Steel structure health monitoring method based on asymmetric anomaly propagation modeling

Abstract

The invention discloses a steel structure health monitoring method based on asymmetric anomaly propagation modeling, and relates to the technical field of steel structure monitoring, and the method specifically comprises the steps: obtaining multi-modal data based on a sensor group arranged on a steel structure; mapping the multi-modal data to corresponding nodes or edges in the steel structure topological graph to form multi-modal features of each node; inputting the multi-modal features into a graph neural network model, modeling a propagation mode of the multi-modal features in a topological graph, identifying an asymmetric propagation path represented by an abnormal signal in a structure, outputting an abnormal state label of a node, calculating a discrimination contribution degree of each modal feature to the abnormal state label, and then performing multi-stage screening; and inputting the abnormal state label and the modal features reserved after screening into a state discrimination model, and outputting corresponding early warning information. The method has the advantages that the position of the abnormal source and the influence range of the abnormal source can be identified more accurately, and the problem of misjudgment caused by assuming a symmetric path in damage propagation modeling in a traditional method is solved.

Description

Technical Field

[0001] This invention relates to the field of steel structure monitoring technology, and in particular to a method for monitoring the health of steel structures based on asymmetric anomaly propagation modeling. Background Technology

[0002] With the continuous expansion of urban infrastructure, steel structures are widely used in bridges, high-rise buildings, factories, and various industrial structures. To ensure their operational safety throughout their entire life cycle, structural health monitoring systems have become a key supporting means. Existing steel structure health monitoring systems typically rely on multiple types of sensors (such as strain gauges, accelerometers, infrared thermal imagers, acoustic emission sensors, etc.) to collect multimodal data, and use feature extraction and fusion algorithms to achieve structural condition assessment and risk warning.

[0003] However, existing technologies still have shortcomings in the following aspects: First, they lack accurate modeling of the propagation process of abnormal responses in structures. Especially in steel structures with complex multi-component connections, abnormal responses often exhibit asymmetric propagation characteristics, while traditional models often assume symmetrical propagation paths, affecting positioning accuracy. Second, multimodal sensor data has redundancy in spatial distribution and sensing capabilities. Current monitoring systems often "process all modes," lacking dynamic identification and suppression mechanisms for invalid or redundant modes, leading to wasted computing resources and increased false alarm rates. Third, in scenarios where sensor deployment is limited or data is missing, the system lacks the ability to infer the overall state of the structure from local observation data, reducing the system's adaptability and practicality. Summary of the Invention

[0004] To address the problems of inaccurate modeling of abnormal signal propagation, redundant interference in multimodal data, and insufficient structural condition discrimination capability under incomplete data conditions in existing steel structure health monitoring methods, this invention provides a steel structure health monitoring method based on asymmetric anomaly propagation modeling and modal feature optimization. This method constructs a connection topology graph between steel structure components and uses a graph neural network to model the propagation path of abnormal responses in the structure to identify anomaly sources and their impact range. Simultaneously, it filters different modal features based on the discriminative contribution of multimodal sensor data, removing low-efficiency modes to improve feature quality and analysis efficiency. Finally, combining retained modal features and anomaly labels, a state classification model is used to generate the health status results for each node, making it suitable for structural condition assessment and risk warning under conditions of limited sensor deployment or missing data.

[0005] To achieve the above-mentioned objectives, this invention provides a method for health monitoring of steel structures based on asymmetric anomaly propagation modeling and modal feature optimization, the method comprising the following steps: Step S1: Deploy a sensor group on the steel structure to be monitored, and preprocess the data collected by the sensor group to obtain multimodal data; Step S2: Based on the connection relationship and spatial distribution between the components in the steel structure, construct a topology diagram of the steel structure, and map the multimodal data to the corresponding nodes or edges in the topology diagram of the steel structure to form the multimodal features of each node. Step S3: Input the multimodal features into the graph neural network model, model the propagation mode of the multimodal features in the topology graph, identify the asymmetric propagation path of the abnormal signal in the structure, and output the abnormal state label of the node to determine the abnormal source point and its propagation influence range in the steel structure. Step S4: Based on the abnormal state labels obtained in step S3, and combined with the multimodal features in step S2, calculate the contribution of each modal feature to the discrimination of the abnormal state labels, and use it as the modal validity parameter. Then, remove modal features whose validity parameters are lower than the preset threshold, and construct an objective function to select modal features whose mutual information values ​​are higher than the set threshold and whose redundancy is lower than the set upper limit. Step S5: Input the abnormal state labels obtained in step S3 and the modal features selected in step S4 into the state discrimination model to generate the health status classification results of each node in the steel structure, and output the corresponding early warning information according to the set health level standards.

[0006] In step S1, the sensor group includes an acoustic emission sensor, an acceleration sensor, a fiber optic strain sensor, and a fiber optic temperature sensor, which are used to collect acoustic signals, vibration signals, strain data, and temperature data, respectively, and are installed at key connection nodes or stress concentration areas of the steel structure components.

[0007] In step S2, the constructed steel structure topology graph uses the connection relationship of steel structure components as the basis for establishing edges between nodes, uses the connection nodes in the structural layout diagram as nodes in the graph structure, establishes edges using the physical connection relationship between nodes, and represents the graph structure in the form of an adjacency matrix; the mapping method of the multimodal data is as follows: according to the installation correspondence between each sensor and structural component, the collected data is mapped to the corresponding nodes in the topology graph according to the structural node number bound to the sensor, and a multimodal feature vector of each node is constructed.

[0008] In step S3, the graph neural network model employs a heterogeneous graph attention network. The method for constructing the heterogeneous graph attention neural network includes the following steps: a) Use the modal features calculated in step S2 as the initial input features for each node; b) Based on modal type, propagation direction and structural connection information, define a set of multiple edge types and assign a weight vector to each edge, consisting of three physical attributes: propagation delay, frequency drift and signal-to-noise ratio; c) During the propagation of adjacent edge information, the propagation weights of each type of edge are calculated based on the attention mechanism to form a direction-weighted attention coefficient; d) Iteratively update the node representation in the multi-layer neural network and output the abnormal state label of the node.

[0009] In step S3, the process of determining the abnormal state label of the node includes: constructing a function model reflecting the degree of deviation of abnormal propagation based on the modal feature differences, adjacency weights, time-domain response characteristics and frequency drift factors among the nodes in the graph structure, and assigning values ​​to the abnormal state label of the node according to the model to characterize the degree of deviation of its normal propagation behavior from the neighborhood.

[0010] The health status classification model mentioned in step S5 is a structural state discrimination model based on long short-term memory network. The input is the abnormal state label and corresponding modal features of each node. The model introduces a time sliding window mechanism during training to construct a time series input and output the health level label of the node in the future prediction period.

[0011] The beneficial effects of this invention are as follows: By constructing a topological graph structure among steel structure components and combining it with a graph neural network to model the asymmetric propagation path of abnormal responses in the structure, this invention can more accurately identify the location of the abnormal source and its influence range, solving the problem of misjudgment caused by the assumption of symmetric paths in damage propagation modeling in traditional methods; at the same time, the modal validity parameter calculation and modal feature optimization mechanism in this invention can effectively suppress the interference of redundant or invalid modal data, reducing the false alarm rate and computational resource consumption; in addition, the state classification model, combined with the selected high-quality features, enables accurate assessment of the overall structural health status in scenarios with insufficient local sensor data, enhancing the adaptability and stability of the system. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of Embodiment 1 of the present invention. Detailed Implementation

[0013] To clearly illustrate the technical features of this solution, the following detailed implementation method will be used to explain the solution.

[0014] Example 1 This invention provides a method for health monitoring of steel structures based on anomaly diffusion modeling and modal optimization. The method includes the following steps: Step S1: Deploy a sensor group on the steel structure to be monitored, and preprocess the data collected by the sensor group to obtain multimodal data. The sensor group includes an acoustic emission sensor, an acceleration sensor, a fiber optic strain sensor, and a fiber optic temperature sensor, which are used to collect acoustic signals, vibration signals, strain data, and temperature data, respectively, and are installed at key connection nodes or stress concentration areas of the steel structure components.

[0015] Step S2: Based on the connection relationship and spatial distribution between the components in the steel structure, construct a topology diagram of the steel structure, and map the multimodal data to the corresponding nodes or edges in the topology diagram of the steel structure to form the multimodal features of each node.

[0016] The constructed steel structure topology uses the connection relationship of steel structure components as the basis for establishing the edges between nodes. The connection nodes in the structural layout diagram are used as the nodes in the graph structure. The edges are established by the physical connection relationship between the nodes, and the graph structure is represented by the adjacency matrix. The mapping method of multimodal data is as follows: according to the installation correspondence between each sensor and structural component, the collected data is mapped to the corresponding nodes in the topology diagram according to the structural node number bound to the sensor, and the multimodal feature vector of each node is constructed.

[0017] Step S3: Input the multimodal features into the graph neural network model, model the propagation mode of the multimodal features in the topology graph, identify the asymmetric propagation path of the abnormal signal in the structure, and output the abnormal state label of the node to determine the abnormal source point and its propagation influence range in the steel structure.

[0018] In this embodiment, the graph neural network model uses a heterogeneous graph attention network, and its construction method includes the following steps: a) Use the modal features calculated in step S2 as the initial input features for each node; b) Based on modal type, propagation direction and structural connection information, define a set of multiple edge types and assign a weight vector to each edge, consisting of three physical attributes: propagation delay, frequency drift and signal-to-noise ratio; c) During the propagation of adjacent edge information, the propagation weights of each type of edge are calculated based on the attention mechanism to form a direction-weighted attention coefficient; d) Iteratively update the node representation in the multi-layer neural network and output the abnormal state label of the node.

[0019] In S3, the process of judging the abnormal state label of a node includes: constructing a function model that reflects the degree of deviation of abnormal propagation based on the modal feature differences, adjacency weights, time-domain response characteristics and frequency drift factors among nodes in the graph structure, and assigning values ​​to the abnormal state labels of nodes according to the model to characterize the degree of deviation of their normal propagation behavior from the neighborhood.

[0020] Step S4: Based on the abnormal state labels obtained in step S3, and combined with the multimodal features in step S2, calculate the contribution of each modal feature to the discrimination of the abnormal state labels, and use it as the modal validity parameter. Then, remove modal features whose validity parameters are lower than the preset threshold, and construct an objective function to select modal features whose mutual information values ​​are higher than the set threshold and whose redundancy is lower than the set upper limit. Step S5: Input the abnormal state labels obtained in Step S3 and the modal features retained in Step S4 into the state discrimination model to generate the health status classification results of each node in the steel structure, and output corresponding early warning information according to the set health level standards. The health status classification model is a structural state discrimination model based on a long short-term memory network. The inputs are the abnormal state labels and corresponding modal features of each node. During training, the model introduces a time-sliding window mechanism to construct a time-series input and outputs the health level labels of nodes within the future prediction period.

[0021] Example 2 This invention provides a method for health monitoring of steel structures based on anomaly diffusion modeling and modal optimization. The specific method includes: 1. Install sensor arrays on the steel structure to be monitored, and perform synchronous processing, noise reduction and normalization on the data collected by the sensor arrays to obtain multimodal data.

[0022] The sensor group includes an acoustic emission sensor, an acceleration sensor, a fiber Bragg grating strain sensor, and a fiber Bragg grating temperature sensor, which are used to collect acoustic signals, vibration signals, strain data, and temperature data, respectively. They are installed at key connection nodes or stress concentration areas of the steel structure components. The noise reduction process includes using a combined method of wavelet transform and adaptive Kalman filtering to reduce the noise of the collected signals.

[0023] 2. Based on the connection relationships and spatial distribution between the components in the steel structure, construct a topology diagram of the steel structure, and map the normalized multimodal data to the corresponding nodes or edges in the topology diagram of the steel structure to form the multimodal features of each node.

[0024] The steel structure topology graph constructed above uses the connection relationship of steel structure components as the basis for establishing the edges between nodes, uses the connection nodes in the structural layout diagram as the nodes in the graph structure, establishes the edges using the physical connection relationship between nodes, and represents the graph structure in the form of an adjacency matrix; the mapping method of multimodal data is as follows: according to the installation correspondence between each sensor and structural component, the collected data is mapped to the corresponding nodes in the topology graph according to the structural node number bound to the sensor, and a multimodal feature vector of each node is constructed.

[0025] The multimodal features of each node are obtained using the following formula:

[0026] in, Represents a node The fusion modal eigenvalues, Represents a node modal number, Indicates the first m Each modality at the node The collected values, Representing modes m The global average, Representing modes m The amplitude normalization coefficient, Representing modes m The local coefficient of variation, The Softplus function has the following expression: .

[0027] Third, the multimodal features are input into the graph neural network model to model the propagation mode of the multimodal features in the topology graph, identify the asymmetric propagation path of the abnormal signal in the structure, and output the abnormal state label of the node to determine the abnormal source point and its propagation influence range in the steel structure. The graph neural network model employs a heterogeneous graph attention network (HNN), which features multimodal input channels and multi-type edge modeling capabilities, supporting asymmetric information propagation between nodes and modality-dependent weighted aggregation operations. The construction method of the heterogeneous graph attention neural network includes the following steps: a) The modal features obtained in step two are... 𝑖 As the initial input features for each node; b) Based on modal type, propagation direction and structural connection information, define a set of multiple edge types and assign a weight vector to each edge, consisting of three physical attributes: propagation delay, frequency drift and signal-to-noise ratio; c) During the propagation of adjacent edge information, the propagation weights of each type of edge are calculated based on the attention mechanism to form a direction-weighted attention coefficient; d) Iteratively update the node representation in the multi-layer neural network and output the abnormal state label of the node.

[0028] The process of determining the abnormal state label of a node includes: constructing a function model reflecting the degree of deviation in abnormal propagation based on the modal feature differences, adjacency weights, temporal response characteristics, and frequency drift factors among nodes in the graph structure; and assigning an abnormal state label to the target node according to this model to characterize the degree to which it deviates from the normal propagation behavior of its neighborhood. The abnormal state label of a node is calculated based on the following abnormal deviation mapping function:

[0029] in, Represents a node Abnormal deviation values, For nodes The normalization coefficient, For nodes The set of neighboring nodes, For nodes modal aggregation value, For nodes modal mean, For nodes The normalized value of the modal signal-to-noise ratio, For the edge → The propagation weight, This is the frequency drift value. This is the diffusion sensitivity coefficient. To delay the transmission time, It is the thermal-vibration interference factor.

[0030] Fourth, based on the abnormal state labels obtained in step three, and combined with the multimodal features in step two, calculate the contribution of each modal feature to the discrimination of the abnormal state labels, and use it as the modal validity parameter; first, remove modal features whose validity parameters are lower than the preset threshold, and then construct an objective function to select modal features whose mutual information values ​​are higher than the set threshold and whose redundancy is lower than the set upper limit, for subsequent health status judgment.

[0031] Modal validity parameters are calculated using the following function:

[0032] in, V m Representing modes m Validity parameters, Represents a node The eigenvalues ​​of mode m, and Modal m Mean and standard deviation across all nodes For nodes This is an indicator variable for whether something is marked as abnormal; a value of 1 indicates an abnormality, and a value of 0 indicates normal operation. N The number of nodes participating in the calculation. To prevent non-zero constants with a denominator of zero, modal features with validity parameters below a set threshold are removed based on the calculation results, and the remaining modal features are input to the state judgment step.

[0033] Modal features with high mutual information and low redundancy are selected by optimizing the objective function, specifically: Calculate the mutual information value between each modal feature and the node's abnormal state label, and construct an optimization objective function; based on the solution of the optimization objective function, select modal features with mutual information values ​​higher than a preset threshold and low redundancy as the modes retained in step S4; the specific optimization objective function is as follows:

[0034] in, Indicates the first m Feature vectors of each modality This represents the node abnormal status label output in step S3. Represents the mutual information function. This is the redundancy suppression coefficient.

[0035] Fifth, the abnormal state labels obtained in step three and the modal features selected in step four are input into the state discrimination model to generate the health status classification results of each node in the steel structure, and corresponding early warning information is output according to the set health level standards. In this embodiment, the health status classification model is a structural state discrimination model based on a Long Short-Term Memory (LSTM) network. The inputs are the abnormal state labels and corresponding modal features of each node. During the training process, a time sliding window mechanism is introduced to construct a time series input, and the output is the health level label of the node within the future prediction period.

[0036] The determination of the health level label is based on the following structural degradation scoring function:

[0037] in, For nodes The degradation score, For this node at time Abnormal deviation values, For modality m Time series feature values, For modality m The average value within the time window, For modality m The local variation coefficient, α , β These are adjustable weighting coefficients. T This represents the length of the time window.

[0038] In this embodiment, the early warning information is determined based on the health level labels generated by the state discrimination model. Preferably, a three-level early warning standard can be set: when the abnormal state score of a node is lower than the first threshold, it is marked as "normal"; when it is higher than the second threshold, it is marked as "dangerous"; and when it is between the two, it is marked as "early warning". The system can trigger the corresponding level of alarm mechanism accordingly to achieve real-time monitoring and response management of the health status of the steel structure. In addition, it should be noted that the output method of the health status classification results and the implementation of the early warning information triggering mechanism in this embodiment can be adjusted according to the actual needs of the project. For example, the health level can be divided into two, three, or more levels, and the early warning information can also be output in various forms such as text reminders, audible and visual alarms, and remote communication. The specific implementation method can be selected and determined according to the actual application environment, and all of them should be considered within the protection scope of this invention.

[0039] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A steel structure health monitoring method based on asymmetric anomaly propagation modeling, characterized in that, The method comprises the following steps: Step S1, laying out a sensor group on a steel structure to be monitored, and preprocessing data collected by the sensor group to obtain multi-modal data; Step S2, constructing a steel structure topology graph according to the connection relationship and spatial distribution between components in the steel structure, and mapping the multi-modal data to corresponding nodes or edges in the steel structure topology graph to form multi-modal features of each node; Step S3, inputting the multi-modal features into a graph neural network model to model the propagation mode of the multi-modal features in the topology graph, identify the asymmetric propagation path of the abnormal signal in the structure, and output an abnormal state label of the node for determining the abnormal source point and its propagation influence range in the steel structure; Step S4, based on the abnormal state label obtained in step S3, combining the multi-modal features in step S2, calculating the discriminant contribution of each modal feature to the abnormal state label as a modal effectiveness parameter, then eliminating the modal features with effectiveness parameters below a preset threshold, and then constructing an objective function to select modal features with mutual information values higher than a set threshold and redundancy lower than a set upper limit; Step S5, inputting the abnormal state label obtained in step S3 and the modal features selected in step S4 into a state discrimination model to generate health state classification results of each node in the steel structure, and outputting corresponding warning information according to the set health level standard.

2. The method of claim 1, wherein, In the step S1, the sensor group comprises acoustic emission sensors, acceleration sensors, fiber Bragg strain sensors and fiber Bragg temperature sensors, which are respectively used to collect acoustic signals, vibration signals, strain data and temperature data, and are respectively installed at key connection nodes or stress concentration areas of steel structure components.

3. The method of claim 1, wherein, In the step S2, the constructed steel structure topology graph takes the connection relationship of steel structure components as the basis for establishing the edges between nodes, uses the connection nodes in the structure layout as the nodes in the graph structure, establishes the edges between the nodes according to the physical connection relationship, and represents the graph structure in the form of an adjacency matrix; the mapping mode of the multi-modal data is: according to the installation correspondence between each sensor and the structure component, mapping the collected data to the corresponding node in the topology graph according to the structure node number bound by the sensor, and constructing a multi-modal feature vector of each node.

4. The method of claim 1, wherein, In the step S3, the graph neural network model adopts a heterogeneous graph attention network.

5. The method of claim 4, wherein, The construction method of the heterogeneous graph attention neural network comprises the following steps: a) taking the modal features calculated in step S2 as the initial input features of each node; b) defining a multi-type edge set according to the modal type, propagation direction and structure connection information, and assigning a weight vector composed of three physical attributes of propagation delay, frequency drift and signal-to-noise ratio to each edge; c) during the propagation of adjacent edge information, the propagation weights of each type of edge are calculated based on the attention mechanism to form a directional weighted attention coefficient; d) iteratively updating the node representation in the multi-layer neural network, and outputting the abnormal state label of the node.

6. The method of claim 5, wherein, The judgment process of the abnormal state label of the node in the step S3 includes: constructing a function model reflecting the deviation degree of abnormal propagation based on the modal feature difference, the adjacency weight, the time domain response characteristic and the frequency drift factor between nodes in the graph structure, and assigning the abnormal state label of the node according to the model, so as to represent the degree of deviation from the normal propagation behavior of the neighborhood.

7. The method of claim 1, wherein, The health state classification model in the step S5 is a structure state discrimination model based on a long short-term memory network, the input is the abnormal state label of each node and the corresponding modal feature, and the model introduces a time sliding window mechanism in the training process to construct a time sequence input, and outputs the health level label of the node in a future prediction period.

Images