Abnormal bridge monitoring data recognition method based on multi-channel spatiotemporal graph convolutional network

Abstract

The present invention discloses a method for identifying abnormal bridge monitoring data based on a multi-channel spatiotemporal graph convolutional network, the method comprising: defining a mapping matrix, collecting and standardizing multiple types of monitoring data of cable-stayed cable forces, main beam deflections, and cable tower inclinations of a cable-stayed bridge structure health monitoring system, designing a multi-channel spatiotemporal graph convolutional network with several learnable mapping matrices to perform spatiotemporal correlation modeling on multiple types of monitoring data of cable-stayed bridges, using early multiple types of monitoring data for training, obtaining adjacency relationships and calculating diagnostic indicators, and combining diagnostic indicators with adjacency relationships to identify abnormal bridge monitoring data. This method solves the shortcomings of low utilization of monitoring data in data-driven state assessment and confusion between sensor failures and structural variations, realizes the effective utilization of massive monitoring data, and is suitable for the operation of bridge structure health monitoring systems with multiple types of monitoring data.

Description

Technical Field

[0001] The present invention belongs to the field of bridge structure health monitoring, and relates to a method for identifying abnormal bridge monitoring data, and specifically to a method for identifying abnormal bridge monitoring data based on a multi-channel spatiotemporal graph convolutional network. Background Art

[0002] The bridge structure health monitoring system monitors the multi-point and multi-category loads and responses of the bridge in real time, providing data support for the assessment of the bridge structure status. Years of continuous and all-round monitoring have produced dense monitoring data sets. For example, the structural health monitoring system equipped with the Sutong Bridge in my country has a total of 785 sensors of various types, and the amount of data generated each year is about 2.5TB. These data are generated by each sensor at a certain sampling frequency and have both time and space dimension attributes. Mining valuable information from spatiotemporal big data and discovering the correlation patterns between data will help to give full play to the role of the bridge structure health monitoring system and provide scientific guidance for bridge structure operations.

[0003] However, the current data-driven bridge structure status assessment work usually starts with the analysis of a single monitoring item (such as cable force, deflection, etc.), resulting in most of the data being wasted invisibly. At the same time, the sensor array of the bridge structure health monitoring system usually works under outdoor conditions, which will lead to a greater risk of sensor failure over time, further increasing the proportion of abnormal data in the later stage of monitoring. Data-driven status assessment work must distinguish between sensor failure and structural variation, otherwise it will lead to misjudgment. In addition, the transmission lines of a single type of sensor usually face the same environment, so misjudgment of structural variation may occur. Therefore, by fusing multiple types of data and proposing an abnormal data identification method based on the spatiotemporal association of multiple types of data, it can effectively avoid the confusion between sensor failure and structural variation while improving the utilization rate of monitoring data, which is of great significance to data-driven bridge structure status assessment. Summary of the invention

[0004] In order to solve the current problems of low utilization of bridge structure monitoring data and confusion between sensor failure and structural variation, the present invention provides a method for identifying abnormal bridge monitoring data based on a multi-channel spatiotemporal graph convolutional network. This method solves the shortcomings of low utilization of monitoring data and confusion between sensor failure and structural variation in data-driven state assessment, realizes the effective utilization of massive monitoring data, and is suitable for the operation of bridge structure 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 identifying abnormal bridge monitoring data based on a multi-channel spatiotemporal graph convolutional network comprises the following steps:

[0007] Step 1: Define the mapping matrix and use it as a learnable parameter, so that it has the ability to quantitatively reveal the directed correlation between multi-type monitoring data of bridges;

[0008] Step 2: Collect multiple types of monitoring data of cable force, main beam deflection, and cable tower inclination of a cable-stayed bridge structure health monitoring system, and standardize them considering the different scales of the sensor data;

[0009] Step 3: Construct a multi-channel spatiotemporal graph convolutional network with several learnable mapping matrices, including designing the network model and loss function to ensure that the sensors can hierarchically aggregate information from similar and heterogeneous sensors of different orders;

[0010] Step 4: Use early multi-type monitoring data for training, regard each first-order mapping matrix obtained through training as the adjacency relationship of each sensor under normal operation of the cable-stayed bridge, and calculate the regression residual threshold of each sensor based on the regression residual in the training set;

[0011] Step 5: Apply the trained network model to the subsequent multi-type monitoring data, define diagnostic indicators, and calculate the daily diagnostic indicators of each sensor. Sensors with diagnostic indicators that are not zero are considered to be abnormal mode sensors, and sensors with diagnostic indicators that are not zero for a long time are considered to be key attention sensors.

[0012] Step 6: traverse the mode abnormal sensors, establish a key sensor set for each mode abnormal sensor, locate the suspected faulty sensor, and perform data reconstruction or model retraining;

[0013] Step 7: Determine whether the data anomaly pattern is a local anomaly or a global anomaly based on the data reconstruction or model retraining results. The former corresponds to sensor failure, and the latter corresponds to structural variation.

[0014] Compared with the prior art, the present invention has the following advantages:

[0015] 1. The present invention provides a basis for state assessment by means of correlation characterization of multi-type monitoring data, effectively solving the shortcomings of low utilization of monitoring data and difficulty in distinguishing sensor failures from structural variations.

[0016] 2. The present invention is applicable to the operation of bridge structure health monitoring systems with multiple types of monitoring data, providing a scientific basis for the management and decision-making of serving bridges. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of the abnormal bridge monitoring data identification method based on the multi-channel spatiotemporal graph convolutional network of the present invention;

[0018] Figure 2This is a data normalization diagram for step 2 of the embodiment;

[0019] Figure 3 This is a diagram of the multi-channel spatiotemporal graph convolutional network structure of step three of the embodiment;

[0020] Figure 4 The sensor daily diagnostic index graphs in February 2007 and April 2008 in step 5 of the embodiment;

[0021] Figure 5 This is a diagram of the process of locating a suspected faulty sensor in step six of the embodiment;

[0022] Figure 6 The figure is a result of locating the suspected faulty sensor in February 2007 and April 2008 in step 6 of the embodiment;

[0023] Figure 7 This is a graph showing the results of retraining the model in step seven of the embodiment. DETAILED DESCRIPTION

[0024] 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.

[0025] The present invention provides a method for identifying abnormal bridge monitoring data based on a multi-channel spatiotemporal graph convolutional network, the method comprising: defining a mapping matrix, collecting and standardizing multi-type monitoring data of cable-stayed cable forces, main beam deflection, and cable tower inclination of a cable-stayed bridge structure health monitoring system, designing a multi-channel spatiotemporal graph convolutional network with several learnable mapping matrices to perform spatiotemporal correlation modeling on multi-type monitoring data of cable-stayed bridges, using early multi-type monitoring data for training, obtaining adjacency relationships and calculating diagnostic indicators, and combining diagnostic indicators with adjacency relationships to identify abnormal bridge monitoring data. Figure 1 As shown, the specific steps are as follows:

[0026] Step 1: Define the mapping matrix and use it as a learnable parameter so that it has the ability to quantitatively reveal the directed correlations between multi-type monitoring data of bridges.

[0027] In this step, the definition of the mapping matrix is ​​generalized from the adjacency matrix, which is achieved by not restricting the shape of the matrix. The adjacency matrix is ​​defined as a square matrix with a size equal to the total number of vertices in the graph structure, while the mapping matrix is ​​a matrix with row and column scales of the number of sensors of two types (they can be the same type), which can be used to characterize the directed association of sensor pairs in one or two types of monitoring data. For example, the definition To learn the mapping matrix of the directed correlation between the cable forces, define To learn the mapping matrix from the beam deflection to the tower inclination, To learn the mapping matrix of the directed correlation from the tower inclination to the main beam deflection, N c 、N d 、N s They represent the number of cable tension sensors, hydraulic transmitters and inclinometers respectively.

[0028] Step 2: Collect multiple types of monitoring data such as cable force, main beam deflection, and tower inclination of a cable-stayed bridge structure health monitoring system, and standardize them considering the different scales of the sensor data.

[0029] In this step, the z-score method is used for sensor data standardization. The specific formula is:

[0030]

[0031] In the formula, x ij is the value of the j-th data of sensor i before normalization, x i ' j is the normalized value of the jth data of sensor i, μ i is the mean value of all data before normalization of sensor i, σ i is the standard deviation of all data before normalization of sensor i.

[0032] Step 3: Construct a multi-channel spatiotemporal graph convolutional network with several learnable mapping matrices, including designing the network model and loss function to ensure that the sensors can hierarchically aggregate information from similar and heterogeneous sensors of various orders.

[0033] In this step, the network model includes a multi-channel graph convolution layer, a one-dimensional convolution layer, a fully connected layer and a fusion layer. The specific structure is as follows:

[0034] L1-1 (multi-channel graph convolution) layer: the input data scale is N c ×S×1,N d ×S×1,N s ×S×1, each dimension in the three-dimensional data represents vertex, time, and feature. This layer operates on vertex and feature dimensions. The number of input features is 1, the number of output features is 64, and the output data scale is N c ×S×64、N d ×S×64、N s ×S×64;

[0035] L1-2 (one-dimensional convolution) layer: connected to the L1-1 layer, this layer operates in the time and feature dimensions, with an input feature number of 64, a convolution kernel size of 3, a depth of 64, a number of 64, an output feature number of 64, and an output data scale of N c ×S(1) ×64, N d ×S (1) ×64, N s ×S (1) ×64;

[0036] L1-3 (one-dimensional convolution) layer: connected to L1-2 layer, this layer operates in the time and feature dimensions, the input feature number is 64, the convolution kernel size is 3, the depth is 64, the number is 64, the output feature number is 64, and the output data scale size is N c ×S (2) ×64, N d ×S (2) ×64, N s ×S (2) ×64;

[0037] L1-4 (one-dimensional convolution) layer: connected to L1-3 layer, this layer operates in the time and feature dimensions, the input feature number is 64, the convolution kernel size is 3, the depth is 64, the number is 64, the output feature number is 64, and the output data scale size is N c ×1×64,N d ×1×64,N s ×1×64;

[0038] L2-1 (multi-channel graph convolution) layer: connected to L1-2 layer, this layer operates on vertex and feature dimensions, with 64 input features, 64 output features, and output data scale size N c ×S (1) ×64, N d ×S (1) ×64, N s ×S (1) ×64;

[0039] L2-2 (one-dimensional convolution) layer: connected to the L2-1 layer, this layer operates in the time and feature dimensions, with an input feature number of 64, a convolution kernel size of 3, a depth of 64, a number of 64, an output feature number of 64, and an output data scale of N c ×S (2) ×64, N d ×S (2) ×64, N s ×S (2) ×64;

[0040] L2-3 (one-dimensional convolution) layer: connected to the L2-2 layer, this layer operates in the time and feature dimensions, with an input feature number of 64, a convolution kernel size of 3, a depth of 64, a number of 64, an output feature number of 64, and an output data scale of N c ×1×64,N d ×1×64,Ns ×1×64;

[0041] L3-1 (multi-channel graph convolution) layer: connected to L2-2 layer, this layer operates on vertex and feature dimensions, with 64 input features, 64 output features, and an output data scale of N c ×S (2) ×64, N d ×S (2) ×64, N s ×S (2) ×64;

[0042] L3-2 (one-dimensional convolution) layer: connected to the L3-1 layer, this layer operates in the time and feature dimensions, with an input feature number of 64, a convolution kernel size of 3, a depth of 64, a number of 64, an output feature number of 64, and an output data scale of N c ×1×64,N d ×1×64,N s ×1×64;

[0043] L4 (fully connected) layer: connected to L1-4, L2-3, and L3-2. This layer performs linear transformation operations on the feature dimension, and the output data scale is N. c 、N d 、N s ;

[0044] The L5 (fusion) layer is connected to the L4 layer. This layer performs fusion operations on the vertex dimension, and the output data scale is N. c 、N d 、N s ;

[0045] Among them, the activation function of L1~L3 layers is ReLU, and no activation function is set for L4~L5 layers.

[0046] In this step, the loss function formula is:

[0047]

[0048] Where MSE(·) is the mean square error, is the true value of the cable force data, is the true value of the main beam deflection data, is the true value of the tower tilt data, is the regression value of the cable force data, and the rest of the definitions are the same as above. H , P I , P O are three additional penalty items, among which P H To ensure that each sensor effectively utilizes the penalty term of both homogeneous and heterogeneous monitoring data, P ITo avoid the penalty term of the sensor participating in its own regression, P O A penalty term to ensure that the first-order neighborhood information accounts for the main components of the regression.

[0049] Step 4: Use early multi-type monitoring data for training, regard each first-order mapping matrix obtained through training as the adjacency relationship of each sensor under normal operation of the cable-stayed bridge, and calculate the regression residual threshold of each sensor based on the regression residual in the training set.

[0050] In this step, the regression residual threshold of each sensor is determined by the 3σ principle, and the specific formula is:

[0051]

[0052] Where, T ui is the upper threshold of the regression residual of sensor i, T li is the lower threshold of the regression residual of sensor i, is the mean of the daily regression residuals of the training data after normalization of sensor i, is the standard deviation of the daily mean regression residual of the training data after normalization for sensor i.

[0053] Step 5: Apply the trained network model to the subsequent multi-type monitoring data, define diagnostic indicators, and calculate the daily diagnostic indicators of each sensor. Sensors with non-zero diagnostic indicators are regarded as mode abnormal sensors, and sensors with non-zero diagnostic indicators for a long time are regarded as key focus sensors.

[0054] In this step, the diagnostic index is defined as the part of the sensor's daily average regression residual that exceeds the threshold. The specific formula is:

[0055]

[0056] In the formula, I id is the diagnostic index of sensor i on day d, is the daily average regression residual of sensor i on day d. If a sensor has a non-zero diagnostic index, it is considered as a mode abnormal sensor, and the mode abnormal sensor whose diagnostic index is non-zero for a long time is considered as a key concern sensor.

[0057] Step 6: Traverse the mode abnormal sensors, establish a key sensor set for each mode abnormal sensor, locate the suspected faulty sensors, and perform data reconstruction or model retraining.

[0058] In this step, the key sensor set meets the following two conditions: (1) All sensors in the key sensor set are key focus sensors; (2) All key focus sensors have directed edges pointing to the mode abnormal sensors, and the edge weights are greater than or equal to 0.1.

[0059] In this step, the location of the suspected faulty sensor is achieved by combining the adjacency relationship and diagnostic indicators of the observed pattern abnormality sensors. Specifically, there are two paths: (1) If the diagnostic indicator of the short-term abnormal pattern abnormality sensor is the opposite of the diagnostic indicator of the sensor, the key sensor is the suspected faulty sensor; (2) If the diagnostic indicator of the long-term abnormal pattern abnormality sensor (i.e., the key focus sensor) is the opposite of the diagnostic indicator of the sensor, the sensor is the suspected faulty sensor.

[0060] In this step, data reconstruction or model retraining is determined by whether there are directed edges between the suspected faulty sensors. If there are no directed edges, data reconstruction is performed, that is, the measured values ​​are replaced by regression values; if there are directed edges, retraining is performed, that is, the data of the suspected faulty sensors are removed and the multi-channel spatiotemporal graph convolutional network is retrained.

[0061] Step 7: Determine whether the data anomaly pattern is a local anomaly or a global anomaly based on the data reconstruction or model retraining results. The former corresponds to sensor failure, and the latter corresponds to structural variation.

[0062] In this step, the judgment of local anomalies and global anomalies is based on the results of data reconstruction or model retraining. There are three possible results: (1) All or most of the sensor diagnostic indicators are restored to zero; (2) There are still sensors with non-zero diagnostic indicators, but they are concentrated in a specific type of sensor; (3) There are still sensors with non-zero diagnostic indicators, and they are widely distributed in multiple types of sensors. Among them, (1) and (2) are local anomalies, and (3) is a global anomaly.

[0063] Example:

[0064] like Figure 1 As shown in FIG. 1 , the abnormal bridge monitoring data recognition method based on the multi-channel spatiotemporal graph convolutional network includes the following steps:

[0065] Step 1: Define the mapping matrix to learn the correlation between the cable forces Define the mapping matrix from the deflection of the main beam to the inclination of the tower Define the mapping matrix from the inclination of the tower to the deflection of the main beam N c 、N d 、N s They represent the number of cable force sensors, hydraulic transmitters and inclinometers, which are 42, 17 and 3 respectively, for a total of 9 mapping matrices.

[0066] Step 2: Collect multiple types of monitoring data of cable force, main beam deflection, and cable tower inclination of a cable-stayed bridge structure health monitoring system, and standardize the data considering the different scales of each sensor data. The data standardization diagram is shown in the figure below: Figure 2 shown.

[0067] Step 3: Use several learnable mapping matrices to construct a multi-channel spatiotemporal graph convolutional network, including designing the network structure and loss function, such as Figure 3 As shown in Figure 2, the multi-channel spatiotemporal graph convolutional network structure is:

[0068] L1-1 (multi-channel graph convolution) layer: The input data scale is 42×7×1, 17×7×1, and 3×7×1. Each dimension in the three-dimensional data represents vertex, time, and feature respectively. This layer operates on vertex and feature dimensions. The input feature number is 1, the output feature number is 64, and the output data scale is 42×7×64, 17×7×64, and 3×7×64.

[0069] L1-2 (one-dimensional convolution) layer: connected to the L1-1 layer, this layer operates in the time and feature dimensions, with an input feature number of 64, a convolution kernel size of 3, a depth of 64, a quantity of 64, an output feature number of 64, and an output data scale of 42×5×64, 17×5×64, and 3×5×64;

[0070] L1-3 (one-dimensional convolution) layer: connected to the L1-2 layer, this layer operates in the time and feature dimensions, with an input feature number of 64, a convolution kernel size of 3, a depth of 64, a quantity of 64, an output feature number of 64, and an output data scale of 42×3×64, 17×3×64, and 3×3×64;

[0071] L1-4 (one-dimensional convolution) layer: connected to L1-3 layer, this layer operates in the time and feature dimensions, with an input feature number of 64, a convolution kernel size of 3, a depth of 64, a quantity of 64, an output feature number of 64, and an output data scale of 42×1×64, 17×1×64, and 3×1×64;

[0072] L2-1 (multi-channel graph convolution) layer: connected to the L1-2 layer, this layer operates on the vertex and feature dimensions, with an input feature number of 64, an output feature number of 64, and an output data scale of 42×5×64, 17×5×64, and 3×5×64;

[0073] L2-2 (one-dimensional convolution) layer: connected to the L2-1 layer, this layer operates in the time and feature dimensions, with an input feature number of 64, a convolution kernel size of 3, a depth of 64, a quantity of 64, an output feature number of 64, and an output data scale of 42×3×64, 17×3×64, and 3×3×64;

[0074] L2-3 (one-dimensional convolution) layer: connected to the L2-2 layer, this layer operates on the time and feature dimensions, with an input feature number of 64, a convolution kernel size of 3, a depth of 64, a quantity of 64, an output feature number of 64, and an output data scale of 42×1×64, 17×1×64, and 3×1×64;

[0075] L3-1 (multi-channel graph convolution) layer: connected to the L2-2 layer, this layer operates on the vertex and feature dimensions, with an input feature number of 64, an output feature number of 64, and an output data scale of 42×3×64, 17×3×64, and 3×3×64;

[0076] L3-2 (one-dimensional convolution) layer: connected to the L3-1 layer, this layer operates in the time and feature dimensions, with an input feature number of 64, a convolution kernel size of 3, a depth of 64, a number of 64, an output feature number of 64, and an output data scale of N c ×1×64,N d ×1×64,N s ×1×64;

[0077] L4 (fully connected) layer: connected to L1-4, L2-3, and L3-2. This layer performs linear transformation operations on the feature dimension, and the output data scale is N. c 、N d 、N s ;

[0078] L5 (fusion) layer: connected to L4 layer, this layer performs fusion operation on the vertex dimension, and the output data scale is N c 、N d 、N s ;

[0079] Among them, the activation function of L1~L3 layers is ReLU, and no activation function is set for L4~L5 layers.

[0080] Step 4: Use the standardized early multi-type monitoring data (from May 2006 to January 2007) for training, with Epoch (training round) of 100, batch_size (batch size) of 64, and learning_rate (learning rate) of 1e-3. Use the first-order mapping matrices obtained from training as the adjacency relationship of each sensor under normal operation of the cable-stayed bridge, and calculate the regression residual threshold of each sensor.

[0081] Step 5: Apply the trained network model to the later multi-type monitoring data (from February 2007 to May 2009), define the diagnostic indicators, and calculate the daily diagnostic indicators of each sensor. Sensors with diagnostic indicators that are not zero are considered to be abnormal sensors, and sensors with diagnostic indicators that are not zero for a long time are considered to be key attention sensors. The diagnostic indicators for February 2007 and April 2008 are as follows: Figure 4 As shown in the figure, it can be observed that a considerable number of sensor diagnostic indicators are abnormal. Among them, sensors C4, C5, C7, C19, C20, C23, C25, C26, C33, C38, C41, S1, S2, S3, D3, D4, and D5 have long-term abnormal conditions and are considered to be key sensors.

[0082] Step 6: Traverse the abnormal mode sensors, and establish a key sensor set for each abnormal mode sensor to locate the suspected faulty sensor. The positioning method is as follows: Figure 5 The positioning results are shown in Figure 6 As shown, sensors C4, C5, C7, C19, C23, C26, C33, C38, C41, S1, S2, S3, and D3 are considered as suspected faulty sensors. Since there is a correlation between them, the above sensor information is removed and retraining is performed.

[0083] Step 7: Model retraining results are as follows Figure 7 As shown in the figure, it can be found that after eliminating the influence of the suspected faulty sensor, most of the sensor data in February 2007 and April 2008 returned to normal, indicating that the data abnormality pattern during this period was local abnormality, which was related to the sensor and did not represent structural damage or variation; since June 2008, many sensors still have abnormal diagnostic indicators. However, all these sensors are cable force sensors, and their abnormal conditions are usually not continuous. If the abnormal condition is global (structural damage or variation), the hydraulic transmitter should also show a large number of abnormal regression residuals. However, the diagnostic indicators of the hydraulic transmitter are almost all normal, which shows that the data abnormality pattern is still local abnormality at this time.

Claims

1. A method for identifying abnormal bridge monitoring data based on a multi-channel spatiotemporal graph convolutional network, characterized in that The method comprises the following steps: Step 1: Define the mapping matrix and use it as a learnable parameter, so that it has the ability to quantitatively reveal the directed correlation between multi-type monitoring data of bridges; Step 2: Collect multiple types of monitoring data of cable force, main beam deflection, and cable tower inclination of a cable-stayed bridge structure health monitoring system, and standardize them considering the different scales of the sensor data; Step 3: Construct a multi-channel spatiotemporal graph convolutional network with several learnable mapping matrices, including designing the network model and loss function to ensure that the sensors can hierarchically aggregate information from similar and heterogeneous sensors of different orders; Step 4: Use early multi-type monitoring data for training, regard each first-order mapping matrix obtained through training as the adjacency relationship of each sensor under normal operation of the cable-stayed bridge, and calculate the regression residual threshold of each sensor based on the regression residual in the training set; Step 5: Apply the trained network model to the subsequent multi-type monitoring data, define diagnostic indicators, and calculate the daily diagnostic indicators of each sensor. Sensors with diagnostic indicators that are not zero are considered to be mode abnormal sensors, and sensors with diagnostic indicators that are not zero for a long time are considered to be key attention sensors. Step 6: traverse the mode abnormal sensors, establish a key sensor set for each mode abnormal sensor, locate the suspected faulty sensor, and perform data reconstruction or model retraining; Step 7: Determine whether the data anomaly pattern is a local anomaly or a global anomaly based on the data reconstruction or model retraining results. The former corresponds to sensor failure, and the latter corresponds to structural variation.

2. The abnormal bridge monitoring data identification method based on multi-channel spatiotemporal graph convolutional network according to claim 1 is characterized in that In the step 1, define To learn the mapping matrix of the directed correlation between the cable forces, define To learn the mapping matrix from the beam deflection to the tower inclination, To learn the mapping matrix of the directed correlation from the tower inclination to the main beam deflection, N c 、N d 、N s They represent the number of cable tension sensors, hydraulic transmitters and inclinometers respectively.

3. The abnormal bridge monitoring data identification method based on multi-channel spatiotemporal graph convolutional network according to claim 1 is characterized in that In step 2, the z-score method is used for sensor data standardization, and the specific formula is: In the formula, x ij is the value of the j-th data of sensor i before normalization, x i ' j is the normalized value of the jth data of sensor i, μ i is the mean value of all data before normalization of sensor i, σ i is the standard deviation of all data before normalization of sensor i.

4. The abnormal bridge monitoring data identification method based on multi-channel spatiotemporal graph convolutional network according to claim 1 is characterized in that In step 3, the network model includes a multi-channel graph convolution layer, a one-dimensional convolution layer, a fully connected layer and a fusion layer, and the specific structure is: The L1-1 layer is a multi-channel graph convolution layer, and the input data scale is N c ×S×1,N d ×S×1,N s ×S×1, each dimension in the three-dimensional data represents vertex, time, and feature. This layer operates on vertex and feature dimensions. The number of input features is 1, the number of output features is 64, and the output data scale is N c ×S×64、N d ×S×64、N s ×S×64, where N c 、N d 、N s represent the number of cable force sensors, hydraulic transmitters and inclinometers respectively; The L1-2 layer is a one-dimensional convolution layer, which is connected to the L1-1 layer. This layer operates in the time and feature dimensions. The number of input features is 64, the convolution kernel size is 3, the depth is 64, the number is 64, the output feature number is 64, and the output data scale size is N c ×S (1) ×64, N d ×S (1) ×64, N s ×S (1) ×64; The L1-3 layer is a one-dimensional convolution layer, connected to the L1-2 layer. This layer operates in the time and feature dimensions. The number of input features is 64, the convolution kernel size is 3, the depth is 64, the number is 64, the output feature number is 64, and the output data scale size is N c ×S (2) ×64, N d ×S (2) ×64, N s ×S (2) ×64; The L1-4 layer is a one-dimensional convolution layer, connected to the L1-3 layer. This layer operates in the time and feature dimensions. The number of input features is 64, the convolution kernel size is 3, the depth is 64, the number is 64, the output feature number is 64, and the output data scale size is N c ×1×64,N d ×1×64,N s ×1×64; The L2-1 layer is a multi-channel graph convolution layer, which is connected to the L1-2 layer. This layer operates on the vertex and feature dimensions. The number of input features is 64, the number of output features is 64, and the output data scale is N. c ×S (1) ×64, N d ×S (1) ×64, N s ×S (1) ×64; The L2-2 layer is a one-dimensional convolution layer, which is connected to the L2-1 layer. This layer operates in the time and feature dimensions. The number of input features is 64, the convolution kernel size is 3, the depth is 64, the number is 64, the output feature number is 64, and the output data scale size is N. c ×S (2) ×64, N d ×S (2) ×64, N s ×S (2) ×64; The L2-3 layer is a one-dimensional convolution layer, which is connected to the L2-2 layer. This layer operates in the time and feature dimensions. The number of input features is 64, the convolution kernel size is 3, the depth is 64, the number is 64, the output feature number is 64, and the output data scale size is N c ×1×64,N d ×1×64,N s ×1×64; The L3-1 layer is a multi-channel graph convolution layer, which is connected to the L2-2 layer. This layer operates on the vertex and feature dimensions. The number of input features is 64, the number of output features is 64, and the output data scale is N. c ×S (2) ×64, N d ×S (2) ×64, N s ×S (2) ×64; The L3-2 layer is a one-dimensional convolution layer, which is connected to the L3-1 layer. This layer operates in the time and feature dimensions. The number of input features is 64, the convolution kernel size is 3, the depth is 64, the number is 64, the output feature number is 64, and the output data scale size is N c ×1×64,N d ×1×64,N s ×1×64; The L4 layer is a fully connected layer, connected to the L1-4 layer, L2-3 layer, and L3-2 layer. This layer performs linear transformation operations on the feature dimension, and the output data scale is N. c 、N d 、N s ; The L5 layer is the fusion layer, which is connected to the L4 layer. This layer performs fusion operations on the vertex dimension, and the output data scale is N c 、N d 、N s ; Among them, the activation function of L1~L3 layers is ReLU, and no activation function is set for L4~L5 layers; The loss function formula is: Where MSE(·) is the mean square error, is the true value of the cable force data, is the true value of the main beam deflection data, is the true value of the tower tilt data, is the regression value of the cable force data, N c 、N d 、N s Represents the number of cable force sensors, hydraulic transmitters and inclinometers, P H To ensure that each sensor effectively utilizes the penalty term of both homogeneous and heterogeneous monitoring data, P I To avoid the penalty term of the sensor participating in its own regression, P O A penalty term to ensure that the first-order neighborhood information accounts for the main components of the regression.

5. The abnormal bridge monitoring data identification method based on multi-channel spatiotemporal graph convolutional network according to claim 1 is characterized in that In step 4, the regression residual threshold of each sensor is determined by the 3σ principle, and the specific formula is: Where, T ui is the upper threshold of the regression residual of sensor i, T li is the lower threshold of the regression residual of sensor i, is the mean of the daily regression residuals of the training data after normalization of sensor i, is the standard deviation of the daily mean regression residual of the training data after normalization for sensor i.

6. The abnormal bridge monitoring data identification method based on multi-channel spatiotemporal graph convolutional network according to claim 1 is characterized in that In step 5, the diagnostic index is defined as the part of the sensor daily average regression residual that exceeds the threshold, and the specific formula is: In the formula, I id is the diagnostic index of sensor i on day d, is the daily average regression residual of sensor i on day d; if a sensor has a non-zero diagnostic index, it is regarded as a mode abnormal sensor, and the mode abnormal sensor whose diagnostic index is non-zero for a long time is regarded as a key concern sensor.

7. The abnormal bridge monitoring data identification method based on multi-channel spatiotemporal graph convolutional network according to claim 1 is characterized in that In step six, the key sensor set satisfies the following two conditions: (1) all sensors in the key sensor set are key focus sensors; and (2) all key focus sensors have directed edges pointing to the mode abnormal sensors, and the edge weights are all greater than or equal to 0.

1.

8. The abnormal bridge monitoring data identification method based on multi-channel spatiotemporal graph convolutional network according to claim 1 is characterized in that In step six, the location of the suspected faulty sensor is achieved by combining the adjacency relationship and diagnostic indicators of the observed pattern abnormality sensors. Specifically, there are two paths: (1) If the diagnostic indicator of one and only one sensor in the key sensor set of the short-term abnormal pattern abnormality sensor is opposite to the diagnostic indicator of the sensor, then the key sensor is the suspected faulty sensor; (2) If the diagnostic indicator of no sensor in the key sensor set of the long-term abnormal pattern abnormality sensor is opposite to the diagnostic indicator of the sensor, then the sensor is the suspected faulty sensor.

9. The abnormal bridge monitoring data identification method based on multi-channel spatiotemporal graph convolutional network according to claim 1 is characterized in that In step 6, data reconstruction or model retraining is determined based on whether there is a directed edge between the suspected faulty sensors. If there is no directed edge, data reconstruction is performed, i.e., the measured value is replaced by the regression value. If there are directed edges, retraining is performed, that is, the data of the suspected faulty sensors are removed and the multi-channel spatiotemporal graph convolutional network is retrained.

10. The abnormal bridge monitoring data identification method based on multi-channel spatiotemporal graph convolutional network according to claim 1 is characterized in that In step seven, the judgment of local anomalies and global anomalies is based on the results of data reconstruction or model retraining. Specifically, there are three possible results: (1) all or most of the sensor diagnostic indicators are restored to zero; (2) there are still sensors with non-zero diagnostic indicators, but they are concentrated in specific types of sensors; (3) there are still sensors with non-zero diagnostic indicators, and they are widely distributed in multiple types of sensors. Among them, (1) and (2) are local anomalies, and (3) is a global anomaly.

Images