A method for representing spatial directed associations of multi-type monitoring data based on multi-channel spatiotemporal graph convolutional networks

Abstract

The present invention discloses a method for characterizing spatial directed associations of multi-type monitoring data based on a multi-channel spatiotemporal graph convolutional network, the method comprising the following steps: 1. Collecting health monitoring data of a large-span cable-stayed bridge, performing data preprocessing to remove random effects, and obtaining long-term trend data; 2. Standardizing the long-term trend data of each sensor and creating a data set; 3. Modeling the spatiotemporal associations of multi-type structural health monitoring data in the form of a directed graph; 4. Establishing a multi-channel spatiotemporal graph convolutional network model; 5. Designing a multi-channel spatiotemporal graph convolutional network loss function; 6. Training the multi-channel spatiotemporal graph convolutional network model, and after the training is completed, outputting each first-order mapping matrix to obtain spatial directed association characterizations of multi-type monitoring data. The present invention solves the shortcomings of the existing bridge structure status assessment work involving too little data and a single data type, and realizes the effective use of massive monitoring data.

Description

Technical Field

[0001] The present invention belongs to the field of bridge engineering, and relates to a method for characterizing spatial directed associations of multi-type sensor monitoring data, and specifically to a method for characterizing spatial directed associations of multi-type monitoring data based on a multi-channel spatiotemporal graph convolutional network. Background Art

[0002] As a means to maintain structural health and extend the service life of structures, structural health monitoring systems have attracted much attention. Through uninterrupted operation, the sensor array on the structural health monitoring system continuously records a large amount of information including the bridge site environment, vehicle loads, and dynamic and static responses of the structure in real time, providing a basis for bridge structure status assessment based on health monitoring data.

[0003] As an important branch of condition assessment, the widely existing spatiotemporal correlation between load-response and response-response is investigated based on data-driven methods, and a mathematical model that implicitly represents the structural state is established, which effectively avoids the problem of ill-posed equations and non-unique results caused by limited measurement points and incomplete observation data during solution. The core function of this type of model is to describe the mapping relationship between multiple measurement sensors. However, the monitoring data involved in the current bridge structure condition assessment based on monitoring data are mostly single-type monitoring data, and the amount of data involved is still too small compared with the amount of data generated by the health monitoring system of the serving bridge structure, resulting in a relatively low utilization rate of massive monitoring data. At the same time, the single monitoring data ignores the collaborative working mechanism of various components in the bridge structure, making it difficult for the training model to discover the deep-level characteristics of the monitoring data. Therefore, establishing a model that can reveal the spatial directed correlation representation of multiple types of sensors can improve the utilization of massive monitoring data and explore the deep-level characteristics of monitoring data. It is of great significance to the bridge structure condition assessment based on monitoring data, and it is also a difficult problem that needs to be solved urgently. Summary of the invention

[0004] In order to solve the shortcomings of low utilization rate and insufficient correlation mining of existing massive monitoring data, the present invention provides a method for characterizing spatial directed correlations of multi-type monitoring data based on a multi-channel spatiotemporal graph convolutional network. This method quantitatively reveals the spatial directed correlations between any sensor pairs with intuitive and sparse graphs, fully improving the utilization rate and mining degree of massive monitoring data, solving the shortcomings of too little data and single data type involved in existing bridge structure status assessment work, and realizing the effective utilization of massive monitoring data. It is suitable for the operation of health monitoring systems with multiple types of monitoring data.

[0005] The objective of the present invention is achieved through the following technical solutions:

[0006] A method for representing spatial directed associations of multi-type monitoring data based on a multi-channel spatiotemporal graph convolutional network comprises the following steps:

[0007] Step 1: Collect three types of long-term structural health monitoring data: cable force, main beam deflection, and cable tower inclination on any side of the long-span cable-stayed bridge. Assuming that the trend data is a constant value within 10 minutes, perform data preprocessing to remove random effects and obtain long-term trend data.

[0008] Step 2: Standardize the long-term trend data of each sensor and create a dataset of multi-type structural health monitoring data with a 1-hour time window. The training instance length of each sensor in the dataset is 7, including data from the past hour to the current time.

[0009] Step 3: Model the spatiotemporal correlation of multi-type structural health monitoring data in the form of a directed graph, represent each sensor as a vertex on the directed graph, represent the sensor reading as a sequence signal on the vertex, represent the correlation of sensor pairs as a directed edge on the graph, and the strength of the correlation is proportional to the weight of the directed edge;

[0010] Step 4: Generalize the "adjacency matrix" in the field of graph theory to a "mapping matrix", and use several learnable normalized mapping matrices to establish a multi-channel spatiotemporal graph convolutional network model;

[0011] Step 5: Design a multi-channel spatiotemporal graph convolutional network loss function;

[0012] Step 6: Train the multi-channel spatiotemporal graph convolutional network model established in step 4. After the training is completed, output each first-order mapping matrix to obtain the spatial directed correlation representation of multiple types of monitoring data.

[0013] Compared with the prior art, the present invention has the following advantages:

[0014] In view of the low utilization rate of massive monitoring data and the inability to reveal the deep-level characteristics of monitoring data in existing bridge structure status assessment methods based on monitoring data, the present invention proposes a sensor space directed correlation representation method that integrates multiple types of monitoring data. The method takes into account the collaborative working mechanism of several components of the structure and uses several learnable mapping matrices to quantitatively reveal the directed correlation between sensor pairs, thereby providing a basis for subsequent status assessment work. The present invention can also be extended to the multi-type sensor space directed correlation representation of other types of bridges and even other structural health monitoring systems, providing guidance for the normal operation and safe maintenance of the structure. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of the process of the present invention;

[0016] Figure 2 This is a schematic diagram of data preprocessing in step 1 of the embodiment;

[0017] Figure 3This is a schematic diagram of the structure of a multi-channel spatiotemporal graph convolutional network in step 4 of the embodiment;

[0018] Figure 4 Two results of the visualized spatial directed association representation of step six of the embodiment are shown, (a) is a schematic diagram of the spatial directed association representation between the tension sensors, and (b) is a schematic diagram of the spatial directed association representation between the deflection sensor and the tilt sensor. DETAILED DESCRIPTION

[0019] The technical solution of the present invention is further described below in conjunction with the accompanying drawings, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be included in the protection scope of the present invention.

[0020] The present invention provides a method for characterizing spatial directed associations of multi-type monitoring data based on a multi-channel spatiotemporal graph convolutional network. Figure 1 As shown, the method comprises the following steps:

[0021] Step 1: Collect three types of long-term structural health monitoring data: cable tension, main beam deflection, and tower inclination on any side of the long-span cable-stayed bridge. Assuming that the trend data is a constant value within 10 minutes, perform data preprocessing to remove random effects and obtain long-term trend data.

[0022] In this step, data preprocessing to remove random effects is achieved by performing a sliding average operation.

[0023] Step 2: Standardize the long-term trend data of each sensor and create a dataset of multi-type structural health monitoring data with a 1-hour time window. The training instance length of each sensor in the dataset is 7, including data from the past hour to the current time.

[0024] In this step, the standardization of each sensor trend data is achieved by subtracting the mean of the sensor trend data from the trend data and dividing it by the standard deviation of the sensor trend data.

[0025] Step 3: Model the spatiotemporal correlation of multi-type structural health monitoring data in the form of a directed graph. Each sensor is represented as a vertex on the directed graph, the sensor reading is represented as a sequence signal on the vertex, and the correlation between sensor pairs is represented as a directed edge on the graph. The strength of the correlation is proportional to the weight of the directed edge.

[0026] Step 4: Generalize the "adjacency matrix" in the field of graph theory to a "mapping matrix" and use several learnable normalized mapping matrices to establish a multi-channel spatiotemporal graph convolutional network model.

[0027] In this step, the generalization of the "adjacency matrix" to the "mapping matrix" is achieved by not restricting the shape of the matrix. In the field of graph theory, the "adjacency matrix" is defined as a square matrix whose size is equal to the number of graph vertices; the size of the rows and columns of the "mapping matrix" in the present invention is the number of two types of sensors, respectively, and can be used to characterize the correlation between any pair of sensors.

[0028] In this step, the multi-channel spatiotemporal graph convolutional network model includes a multi-channel spatial graph convolution layer, a temporal one-dimensional convolution layer, a fully connected layer, and a fusion layer, which respectively perform multi-channel spatial graph convolution operations, temporal one-dimensional convolution operations, dimensionality reduction operations, and fusion operations. The specific structure is as follows:

[0029] The L1-1 layer is a multi-channel spatial graph convolution layer with an input dimension of N. c ×7×1,N d ×7×1,N s ×7×1, a multi-channel spatial graph convolution operation is performed on the input of each time step. The number of input features of the multi-channel spatial graph convolution layer is 1, the number of output features is 64, and the output dimension is N c ×7×64、N d ×7×64、N s ×7×64, the activation function is ReLU;

[0030] The L1-2 layer, connected to the L1-1 layer, is a temporal one-dimensional convolutional layer with an input dimension of N. c ×7×64、N d ×7×64、N s ×7×64, perform a one-dimensional convolution operation in the time domain on the input of each sensor, the convolution kernel size is 3, the number is 64, and the output dimension is N c ×5×64,N d ×5×64,N s ×5×64, the activation function is ReLU;

[0031] The L1-3 layer, connected to the L1-2 layer, is a temporal one-dimensional convolution layer with an input dimension of N. c ×5×64,N d ×5×64,N s ×5×64, perform a one-dimensional convolution operation in the time domain on the input of each sensor, the convolution kernel size is 3, the number is 64, and the output dimension is N c ×3×64,N d ×3×64,N s ×3×64, the activation function is ReLU;

[0032] The L1-4 layer, connected to the L1-3 layer, is a temporal one-dimensional convolution layer with an input dimension of N. c ×3×64,N d ×3×64,Ns ×3×64, perform a one-dimensional convolution operation in the time domain on the input of each sensor, the convolution kernel size is 3, the number is 64, and the output dimension is N c ×1×64,N d ×1×64,N s ×1×64, the activation function is ReLU;

[0033] The L2-1 layer, connected to the L1-2 layer, is a multi-channel spatial graph convolution layer with an input dimension of N. c ×5×64,N d ×5×64,N s ×5×64, a multi-channel spatial graph convolution operation is performed on the input of each time step. The number of input features and output features of the multi-channel spatial graph convolution layer is 64, and the output dimension is N c ×5×64,N d ×5×64,N s ×5×64, the activation function is ReLU;

[0034] The L2-2 layer, connected to the L2-1 layer, is a temporal one-dimensional convolution layer with an input dimension of N. c ×5×64,N d ×5×64,N s ×5×64, perform a one-dimensional convolution operation in the time domain on the input of each sensor, the convolution kernel size is 3, the number is 64, and the output dimension is N c ×3×64,N d ×3×64,N s ×3×64, the activation function is ReLU;

[0035] The L2-3 layer, connected to the L2-2 layer, is a temporal one-dimensional convolution layer with an input dimension of N. c ×3×64,N d ×3×64,N s ×3×64, perform a one-dimensional convolution operation in the time domain on the input of each sensor, the convolution kernel size is 3, the number is 64, and the output dimension is N c ×1×64,N d ×1×64,N s ×1×64, the activation function is ReLU;

[0036] The L3-1 layer, connected to the L2-2 layer, is a multi-channel spatial graph convolution layer with an input dimension of N. c ×3×64,N d ×3×64,N s ×3×64, a multi-channel spatial graph convolution operation is performed on the input of each time step. The number of input features and output features of the multi-channel spatial graph convolution layer is 64, and the output dimension is N c ×3×64,Nd ×3×64,N s ×3×64, the activation function is ReLU;

[0037] The L3-2 layer, connected to the L3-1 layer, is a temporal one-dimensional convolution layer with an input dimension of N. c ×3×64,N d ×3×64,N s ×3×64, perform a one-dimensional convolution operation in the time domain on the input of each sensor, the convolution kernel size is 3, the number is 64, and the output dimension is N c ×1×64,N d ×1×64,N s ×1×64, the activation function is ReLU;

[0038] The L4 layer is connected to the L1-4 layer, the L2-3 layer, and the L3-2 layer. It is a fully connected layer with an input dimension of N. c ×1×64,N d ×1×64,N s ×1×64, performs linear transformation operations, and the output dimension is N c 、N d 、N s , no activation function;

[0039] The L5 layer is connected to the L4 layer and is a fusion layer. The input dimension is N c 、N d 、N s , perform fusion operation, the output dimension is N c 、N d 、N s , no activation function;

[0040] In this step, the formula for the multi-channel spatial graph convolution operation is:

[0041]

[0042] In the formula, are the cable forces, main beam deflections, and tower tilt inputs at each time step in the l-th multi-channel spatial graph convolution layer, and their dimensions are N c ×F (l) 、N d ×F (l) 、N s ×F (l) , N c 、N d 、N s are the number of cable force sensors, deflection sensors and tilt sensors considered, respectively, and F (l) is the number of input features of the l-th multi-channel spatial graph convolutional layer, F (1) =1, F(2) =64, F (3) =64; It is the lth-order mapping matrix of the directed correlation between the cable forces and the cable forces in space, with a dimension of N c ×N c ; It is the lth-order mapping matrix for learning the directed spatial correlation between the main beam deflection and the cable force, with a dimension of N c ×N d ; All mapping matrices are row normalized; is the weight matrix to be trained for the l-th multi-channel spatial graph convolution layer, with a dimension of F (l) ×C (l) , C (l) is the number of output features of the l-th multi-channel spatial graph convolutional layer, C (l) ≡64; is the cable force, deflection, and tilt output of the l-th multi-channel spatial graph convolution layer at each time step, and their dimensions are N c ×C (l) 、N d ×C (l) 、N s ×C (l) ;|| is the splicing operation; It is a 1×1 convolution kernel that reduces the dimension of multi-channel information after three graph convolution operations, and the number of channels is 3; the rest of the definitions are the same as above.

[0043] In this step, the formula for the fusion operation is:

[0044]

[0045] In the formula, is the model output, is the regression value of the cable force of the l-th order neighboring vertex; the rest of the definitions are the same as above; k l is a learnable parameter representing the contribution of the l-th order neighborhood vertex to the target vertex regression.

[0046] Step 5. Design the loss function of the multi-channel spatiotemporal graph convolutional network, and use the mean square error of three types of data and the sum of three penalty terms as the loss function. The functions of the three penalty terms are to appropriately increase the contribution of heterogeneous monitoring data, avoid the participation of each vertex information in its own regression, and ensure that the first-order neighborhood vertices dominate.

[0047] In this step, the loss function of the multi-channel spatiotemporal graph convolutional network is:

[0048]

[0049] Where, MSE(·) is the mean square error; PH In order to appropriately increase the penalty items for the contribution of heterogeneous monitoring data, the specific ones are:

[0050]

[0051] In the formula, λ h,l,dc is the penalty term for increasing the contribution of the main beam deflection to the cable force regression in the l-th multi-channel spatial graph convolution layer; is the convolution kernel parameter from the cable force to the cable force in the lth multi-channel spatial graph convolution layer; the rest of the definitions are the same as above.

[0052] P I In order to avoid the penalty item of each vertex information acting on itself, the specific items are:

[0053]

[0054] In the formula, λ I,cc To avoid the penalty term where the cable force of the measuring point returns to itself along the cable force of the inclined cable; is the first-order mapping matrix that reveals the directed correlation between the cable forces and the cable forces; tr(·) is the trace of the matrix. O To ensure that the penalty term of the first-order neighborhood vertices dominates, it is:

[0055]

[0056] In the formula, λ o,l A penalty term to ensure that the first-order neighbor vertices dominate over the l-th-order neighbor vertices.

[0057] Step 6: Train the multi-channel spatiotemporal graph convolutional network model. After the training is completed, output each first-order mapping matrix to obtain the spatial directed correlation representation of multiple types of monitoring data.

[0058] Example:

[0059] This embodiment provides a method for characterizing spatial directed associations of multi-type monitoring data based on a multi-channel spatiotemporal graph convolutional network, the method comprising the following steps:

[0060] Step 1: Figure 2 As shown in the figure, the structural health monitoring data of cable tension, main beam deflection and tower inclination of a cable-stayed bridge health monitoring system in the first two years were preprocessed. The data of 42 cable tension sensors, 17 hydraulic transmitters and 3 inclinometers were selected. It was assumed that the trend data was a constant value within 10 minutes. Data preprocessing was performed to remove random effects and obtain long-term trend data.

[0061] Step 2: The long-term trend data of each sensor is normalized by subtracting the mean of the trend data of the sensor and then dividing it by the standard deviation of the trend data of the sensor. A data set is created with a time window of 1 hour. The length of the training instance of each sensor in the data set is 7, including data from the past hour to the current time.

[0062] Step 3: Model the spatiotemporal correlation of multi-type structural health monitoring data in the form of a directed graph. Each sensor is represented as a vertex on the directed graph, the sensor reading is represented as a sequence signal on the vertex, and the correlation between sensor pairs is represented as a directed edge on the graph. The strength of the correlation is proportional to the weight of the directed edge.

[0063] Step 4: Generalize the “adjacency matrix” in the field of graph theory to a “mapping matrix” and use several learnable normalized mapping matrices to establish a multi-channel spatiotemporal graph convolutional network model, such as Figure 3 As shown, it includes a multi-channel spatial graph convolution layer, a temporal one-dimensional convolution layer, a fully connected layer and a fusion layer, which respectively perform multi-channel spatial graph convolution operations, temporal one-dimensional convolution operations, dimensionality reduction operations and fusion operations.

[0064] Step 5. Design the loss function of the multi-channel spatiotemporal graph convolutional network, and use the mean square error of three types of data and the sum of three penalty terms as the loss function. The three penalty terms are to appropriately increase the contribution of heterogeneous monitoring data, avoid the participation of each vertex information in its own regression, and ensure that the first-order neighborhood vertices dominate.

[0065] Step 6: Train the multi-channel spatiotemporal graph convolutional network model established in step 4. After the training is completed, output each first-order mapping matrix to obtain the spatial directed correlation representation of multiple types of monitoring data, such as Figure 4 shown.

Claims

1. A method for representing spatial directed associations of multi-type monitoring data based on a multi-channel spatiotemporal graph convolutional network, characterized in that The method comprises the following steps: Step 1: Collect three types of long-term structural health monitoring data: cable force, main beam deflection, and cable tower inclination on any side of the long-span cable-stayed bridge. Assuming that the trend data is a constant value within 10 minutes, perform data preprocessing to remove random effects and obtain long-term trend data. Step 2: Standardize the long-term trend data of each sensor and create a dataset of multi-type structural health monitoring data with a 1-hour time window. The training instance length of each sensor in the dataset is 7, including data from the past hour to the current time. Step 3: Model the spatiotemporal correlation of multi-type structural health monitoring data in the form of a directed graph, represent each sensor as a vertex on the directed graph, represent the sensor reading as a sequence signal on the vertex, represent the correlation of sensor pairs as a directed edge on the graph, and the strength of the correlation is proportional to the weight of the directed edge; Step 4: Generalize the "adjacency matrix" in the field of graph theory to a "mapping matrix", and use several learnable normalized mapping matrices to establish a multi-channel spatiotemporal graph convolutional network model. The multi-channel spatiotemporal graph convolutional network model includes a multi-channel spatial graph convolution layer, a temporal one-dimensional convolution layer, a fully connected layer, and a fusion layer. The specific structure is: The L1-1 layer is a multi-channel spatial graph convolution layer with an input dimension of N. c ×7×1,N d ×7×1,N s ×7×1, a multi-channel spatial graph convolution operation is performed on the input of each time step. The number of input features of the multi-channel spatial graph convolution layer is 1, the number of output features is 64, and the output dimension is N c ×7×64、N d ×7×64、N s ×7×64, the activation function is ReLU; The L1-2 layer, connected to the L1-1 layer, is a temporal one-dimensional convolutional layer with an input dimension of N. c ×7×64、N d ×7×64、N s ×7×64, perform a one-dimensional convolution operation in the time domain on the input of each sensor, the convolution kernel size is 3, the number is 64, and the output dimension is N c ×5×64,N d ×5×64,N s ×5×64, the activation function is ReLU; The L1-3 layer, connected to the L1-2 layer, is a temporal one-dimensional convolution layer with an input dimension of N. c ×5×64,N d ×5×64,N s ×5×64, perform a one-dimensional convolution operation in the time domain on the input of each sensor, the convolution kernel size is 3, the number is 64, and the output dimension is N c ×3×64,N d ×3×64,N s ×3×64, the activation function is ReLU; The L1-4 layer, connected to the L1-3 layer, is a temporal one-dimensional convolution layer with an input dimension of N. c ×3×64,N d ×3×64,N s ×3×64, perform a one-dimensional convolution operation in the time domain on the input of each sensor, the convolution kernel size is 3, the number is 64, and the output dimension is N c ×1×64,N d ×1×64,N s ×1×64, the activation function is ReLU; The L2-1 layer, connected to the L1-2 layer, is a multi-channel spatial graph convolution layer with an input dimension of N. c ×5×64,N d ×5×64,N s ×5×64, a multi-channel spatial graph convolution operation is performed on the input of each time step. The number of input features and output features of the multi-channel spatial graph convolution layer is 64, and the output dimension is N c ×5×64,N d ×5×64,N s ×5×64, the activation function is ReLU; The L2-2 layer, connected to the L2-1 layer, is a temporal one-dimensional convolutional layer with an input dimension of N. c ×5×64,N d ×5×64,N s ×5×64, perform a one-dimensional convolution operation in the time domain on the input of each sensor, the convolution kernel size is 3, the number is 64, and the output dimension is N c ×3×64,N d ×3×64,N s ×3×64, the activation function is ReLU; The L2-3 layer, connected to the L2-2 layer, is a temporal one-dimensional convolution layer with an input dimension of N. c ×3×64,N d ×3×64,N s ×3×64, perform a one-dimensional convolution operation in the time domain on the input of each sensor, the convolution kernel size is 3, the number is 64, and the output dimension is N c ×1×64,N d ×1×64,N s ×1×64, the activation function is ReLU; The L3-1 layer, connected to the L2-2 layer, is a multi-channel spatial graph convolution layer with an input dimension of N. c ×3×64,N d ×3×64,N s ×3×64, a multi-channel spatial graph convolution operation is performed on the input of each time step. The number of input features and output features of the multi-channel spatial graph convolution layer is 64, and the output dimension is N c ×3×64,N d ×3×64,N s ×3×64, the activation function is ReLU; The L3-2 layer, connected to the L3-1 layer, is a temporal one-dimensional convolution layer with an input dimension of N. c ×3×64,N d ×3×64,N s ×3×64, perform a one-dimensional convolution operation in the time domain on the input of each sensor, the convolution kernel size is 3, the number is 64, and the output dimension is N c ×1×64,N d ×1×64,N s ×1×64, the activation function is ReLU; The L4 layer is connected to the L1-4 layer, the L2-3 layer, and the L3-2 layer. It is a fully connected layer with an input dimension of N. c ×1×64,N d ×1×64,N s ×1×64, performs linear transformation operations, and the output dimension is N c 、N d 、N s , no activation function; The L5 layer is connected to the L4 layer and is a fusion layer. The input dimension is N c 、N d 、N s , perform fusion operation, the output dimension is N c 、N d 、N s , no activation function; Step 5: Design a multi-channel spatiotemporal graph convolutional network loss function; Step 6: Train the multi-channel spatiotemporal graph convolutional network model established in step 4. After the training is completed, output each first-order mapping matrix to obtain the spatial directed correlation representation of multiple types of monitoring data.

2. According to claim 1, the method for characterizing spatial directed associations of multi-type monitoring data based on a multi-channel spatiotemporal graph convolutional network is characterized in that In the step 1, data preprocessing to remove random effects is achieved by performing a sliding average operation.

3. According to claim 1, the method for characterizing spatial directed associations of multi-type monitoring data based on a multi-channel spatiotemporal graph convolutional network is characterized in that In the step 2, the standardization of each sensor trend data is achieved by subtracting the mean of the sensor trend data from the trend data and dividing the result by the standard deviation of the sensor trend data.

4. The method for characterizing spatial directed associations of multi-type monitoring data based on a multi-channel spatiotemporal graph convolutional network according to claim 1 is characterized in that In step 4, generalizing the "adjacency matrix" to the "mapping matrix" is achieved by not restricting the shape of the matrix. The "adjacency matrix" is defined as a square matrix whose size is equal to the number of graph vertices. The sizes of the rows and columns of the "mapping matrix" are respectively the numbers of the two types of sensors, which are used to characterize the correlation between any pair of sensors.

5. The method for characterizing spatial directed associations of multi-type monitoring data based on a multi-channel spatiotemporal graph convolutional network according to claim 1 is characterized in that The formula for the multi-channel spatial graph convolution operation is: In the formula, are the cable forces, main beam deflections, and tower tilt inputs at each time step in the l-th multi-channel spatial graph convolution layer, and their dimensions are N c ×F (l) 、N d ×F (l) 、N s ×F (l) , N c 、N d 、N s are the number of cable force sensors, deflection sensors and tilt sensors considered, respectively, and F (l) is the number of input features of the l-th multi-channel spatial graph convolutional layer, F (1) =1, F (2) =64, F (3) =64; It is the lth-order mapping matrix of the directed correlation between the cable forces and the cable forces in space, with a dimension of N c ×N c ; It is the lth-order mapping matrix for learning the directed spatial correlation between the main beam deflection and the cable force, with a dimension of N c ×N d ; All mapping matrices are row normalized; is the weight matrix to be trained for the l-th multi-channel spatial graph convolution layer, with a dimension of F (l) ×C (l) , C (l) is the number of output features of the l-th multi-channel spatial graph convolutional layer, C (l) ≡64; is the cable force, deflection, and tilt output of the l-th multi-channel spatial graph convolution layer at each time step, and their dimensions are N c ×C (l) 、N d ×C (l) 、N s ×C (l) ;|| is the splicing operation; It is a 1×1 convolution kernel that reduces the dimension of multi-channel information after three graph convolution operations, and the number of channels is 3.

6. The method for characterizing spatial directed associations of multi-type monitoring data based on a multi-channel spatiotemporal graph convolutional network according to claim 1 is characterized in that The formula for the fusion operation is: In the formula, is the model output, is the regression value of the cable force of the l-th order neighboring vertex; k l is a learnable parameter representing the contribution of the l-th order neighborhood vertex to the target vertex regression.

7. The method for characterizing spatial directed associations of multi-type monitoring data based on a multi-channel spatiotemporal graph convolutional network according to claim 1 is characterized in that In step 5, the loss function of the multi-channel spatiotemporal graph convolutional network is: Where, MSE(·) is the mean square error; P H To appropriately increase the penalty term for the contribution of heterogeneous monitoring data; P I To avoid the penalty term of each vertex information acting on itself; P O A penalty term to ensure that the first-order neighborhood vertices dominate.

8. The method for characterizing spatial directed associations of multi-type monitoring data based on a multi-channel spatiotemporal graph convolutional network according to claim 7 is characterized in that The P H , P I , P O Specifically: In the formula, λ h,l,dc is the penalty term for increasing the contribution of the main beam deflection to the cable force regression in the l-th multi-channel spatial graph convolution layer; is the convolution kernel parameter from the cable force to the cable force in the l-th multi-channel spatial graph convolution layer; In the formula, λ I,cc To avoid the penalty term where the measuring point is along the cable force-the cable force regresses to itself; is the first-order mapping matrix that reveals the directed correlation between the cable forces and the cable forces; tr(·) is the trace of the matrix; In the formula, λ o,l A penalty term to ensure that the first-order neighbor vertices dominate over the l-th-order neighbor vertices.

Images