, lag propagation, and magnetic field coupling propagation during message passing. During the message passing process, a spatiotemporal propagation constraint parameter set is introduced to continuously apply propagation constraints to the message passing process between graph nodes in the dynamic heterogeneous spatiotemporal association graph. During message passing and node state update, the message passing input is the node path message after multi-channel aggregation and the application of propagation constraints. Based on updating intermediate states, perform node state update operations on each graph node to generate a spatiotemporal decoding state corresponding to the corresponding sliding time window; Arrange the spatiotemporal decoding states generated under each sliding time window according to the time index order of the sliding time window to generate a spatiotemporal decoding state sequence.

8. The intelligent detection method for spatiotemporal decoding of magnetic field signals based on graph neural networks according to claim 1, characterized in that, The generation of the spatiotemporal decoding discriminant specifically includes: Under the constraints of dynamic heterogeneous spatiotemporal correlation graph, based on the spatiotemporal correlation relationships formed between graph nodes through time edges, spatial edges, coupling edges and lag edges, consistency calculation is performed on the spatiotemporal decoding states corresponding to the interconnected graph nodes in the spatiotemporal decoding state sequence to generate structural consistency index. Under different time index conditions corresponding to the same graph node, the spatiotemporal decoding state corresponding to the graph node in the spatiotemporal decoding state sequence changes with time, and a time series consistency index is generated. Based on structural consistency index and temporal consistency index, the node association relationships that violate the spatiotemporal propagation constraints in the dynamic heterogeneous spatiotemporal association graph are identified, and the constraint violation degree is generated. The structural consistency index and constraint violation degree are jointly aggregated and processed to generate a spatiotemporal decoding discriminant.

9. The intelligent detection method for spatiotemporal decoding of magnetic field signals based on graph neural networks according to claim 1, characterized in that, The generation of the spatiotemporal decoding intelligent detection result of the magnetic field signal specifically includes: Based on the spatiotemporal decoding discrimination value, the nodes and their relationships in the dynamic heterogeneous spatiotemporal correlation graph are judged, the node association regions that meet the anomaly judgment conditions of the spatiotemporal decoding discrimination value are identified, and the structural anomalies of the magnetic field signal are determined. In cases where a structural anomaly is determined to exist in the magnetic field signal, the correlation paths between the anomaly nodes are traced in time index order along the spatiotemporal propagation relationship in the dynamic heterogeneous spatiotemporal correlation graph to generate counterexample trajectories. Based on the counterexample trajectory, the nodes and edges related to the counterexample trajectory in the dynamic heterogeneous spatiotemporal correlation graph are filtered to construct a candidate anomaly subgraph containing the nodes and relationships related to the counterexample trajectory; The candidate abnormal subgraph is inverted. Based on the constraints of minimizing the number of nodes and simplifying the relationships, nodes and relationships in the candidate abnormal subgraph that do not constitute the minimum contradiction structure are gradually pruned to generate the minimum contradiction subgraph. Based on the minimum contradiction subgraph, the abnormal nodes, corresponding time index intervals and edge type information contained therein are collected to generate the spatiotemporal decoding intelligent detection results of magnetic field signals.

Images