Bridge support maintenance decision-making method and system combining the Internet of Things and big data

Abstract

The present application provides a bridge bearing maintenance decision method and system that combines the Internet of Things and big data. The method uses an initialized bearing state decision algorithm to perform feature mapping on local sensor network samples to obtain a local two-dimensional characterization array, and performs feature mapping on target bearing units to obtain a bearing one-dimensional characterization array; the local two-dimensional characterization array is aggregated to obtain a sample local one-dimensional characterization array; a domain deviation error value is obtained based on a consistency score between the sample local one-dimensional characterization array and the bearing one-dimensional characterization array in a feature domain, and a target deviation error value is determined based on the domain deviation error value; a feature consistency error value is determined based on a commonality measurement result between the sample local one-dimensional characterization array and the bearing one-dimensional characterization array, a comprehensive error value is generated based on the target deviation error value and the feature consistency error value, and the initialized bearing state decision algorithm is repeatedly debugged using the comprehensive error value to increase the algorithm's feature learning capability and bearing state decision-making capability.

Description

Technical Field

[0001] The present disclosure relates to the field of data processing, and more specifically, to a bridge bearing maintenance decision-making method and system combining the Internet of Things and big data. Background Art

[0002] In modern infrastructure construction, bridges are key components of transportation hubs, and their structural safety and long-term stability are of vital importance. As an important component connecting the superstructure and substructure of the bridge, bridge bearings bear the load of the bridge and ensure the normal operation of the bridge under various environmental conditions. However, due to the long-term influence of various factors such as complex loads, environmental erosion, and material aging, bridge bearings may suffer from various diseases and damages, such as bearing deformation, rubber aging, and steel plate rust. If these problems are not maintained and handled in a timely and effective manner, they will seriously affect the service life and safety of the bridge, and may even cause damage and collapse of the bridge structure.

[0003] With the rapid development of Internet of Things and Big Data technologies, although some sensors have been applied in the field of bridge monitoring to obtain real-time data of bridge structures, there are relatively few studies on bridge bearing maintenance decisions. Most of the existing analysis methods based on sensor data remain at simple data statistics and trend analysis, lacking in-depth evaluation of bearing status and accurate decision-making. For example, the state of the bearing is judged only by monitoring the changes in single parameters such as bearing displacement and pressure, without fully considering the relationship between the bearing and the surrounding environment and other bearings, making it difficult to accurately identify the true state and potential risks of the bearing. Summary of the invention

[0004] In view of this, the present application provides a bridge bearing maintenance decision-making method and system that combines the Internet of Things and big data.

[0005] According to one aspect of an embodiment of the present disclosure, a bridge bearing maintenance decision method combining the Internet of Things and big data is provided, the method comprising: obtaining a bearing unit to be analyzed and a target local sensor network corresponding to the bearing unit to be analyzed in a target bridge bearing Internet of Things sensor network; using a bearing state decision algorithm after debugging and convergence, performing feature mapping on the target local sensor network to obtain a target local two-dimensional representation array, and performing feature mapping on the bearing unit to be analyzed to obtain a target bearing one-dimensional representation array; aggregating the target local two-dimensional representation array to obtain a target local one-dimensional representation array, and determining a neighborhood deviation coefficient based on a consistency score between the target local one-dimensional representation array and the target bearing one-dimensional representation array in a feature domain; and determining whether the bearing unit to be analyzed is a bearing unit to be maintained based on the neighborhood deviation coefficient. Among them, the debugging steps of the support state decision algorithm include: obtaining a target support unit and a local sensor network sample corresponding to the target support unit in the bridge support IoT sensor network sample; using the initialized support state decision algorithm, feature mapping the local sensor network sample to obtain a local two-dimensional characterization array, and feature mapping the target support unit to obtain a support one-dimensional characterization array; aggregating the local two-dimensional characterization array to obtain a sample local one-dimensional characterization array; obtaining a domain deviation error value based on the consistency score between the sample local one-dimensional characterization array and the support one-dimensional characterization array in the feature domain, and determining a target deviation error value based on the domain deviation error value; determining a feature consistency error value based on the commonality measurement result between the sample local one-dimensional characterization array and the support one-dimensional characterization array; generating a comprehensive error value based on the target deviation error value and the feature consistency error value, and using the comprehensive error value to repeatedly debug the initialized support state decision algorithm until the preset debugging end condition is met to obtain the support state decision algorithm.

[0006] According to another aspect of an embodiment of the present disclosure, a computer system is provided, including: a processor; and a memory, wherein the memory stores a computer-readable code, and when the computer-readable code is executed by the processor, the processor executes the method described above.

[0007] The bridge bearing maintenance decision method combining the Internet of Things and big data provided in the present application obtains the target bearing unit and the local sensor network sample corresponding to the target bearing unit in the bridge bearing Internet of Things sensor network sample in the algorithm training stage, and generates sample data for metric learning. Then, feature mapping is performed on the local sensor network sample and the target bearing unit respectively to obtain a local two-dimensional representation array and a bearing one-dimensional representation array. The local two-dimensional representation array is integrated to obtain a sample local one-dimensional representation array, and the sample local one-dimensional representation array can be used as the environmental feature data of the target bearing unit. The consistency score of the sample local one-dimensional representation array and the support one-dimensional representation array in the feature domain is used to obtain the domain deviation error value, and the domain deviation error value can evaluate the consistency between the target bearing unit and the local sensor network sample. Furthermore, the present application also determines the feature consistency error value based on the commonality measurement results of the sample local one-dimensional representation array and the support one-dimensional representation array at the feature level, and the feature consistency error value can balance the feature domain during feature mapping. The initialization support state decision algorithm is repeatedly debugged without labels using a comprehensive error value jointly generated by the domain deviation error value and the feature consistency error value, thereby increasing the state estimation performance of the algorithm for the mapping feature. Furthermore, the present application adopts a label-free debugging method to avoid the use of a large number of labels of support units to be maintained, which can overcome the defect of poor algorithm training quality caused by a small number of support units to be maintained. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 It is a schematic diagram of the composition of an application environment provided by this application;

[0009] Figure 2 It is a flowchart of a bridge bearing maintenance decision-making method combining the Internet of Things and big data provided by this application;

[0010] Figure 3 It is a structural diagram of a computer system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0011] In order to facilitate a clearer understanding of the present application, the application environment of the bridge support maintenance decision method combining the Internet of Things and big data for implementing the present application is first introduced. Figure 1 As shown, the application environment includes a computer system 10 and a sensor cluster. The sensor cluster may include one or more sensors, and the number of sensors is not limited. Figure 1 As shown, the sensor cluster may specifically include sensor 1, sensor 2, ..., sensor n; it can be understood that sensor 1, sensor 2, sensor 3, ..., sensor n can all be connected to the computer system 10 through a network, so that each sensor can exchange data with the computer system 10 through the network connection.

[0012] It is understandable that the computer system 10 may refer to a device that executes the bridge bearing maintenance decision method combining the Internet of Things and big data in this application, which may be an independent physical server, or a server cluster or distributed system composed of at least two physical servers. The sensor may be a pressure sensor, a displacement sensor, a temperature sensor, etc., but is not limited thereto. Each sensor and the computer system may be directly or indirectly connected via wired or wireless communication, and this application does not limit this.

[0013] For further information, see Figure 2 , is a flow chart of a bridge support maintenance decision-making method combining the Internet of Things and big data provided by an embodiment of the present application. Figure 2 As shown, this method can be Figure 1 The computer system 10 in the embodiment is used for execution, wherein the bridge support maintenance decision-making method combining the Internet of Things and big data may include the following steps:

[0014] Step S100: Acquire a support unit to be analyzed and a target local sensor network corresponding to the support unit to be analyzed in the target bridge support IoT sensor network.

[0015] The target bridge bearing IoT sensor network is a graph structure that contains numerous vertices and edges connecting these vertices. In this network, the vertices represent the various bridge bearing units, while the edges represent certain associations between these bearing units, such as physical adjacency, data transmission, etc. For example, in a large bridge, bearing units at different locations are connected together through sensors to form a network. These sensors can collect various state data of the bearing units in real time, such as displacement, pressure, temperature, etc., and transmit these data to the computer system for analysis and processing. This network can be represented by a graph structure in graph theory, in which the attributes of the vertices can be various characteristic data of the bearing units, and the weights of the edges represent the strength of association between the bearing units.

[0016] The computer system obtains the support units to be analyzed in the IoT sensor network of the target bridge support, where the support units to be analyzed are vertices in the graph structure. The computer system can determine the support units to be analyzed in a variety of ways. One way is to filter according to pre-set rules or conditions. For example, the support units to be analyzed can be selected based on factors such as the location information of the support units, historical maintenance records, and real-time monitoring data. Assuming that in a bridge, some support units are located in key positions, such as the main span of the bridge, or according to historical data, some support units have a high probability of failure, then these support units can be used as support units to be analyzed.

[0017] The target local sensor network corresponding to the support unit to be analyzed refers to the sensor network within a certain range around the support unit to be analyzed. This range can be determined according to specific application requirements and actual conditions. The process of obtaining the target local sensor network can be implemented in the following ways:

[0018] Graph-based traversal algorithm: Graph traversal algorithms such as depth-first search (DFS) or breadth-first search (BFS) can be used to determine the target local sensor network. Taking breadth-first search as an example, starting from the support unit to be analyzed, its adjacent support units are visited in a certain hierarchical order until the preset range is reached or a specific stop condition is met. For example, the preset range can be a subnetwork composed of all support units that are no more than 3 vertices away from the support unit to be analyzed. The specific algorithm implementation can be assisted by the queue data structure, and its basic steps are as follows:

[0019] Put the support unit to be analyzed into the queue; when the queue is not empty, take out the head element, visit all the adjacent vertices of the element that have not been visited, and put these adjacent vertices into the queue; repeat the above steps until the queue is empty or the preset stop condition is reached.

[0020] Alternatively, the method based on distance metric: the target local sensor network can be determined based on the distance metric between the support units. For example, a distance function d(u,v) can be defined to calculate the distance between two support units u and v, where d(u,v) can be a physical distance, data similarity, etc. Then, all support units whose distance to the support unit to be analyzed is less than a certain threshold T are selected to form the target local sensor network. The specific calculation can use different formulas according to different application scenarios. For example, if physical distance is used as the metric, assuming that the coordinates of support unit u are , the coordinates of the support element v are , then their physical distance can be calculated by the following formula:

[0021] ;

[0022] If data similarity is used as a metric, various similarity metrics can be used, such as cosine similarity, Euclidean distance, etc. Taking cosine similarity as an example, assuming that the feature vectors of support units u and v are , then the cosine similarity between them can be calculated by the following formula: For example, in order to better understand the process of step S100, a simple example is given below. Assume that there is a bridge whose support IoT sensor network consists of 10 support units. The connection relationship between these support units can be represented by a graph structure as shown below:

[0023] Vertex (support element): A, B, C, D, E, F, G, H, I, J;

[0024] Edges (connection relationships): (A,B), (A,C), (B,D), (B,E), (C,F), (C,G), (D,H), (E,I), (F,J);

[0025] Assume that the support unit to be analyzed is A, and obtain the corresponding target local sensor network. If the breadth-first search algorithm is used, and the preset range is the subnetwork composed of all support units whose distance from the support unit to be analyzed is no more than 2 vertices, the operation process of the computer system is as follows:

[0026] First, the support element A to be analyzed is placed in the queue;

[0027] Take out the first element A, visit its adjacent vertices B and C, and put them into the queue;

[0028] Take out the first element B, visit its adjacent vertices D and E, and put them into the queue;

[0029] Take out the first element C, visit its adjacent vertices F and G, and put them into the queue;

[0030] At this time, the elements in the queue are D, E, F, and G. Since the distance between these vertices and the support unit A to be analyzed does not exceed 2 vertices, they constitute the target local sensing network, that is, the vertices contained in the target local sensing network are A, B, C, D, E, F, and G.

[0031] Step S200: using the support state decision algorithm after debugging and convergence, feature mapping is performed on the target local sensor network to obtain a target local two-dimensional representation array, and feature mapping is performed on the support unit to be analyzed to obtain a target support one-dimensional representation array.

[0032] The bearing state decision algorithm is a debugged and converged algorithm, which aims to extract meaningful feature information from the input data for subsequent analysis and decision-making of the bridge bearing state. The algorithm maps the characteristics of the target local sensor network and the bearing unit to be analyzed, converting the original data into a more representative and analyzable feature array, providing a basis for subsequent analysis.

[0033] When the computer system performs feature mapping on the target local sensor network, it aims to convert various information and features in the network into a two-dimensional representation array, which can reflect the contextual characteristics of the target local sensor network. The specific process is as follows:

[0034] Before feature mapping, the computer system preprocesses the data of the target local sensor network. This may include data cleaning, normalization and other operations to ensure the quality and consistency of the data. For example, for various data collected by sensors, such as displacement, pressure, temperature, etc., there may be noise, outliers and other problems. The computer system can clean the data by filtering, removing outliers and other methods. At the same time, in order to eliminate the dimensional differences between different data, the computer system can normalize the data and map the data to a specific interval, such as [0,1] or [-1,1].

[0035] Various features are extracted from the target local sensor network. These features may include physical features of the support unit (such as position, size, etc.), data features collected by the sensor (such as mean, variance, spectral features, etc.), and network structure features (such as vertex degree, connection relationship, etc.). For example, for a target local sensor network composed of multiple support units, the computer system can calculate the degree of each support unit, that is, the number of edges connected to the support unit, as a feature. At the same time, the computer system can also calculate the mean and variance of the data collected by the sensor as data features.

[0036] After extracting various features of the target local sensor network, the computer system uses the support state decision algorithm after debugging and convergence to map these features into a two-dimensional representation array. This process can be implemented in a variety of ways, such as using neural networks, deep learning models, etc. Taking neural networks as an example, the computer system can use the extracted features as the input of the neural network, and finally obtain a two-dimensional representation array as the output through calculation and transformation of multiple layers of neurons. Assuming that the target local sensor network contains n support units, each support unit has m features, then the computer system can form these features into an n×m matrix as the input of the neural network. After calculation by the neural network, a p×q two-dimensional representation array is output, where p and q are determined according to specific application requirements and algorithm design.

[0037] Assume that there is a target local sensor network, which contains three support units A, B, and C. Each support unit has two features, namely displacement d and pressure p. The computer system first preprocesses the collected data, removes noise and outliers, and normalizes the data to the [0,1] interval. Then, the computer system extracts the features of each support unit and obtains the following feature matrix: ; Assume that the support state decision algorithm after debugging convergence uses a simple neural network, which contains a hidden layer with 4 neurons. The computer system uses the feature matrix as the input of the neural network, and after calculation and transformation of the hidden layer, it finally obtains a two-dimensional representation array: ; Among them, fij Represents the elements in the i-th row and j-th column of the two-dimensional representation array. These elements are obtained through the calculation of the neural network and reflect the contextual characteristics of the target local sensor network.

[0038] When the computer system performs feature mapping on the support unit to be analyzed, it aims to convert various features of the support unit to be analyzed into a one-dimensional representation array. The specific implementation method and process are similar to feature mapping of the target local sensor network, including data preprocessing, feature extraction and feature mapping.

[0039] The computer system preprocesses the data of the support unit to be analyzed, such as cleaning and normalization, and then extracts various features of the support unit to be analyzed, which may include the physical features of the support unit, the data features collected by the sensor, etc. For example, for the support unit A to be analyzed, the computer system can extract its data features such as displacement, pressure, temperature, and its physical features such as position and size. The extracted features are mapped to a one-dimensional representation array using the support state decision algorithm after debugging and convergence. This process can also be implemented through neural networks, deep learning models, etc. Assuming that the support unit to be analyzed has k features, these features can be combined into a 1×k vector as the input of the neural network. After calculation by the neural network, a one-dimensional representation array of length l is output, where l is determined according to specific application requirements and algorithm design.

[0040] Assume that the support element A to be analyzed has three characteristics, namely displacement ,pressure and temperature After the computer system preprocesses these data, the feature vector is obtained: ; Assume that the support state decision algorithm after debugging convergence uses a simple neural network, which contains a hidden layer with 3 neurons. The computer system uses the feature vector as the input of the neural network, and after calculation and transformation of the hidden layer, it finally obtains a one-dimensional representation array: ; Among them, g i Represents the i-th element in the one-dimensional characterization array, which is obtained through the calculation of the neural network and reflects the characteristics of the support unit to be analyzed.

[0041] Step S300: Aggregate the target local two-dimensional representation array to obtain the target local one-dimensional representation array, and determine the neighborhood deviation coefficient according to the consistency score between the target local one-dimensional representation array and the target support one-dimensional representation array in the feature domain.

[0042] The purpose of the computer system aggregating the target local two-dimensional representation array is to integrate and refine the information in the two-dimensional array and convert it into a one-dimensional representation array for subsequent comparison and analysis with the target support one-dimensional representation array. The aggregation operation can be implemented in many ways, and the following are some of the ways and their examples:

[0043] Mean aggregation: Calculate the mean of each column element in the target local two-dimensional representation array, and combine these means into a one-dimensional array as the target local one-dimensional representation array. Assume that the target local two-dimensional representation array is A, and its dimension is m×n, where m represents the number of rows and n represents the number of columns. Then the calculation formula for mean aggregation is: ; Among them, a i represents the i-th element in the target local one-dimensional representation array, Represents the element in the j-th row and i-th column of the target local two-dimensional representation array A.

[0044] For example, suppose there is a target local two-dimensional representation array A as follows: ; Using the mean aggregation operation, the computer system calculates the mean of each column:

[0045] .

[0046] The target local one-dimensional representation array is obtained as .

[0047] Max pooling aggregation: The computer system selects the maximum value of each column in the target local two-dimensional representation array, and forms these maximum values ​​into a one-dimensional array as the target local one-dimensional representation array. Assume that the target local two-dimensional representation array is A, and its dimension is m×n. Then the calculation formula of the maximum pooling aggregation is: ; Among them, a i Represents the i-th element in the target local one-dimensional representation array, A ji Represents the element in the j-th row and i-th column of the target local two-dimensional representation array A.

[0048] For example, for the above target local two-dimensional representation array A, through the maximum pooling aggregation operation, the computer system selects the maximum value of each column; the target local one-dimensional representation array is obtained as .

[0049] After obtaining the target local one-dimensional representation array, the computer system needs to calculate the consistency score between it and the target support one-dimensional representation array in the feature domain. The consistency score can measure the similarity of the two arrays in the feature space, such as cosine similarity, Euclidean distance, etc.

[0050] The computer system determines the neighborhood deviation coefficient based on the calculated consistency score. The neighborhood deviation coefficient is used to describe the degree of inconsistency between the support unit to be analyzed and the surrounding support units. The specific determination method can be designed according to the application requirements and actual conditions. One way is to normalize the consistency score so that its value range is between [0, 1], and then subtract the normalized consistency score from 1 to obtain the neighborhood deviation coefficient. For example, assuming that the consistency score calculated by cosine similarity is 0.8, it is still 0.8 after normalization to the interval [0, 1]. The neighborhood deviation coefficient is: 1-0.8=0.2. It means that the degree of inconsistency between the support unit to be analyzed and the surrounding support units is 0.2.

[0051] Step S400: Determine whether the support unit to be analyzed is a support unit to be maintained according to the neighborhood deviation coefficient.

[0052] The neighborhood deviation coefficient is an important indicator that describes the degree of inconsistency between the support unit to be analyzed and the surrounding support units. It reflects the uniqueness or abnormality of the support unit in the local sensor network. In actual bridge structures, the state of a support unit usually has a certain correlation with the support units around it. If the neighborhood deviation coefficient of a support unit is large, it means that its state is significantly different from that of the surrounding support units, which may indicate that there are potential problems with the support unit and further attention and evaluation is required. When the computer system determines whether the support unit to be analyzed is a support unit to be maintained, for example, a threshold is set and the neighborhood deviation coefficient is compared with the threshold. The determination of this threshold requires comprehensive consideration of multiple factors, such as the type of bridge, service life, historical maintenance records, and actual operating environment. For example, the following method is used:

[0053] Fixed threshold method: The computer system sets a fixed threshold T in advance based on the relevant information and empirical data of the bridge. When the neighborhood deviation coefficient D is greater than the threshold T, that is, D>T, the computer system determines that the support unit to be analyzed is a support unit to be maintained; conversely, when D ≤T, it is determined that the support unit does not need maintenance temporarily.

[0054] For example, for a specific type of bridge, based on historical maintenance data and expert experience, the threshold value of the neighborhood deviation coefficient is set to T = 0.6. Assume that the computer system calculates the neighborhood deviation coefficient D = 0.8 of a proposed support unit. Since 0.8>0.6, the computer system determines that the proposed support unit is a support unit to be maintained, and maintenance work needs to be arranged in time.

[0055] Dynamic threshold method: Considering that the state of bridge bearings may change with time, environment and other factors, the dynamic threshold method can be used to determine the bearing units to be maintained. This method dynamically adjusts the threshold based on the analysis results of real-time monitoring data and historical data. For example, the statistical distribution of the neighborhood deviation coefficients of all bearing units over a period of time, such as the mean and standard deviation , to determine the dynamic threshold T. A dynamic threshold calculation formula is: ; Wherein, k is a coefficient determined according to actual needs and experience, and is usually between 1 and 3. When the neighborhood deviation coefficient D is greater than the dynamic threshold T, the support unit to be analyzed is determined to be a support unit to be maintained.

[0056] In practical applications, in order to improve the accuracy and reliability of judgment, the computer system can also consider other factors combined with the neighborhood deviation coefficient for comprehensive judgment. For example, in addition to the neighborhood deviation coefficient, factors such as the service life of the support unit, historical maintenance records, and real-time monitored physical parameters (such as displacement, pressure, temperature, etc.) can also be considered. By establishing a comprehensive evaluation model, these factors are weighted and integrated to obtain a comprehensive evaluation index, and then based on this index, it is determined whether the support unit to be analyzed is a support unit to be maintained.

[0057] Assume that the calculation formula of the comprehensive evaluation index I is: ;

[0058] Where D is the neighborhood deviation coefficient, Y is the service life of the support unit, H is the historical maintenance record (which can be quantified by the number of maintenance times, etc.), and P is the comprehensive score of the physical parameters monitored in real time. are the weights of each factor, and satisfy The computer system sets a threshold for a comprehensive evaluation indicator. , when I> , the support unit to be analyzed is determined to be a support unit to be maintained.

[0059] For example, for a support unit to be analyzed, the neighborhood deviation coefficient D = 0.7, the service life Y = 10 years, the historical maintenance record H = 2 times, the comprehensive score of the real-time monitored physical parameters P = 0.8, and the weights of each factor are w1 = 0.3, w2 = 0.2, w3 = 0.2, and w4 = 0.3. Then the comprehensive evaluation index I = 0.3 × 0.7 + 0.2 × 10 + 0.2 × 2 + 0.3 × 0.8 = 2.77. Assume that the threshold value of the comprehensive evaluation index I is set to T =2.5, since 2.77>2.5, the computer system determines that the support unit to be analyzed is a support unit to be maintained.

[0060] In the embodiment of the present application, the debugging steps of the support state decision algorithm include: Step S10: Obtaining the target support unit and the local sensor network sample corresponding to the target support unit in the bridge support IoT sensor network sample.

[0061] The process of the computer system obtaining the target bearing unit in the bridge bearing IoT sensor network sample requires traversing each bearing unit in the entire network sample and randomly selecting according to certain rules. This process can be implemented in many ways, the following is one way:

[0062] Random sampling technology: The computer system assigns a unique identifier, such as a number, to each support unit in the bridge support IoT sensor network sample. Then, a random number is generated within the number range of all support units through a random number generation algorithm, and the support unit corresponding to the random number is the selected target support unit. For example, suppose there are 100 support units in the bridge support IoT sensor network sample, numbered from 1 to 100. The computer system generates a random number between 1 and 100 through a random number generation algorithm. Assuming that the generated random number is 35, the support unit numbered 35 is determined as the target support unit.

[0063] To obtain local sensor network samples, it is necessary to determine local sensor network positive samples and local sensor network negative samples respectively. The specific process is as follows:

[0064] The computer system takes the selected target support unit as the starting unit and wanders to the surrounding support units until the preset wandering number is met. The local sensor network positive samples corresponding to the target support unit are obtained according to the wandering support units. This process can be implemented by using the graph traversal algorithm, including the depth-first search (DFS) and breadth-first search (BFS) algorithms. Here, the breadth-first search algorithm is used as an example to illustrate: the computer system puts the target support unit into a queue. Then, the head element is taken from the queue, all the unvisited adjacent vertices of the element (i.e., the surrounding support units) are visited, and these adjacent vertices are placed in the queue. Repeat this process until the preset wandering number is met or the queue is empty. For example, suppose the target support unit is A, and its adjacent support units are B, C, and D. The computer system first puts A into the queue, then takes A out, visits its adjacent vertices B, C, and D, and puts them into the queue. Then, the head element B is taken out, and the unvisited adjacent vertices of B are visited, and so on, until the preset wandering number is met. After sampling and obtaining the local sensor network, the computer system shields the characteristics of the target support unit. This is to exclude the influence of the characteristics of the target support unit itself on the characteristics of the local sensor network samples in the subsequent analysis, so that the local sensor network positive samples can better reflect the characteristics of the surrounding support units. For example, assuming that the characteristics of the target support unit include displacement, pressure, temperature and other data, the computer system can set these characteristic data to specific values ​​(such as 0) or delete these characteristic data, thereby achieving characteristic shielding.

[0065] The computer system determines the local sensor network negative samples corresponding to the target support unit based on the local sensor network positive samples of the remaining target support units that are different from the target support unit. Specifically, after the computer system obtains multiple target support units and their corresponding local sensor network positive samples, for a specific target support unit, the local sensor network positive samples of other target support units are used as the local sensor network negative samples of the target support unit. For example, suppose there are three target support units A, B, and C, and the computer system obtains their local sensor network positive samples SA, SB, and SC respectively. Then for the target support unit A, SB and SC are its local sensor network negative samples; for the target support unit B, SA and SC are its local sensor network negative samples; for the target support unit C, SA and SB are its local sensor network negative samples.

[0066] Through step S10, the computer system can accurately obtain the target support unit and the corresponding local sensor network sample in the bridge support IoT sensor network sample, providing the necessary data basis for subsequent algorithm debugging. For example, in an actual bridge support IoT sensor network sample, the computer system selects the target support unit X by random sampling, and then uses the breadth-first search algorithm to walk to obtain the local sensor network containing the support units around X, and obtains the local sensor network positive sample after feature shielding. Then, the local sensor network negative sample of X is determined based on the local sensor network positive samples of other target support units, thereby completing the data acquisition process of step S10 and preparing for subsequent feature mapping, error calculation, and algorithm debugging steps.

[0067] Step S20: using the initialization support state decision algorithm, feature mapping is performed on the local sensor network samples to obtain a local two-dimensional representation array, and feature mapping is performed on the target support unit to obtain a one-dimensional support representation array.

[0068] The computer system first obtains a two-dimensional array of sample degrees, that is, a degree matrix, representing the degrees of each support unit in the local sensor network sample. In graph theory, the degree of a vertex refers to the number of edges connected to the vertex. For the graph structure of the local sensor network sample, each support unit can be regarded as a vertex, and its degree reflects the degree of connection between the support unit and other surrounding support units. The computer system traverses each support unit in the local sensor network sample and counts the number of edges connected to it, thereby obtaining the degree of each support unit. Arranging these degree values ​​in a certain order constitutes a two-dimensional array of sample degrees. For example, suppose there is a local sensor network sample containing 4 support units A, B, C, and D, where A is connected to B and C, B is connected to A and D, C is connected to A, and D is connected to B. Then the degree of A is 2, the degree of B is 2, the degree of C is 1, and the degree of D is 1. The corresponding two-dimensional array of sample degrees (degree matrix) D can be expressed as: .

[0069] After obtaining the sample degree two-dimensional array, the computer system uses the array to perform standardization operations on the connection relationship between the support units in the sample connectivity array (adjacency matrix) of the local sensor network sample to obtain a sample dependency two-dimensional array used to represent the dependency between each support unit in the local sensor network sample. The sample connectivity array (adjacency matrix) is used to represent the connection relationship between each support unit in the local sensor network sample. If there is an edge connecting two support units, the corresponding matrix element is 1, otherwise it is 0. For the local sensor network sample in the above example, its sample connectivity array (adjacency matrix) A is: The purpose of the normalization operation is to eliminate the influence of the difference in the degree of different support units on the connection relationship, so that the dependence between the support units can be compared on a relatively fair scale. One normalization method is to divide each element in the sample connectivity array by the square root of the degree value of the corresponding row and column. The specific formula is as follows: ;in, Represents the element in the i-th row and j-th column of the two-dimensional array L of sample dependence, represents the element in the i-th row and j-th column of the sample connectivity array A, D ii and D jj They represent the elements in the i-th row, i-th column and the j-th row, j-th column in the sample degree two-dimensional array D. For the above example, after calculation, the sample dependency two-dimensional array L is:

[0070] ;

[0071] After obtaining the sample dependency two-dimensional array, an integration operation is performed on it and the sample initial feature array of the local sensor network sample to obtain a local two-dimensional characterization array. The sample initial feature array contains the initial feature information of each support unit in the local sensor network sample, such as the physical properties of the support unit, the data collected by the sensor, etc. The purpose of the integration operation is to combine the dependency information between the support units with their initial feature information to more comprehensively describe the characteristics of the local sensor network sample. One integration operation method is to repeatedly integrate the array members in the sample initial feature array according to the dependency between the support units in the sample dependency two-dimensional array until the number of integration operations meets the set maximum number. Assuming that the sample initial feature array is X, and its dimension matches the sample dependency two-dimensional array L, the integration operation can be iteratively performed using the following formula: ;in, represents the result after the kth integration operation, Represents an activation function, which is used to introduce nonlinear transformation. For example, a commonly used activation function can be the ReLU function, which is defined as Assuming that the maximum number of integrations is set to 3, the initial sample initial feature array , then the result after the first integration operation is: ; Continue the second and third integration operations according to the above formula, and finally use the integration operation result obtained by the latest integration operation as the local two-dimensional representation array.

[0072] The computer system uses the initial support state decision algorithm to perform feature mapping on the target support unit to obtain a one-dimensional characterization array of the support. This process is similar to the principle of feature mapping for local sensor network samples, but since the target support unit is a single vertex, its feature mapping is relatively simple. The computer system first extracts the initial features of the target support unit, which may include various physical properties of the support unit and data collected by the sensor. Then, these initial features are converted into a one-dimensional characterization array by initializing the feature mapping function in the support state decision algorithm. For example, assuming that the initial features of the target support unit include displacement d, pressure p and temperature t, the feature mapping function can be a simple linear transformation function f: f(x)=w1d+w2p+w3t+b; where w1, w2, w3 are weight parameters and b is a bias term. By adjusting the values ​​of the weight parameters and the bias term, the feature mapping function can be adapted to different application requirements and data characteristics. After the calculation of the feature mapping function, the result is a one-dimensional characterization array of the target support unit.

[0073] Step S30: Aggregate the local two-dimensional representation array to obtain a sample local one-dimensional representation array.

[0074] The local two-dimensional characterization array contains the complex relationship and characteristic information between each support unit in the local sensor network sample, but in order to facilitate subsequent analysis and calculation, it is necessary to aggregate these two-dimensional information and convert them into a one-dimensional sample local one-dimensional characterization array. This can simplify the data structure while retaining key information, making it easier for the computer system to perform unified processing and analysis. For example, when analyzing the state of a bridge support, a one-dimensional characterization array can more intuitively reflect the overall characteristics of the support unit in the local area, which is helpful for comparison and correlation analysis with the characteristics of the target support unit.

[0075] The computer system can use a variety of implementation methods to aggregate the local two-dimensional representation array. The following are several methods and examples: Mean aggregation is an aggregation method. The computer system calculates the mean of each column element in the local two-dimensional representation array, and combines these means into a one-dimensional array as the sample local one-dimensional representation array. Assume that the local two-dimensional representation array is H, and its dimension is m×n, where m represents the number of rows, corresponding to the number of support units in the local sensor network sample, and n represents the number of columns, corresponding to the characteristic dimension of each support unit. The calculation formula of mean aggregation is as follows: ; Among them, h i represents the i-th element in the local one-dimensional representation array of the sample, H ji Represents the element in the j-th row and i-th column of the local two-dimensional representation array H.

[0076] For example, suppose there is a local two-dimensional representation array H as follows: ; Using the mean aggregation operation, the computer system calculates the mean of each column: ; The local one-dimensional representation array of the sample is obtained as .

[0077] Another way is maximum pooling aggregation, which selects the maximum value of each column in the local two-dimensional representation array and combines these maximum values ​​into a one-dimensional array as the local one-dimensional representation array of the sample. This method can highlight the maximum eigenvalue of each support unit in each feature dimension in the local sensor network sample, which has certain advantages in capturing key information and outliers.

[0078] Another way is weighted aggregation, which considers the importance difference of elements in different rows in the local two-dimensional representation array, assigns a weight to each element, and then calculates the weighted sum to obtain the local one-dimensional representation array of the sample.

[0079] Through the above aggregation operations, the computer system converts the local two-dimensional characterization array into a local one-dimensional characterization array of the sample, which provides an important data basis for calculating the domain deviation error value, feature consistency error value and debugging the initialization support state decision algorithm in subsequent steps.

[0080] Step S40: obtaining a domain deviation error value according to the consistency score between the local one-dimensional characterization array of the sample and the one-dimensional characterization array of the support in the feature domain, and determining a target deviation error value based on the domain deviation error value.

[0081] The computer system performs a dot multiplication operation on the one-dimensional representation array of the support and the flip array of the local one-dimensional representation array of the positive samples of the local sensor network to obtain a first integrated representation array. Assume that the one-dimensional representation array of the support is , the local one-dimensional representation array of the positive samples of the local sensor network is , whose flip array is The formula for the dot product operation is: ;in, represents the first integrated representation array. For example, if .

[0082] When the target support unit corresponds to multiple local sensor network positive samples, the computer system activates the first integrated representation array corresponding to each local sensor network positive sample. The activation function can be a ReLU function, that is, . Calculate the mean of all activation results to get the first deviation coefficient. Assume that the results after activation are , then the mean is calculated as: ; This is the first deviation coefficient.

[0083] The computer system performs a dot multiplication operation on the one-dimensional representation array of the support and the flip array of the sample local one-dimensional representation array of the negative sample of the local sensor network to obtain a second integrated representation array. The calculation process is similar to the dot multiplication operation based on the positive sample. Assume that the sample local one-dimensional representation array of the negative sample of the local sensor network is , whose flip array is , then the second integrated representation array .

[0084] The computer system performs reasoning based on the second integrated representation array to obtain a second deviation coefficient, and the process is similar to obtaining the first deviation coefficient. The second integrated representation array corresponding to all negative samples is activated and averaged to obtain the second deviation coefficient.

[0085] The computer system determines the domain deviation error value based on the first deviation coefficient and the second deviation coefficient. One way is to calculate the absolute value of the difference between the two, that is, the domain deviation error value For example, if the second deviation coefficient , then the domain deviation error value .

[0086] In some cases, other factors need to be considered to determine the target deviation error value. For example, when the support unit deviation error value is introduced, the computer system determines the target deviation error value based on the field deviation error value and the support unit deviation error value. Assume that the support unit deviation error value is E s , the target deviation error value E can be determined by weighted summation t , the formula is: ;in, is a weight parameter used to balance the contribution of the field deviation error value and the support unit deviation error value to the target deviation error value. For example, if ,but: .

[0087] Through step S40, the computer system can obtain the domain deviation error value and the target deviation error value, which reflect the consistency degree and the comprehensive deviation between the target support unit and the local sensor network sample.

[0088] Step S50: Determine the feature consistency error value based on the commonality measurement results between the local one-dimensional characterization array of the sample and the one-dimensional characterization array of the support.

[0089] The computer system first determines the positive commonality measurement results between the sample local one-dimensional representation array of the positive sample of the local sensor network and the one-dimensional representation array of the support. The positive commonality measurement results reflect the similarity between the target support unit and the positive sample of the local sensor network at the feature level. The measurement method is, for example, cosine similarity calculation or Euclidean distance calculation, which will not be described here. Next, the negative commonality measurement results between the sample local one-dimensional representation array of the negative sample of the local sensor network and the one-dimensional representation array of the support are determined. The calculation method is similar to the positive commonality measurement result, which will not be described here.

[0090] After obtaining the positive commonality measurement results and the negative commonality measurement results, the computer system determines the feature consistency error value based on the comparison error between the two. The comparison error reflects the degree of difference between the positive sample and the negative sample and the target support unit at the feature level. One method of calculating the comparison error is to calculate the absolute value of the difference between the two.

[0091] The feature consistency error value can be obtained by normalizing or weighting the contrast error. For example, simple normalization can limit the contrast error to the interval [0, 1].

[0092] Through step S50, the computer system can determine the feature consistency error value, which can reflect the consistency of the local one-dimensional characterization array of the sample and the one-dimensional characterization array of the support at the feature level. During the algorithm debugging process, the feature consistency error value can be used as an important indicator to evaluate the mapping effect of the algorithm on the feature and guide the further optimization and adjustment of the algorithm. For example, if the feature consistency error value is large, it means that the algorithm may not have captured the feature relationship between the target support unit and the local sensor network sample well during the feature mapping process, and the algorithm needs to be improved to improve the accuracy and consistency of the feature mapping.

[0093] Step S60: Generate a comprehensive error value according to the target deviation error value and the feature consistency error value, and use the comprehensive error value to repeatedly debug the initialized support state decision algorithm until the preset debugging end condition is met to obtain the support state decision algorithm.

[0094] The computer system obtains a first weight parameter representing the contribution of the target deviation error value in the debugging process, and the first weight parameter gradually increases as the debugging process progresses. One method for determining the first weight parameter is to calculate it based on the target iteration critical number and the number of debuggings that have been completed in the current debugging process. Specifically, the computer system first obtains the target iteration critical number and the number of debugging times n that have been completed in the current debugging phase, and then the ratio of the number of completed debugging times to the target iteration critical number is determined as the first weight parameter , the calculation formula is: ; For example, assuming the target iteration critical number =100, at a certain moment, the number of debugging times completed in the current debugging phase is n=30, then the first weight parameter =30 / 100=0.3. As the debugging process progresses, when n gradually increases, The value of will also gradually increase, reflecting the increasing importance of the target deviation error value in the later stage of debugging.

[0095] The computer system obtains a second weight parameter representing the contribution of the feature consistency error value in the debugging process, and the second weight parameter gradually decreases as the debugging process progresses. Since the sum of the first weight parameter and the second weight parameter is usually 1, the second weight parameter β can be calculated by the following formula: β=1-α; using the above example, when α=0.3, the second weight parameter β=1-0.3=0.7. As the debugging process progresses, α gradually increases and β gradually decreases accordingly, which reflects that in the early stage of debugging, the feature consistency error value has a greater impact on the comprehensive error value, and as the debugging proceeds, the impact of the target deviation error value gradually dominates.

[0096] The computer system calculates the target deviation error value E according to the first weight parameter and the second weight parameter. target and feature consistency error value E consistency Fusion is performed to obtain the comprehensive error value E total The fusion method is weighted summation, and the calculation formula is: E total =αE target +βE consistency ; For example, suppose that during a certain debugging, the target deviation error value E target =0.5, feature consistency error value E consistency =0.4, the first weight parameter α=0.3, the second weight parameter β=0.7, then the comprehensive error value is: E total =0.3×0.5+0.7×0.4=0.15+0.28=0.43.

[0097] The computer system uses the calculated comprehensive error value to repeatedly debug the initial support state decision algorithm. During each debugging process, the computer system adjusts the algorithm parameters according to the comprehensive error value so that the comprehensive error value gradually decreases. The specific debugging method can use optimization algorithms such as the gradient descent method to calculate the gradient of the comprehensive error value to the algorithm parameters, and then update the parameters in the direction of the gradient. For example, for a simple linear regression model y=w1x1+w2x2+b (where w1 and w2 are weight parameters, x1 and x2 are input features, and b is a bias term), assuming the comprehensive error value E total The gradient of parameter w1 is , the learning rate is η, then the update formula of parameter w1 is: The computer system continuously repeats the above process, that is, calculating the comprehensive error value and adjusting the algorithm parameters according to the comprehensive error value, until the preset debugging end condition is met. The preset debugging end condition can be that the comprehensive error value is less than a certain threshold, or the number of debugging times reaches the maximum number of iterations, etc. For example, the threshold of the comprehensive error value is set to 0.05, and the maximum number of iterations is 200. When the comprehensive error value E total When it is less than 0.05 or the number of debugging times reaches 200, the debugging process ends, and the support state decision algorithm obtained at this time is the algorithm after debugging and optimization.

[0098] As an implementation mode, the local sensor network sample includes a local sensor network positive sample and a local sensor network negative sample; step S10, obtaining a target support unit and a local sensor network sample corresponding to the target support unit in the bridge support IoT sensor network sample, including:

[0099] Step S11: each support unit in the bridge support IoT sensor network sample walks in sequence, and each wandering support unit is randomly selected to obtain a target support unit;

[0100] Step S12: starting from the target support unit in the bridge support IoT sensor network sample and moving toward the surrounding support units until a preset number of moves is met, and then obtaining a local sensor network active sample corresponding to the target support unit based on the moved support units;

[0101] Step S13: Determine the local sensor network negative samples corresponding to the target support unit according to the local sensor network positive samples of the remaining target support units different from the target support unit.

[0102] In step S11, the computer system traverses all the support units in the bridge support IoT sensor network sample and randomly selects one as the target support unit. This process of wandering and random selection is to comprehensively and unbiasedly cover all vertices in the network sample, ensuring that each support unit has the opportunity to be selected as the target support unit, so that the local sensor network sample obtained subsequently is more representative.

[0103] The computer system can use a graph traversal algorithm to implement the sequential wandering of each support unit. Graph traversal algorithms include depth-first search (DFS) and breadth-first search (BFS). Here, the breadth-first search algorithm is used as an example for explanation. The breadth-first search algorithm starts from a starting vertex, visits all neighbor vertices of the vertex in hierarchical order, and then visits the neighbor vertices of these neighbor vertices in turn until the entire graph is traversed or a specific termination condition is reached. In specific implementation, the computer system can use a queue to assist in the breadth-first search. First, put the starting vertex into the queue. Then, take the head vertex from the queue, visit all the neighbor vertices of the vertex that have not been visited, and put these neighbor vertices into the queue. Repeat this process until the queue is empty.

[0104] For example, suppose there is a simple bridge support IoT sensor network sample, which contains 5 support units, marked as A, B, C, D, E, and the connection relationship between them is as follows: A is connected to B and C, B is connected to A and D, C is connected to A and E, D is connected to B, and E is connected to C. The process of breadth-first search with A as the starting vertex is as follows:

[0105] First, put A into the queue, and the queue is [A]. Take out the first vertex A, visit its neighbor vertices B and C, and put them into the queue. The queue is [B, C]. Take out the first vertex B, visit its unvisited neighbor vertex D, and put D into the queue. The queue is [C, D]. Take out the first vertex C, visit its unvisited neighbor vertex E, and put E into the queue. The queue is [D, E]. Take out the first vertex D, because its neighbor vertices have been visited, no operation is performed, and the queue is [E]. Take out the first vertex E, because its neighbor vertices have been visited, no operation is performed, and the queue is empty, and the traversal ends. In this way, the computer system can roam each support unit in the bridge support IoT sensor network sample in turn.

[0106] Alternatively, the implementation method of randomly selecting the target support unit is as follows: after completing the walk of all support units, the computer system needs to randomly select one of these support units as the target support unit. A simple method is to assign a unique number to each support unit, and then use a random number generation algorithm to generate a random number within the number range. The support unit corresponding to the random number is the target support unit. For example, for the above network sample containing 5 support units, A, B, C, D, and E are assigned numbers 1, 2, 3, 4, and 5 respectively. Assuming that the random number generated by the random number generation algorithm is 3, the corresponding support unit C is selected as the target support unit.

[0107] In step S12, the computer system starts from the selected target support unit and moves to the surrounding support units until the number of the wandering support units reaches the preset wandering number. The support units passed during the wandering process constitute the local sensor network positive samples corresponding to the target support unit. This wandering method is intended to obtain the local network structure information around the target support unit, so as to subsequently analyze the relationship between the target support unit and the surrounding environment.

[0108] Implementation of the walking process: The walking process can also be implemented using a graph traversal algorithm, such as a breadth-first search algorithm or a depth-first search algorithm. Here, the depth-first search algorithm is used as an example. The depth-first search algorithm starts from the starting vertex and visits vertices as deeply as possible along a path until it cannot continue or a specific condition is met, then backtracks to the previous step and continues to explore other unvisited paths. In specific implementations, the computer system can use recursion or a stack to assist in implementing the depth-first search.

[0109] For example, assuming that the target support unit is C, the preset number of walks is 4, and the network structure is still the simple network containing 5 support units mentioned above. The depth-first search process is as follows:

[0110] Starting from the target support unit C, first visit one of its neighbor vertices, assuming it is A. Starting from A, visit its unvisited neighbor vertex, assuming it is B. Starting from B, visit its unvisited neighbor vertex, assuming it is D. At this point, 4 support units (C, A, B, D) have been wandered, meeting the preset wandering number, and the wandering ends. The support units C, A, B, and D passed during the wandering process constitute the local sensor network positive samples corresponding to the target support unit C.

[0111] In step S13, the computer system determines the local sensor network negative samples corresponding to the current target support unit based on the local sensor network positive samples of other target support units except the current target support unit. The local sensor network negative samples are obtained to provide comparative information in subsequent analysis to help the computer system better understand the difference between the target support unit and other areas.

[0112] Determine the implementation method of the local sensor network negative sample: Assume that there are multiple target support units in the bridge support IoT sensor network sample, namely C1, C2, C3, etc. For the target support unit C1, its local sensor network positive sample is S1. Then, the computer system can use the local sensor network positive samples (such as S2, S3, etc.) of other target support units (such as C2, C3, etc.) as the local sensor network negative samples of the target support unit C1.

[0113] For example, suppose there are three target support units C1, C2, and C3, and their corresponding local sensor network positive samples are For the target support unit C1, its local sensor network negative samples can be S2 and S3; for the target support unit C2, its local sensor network negative samples can be S1 and S3; for the target support unit C3, its local sensor network negative samples can be S1 and S2.

[0114] Through steps S11 to S13, the computer system can accurately obtain the target bearing unit and the corresponding local sensor network positive samples and negative samples in the bridge bearing IoT sensor network sample. These data provide a basis for subsequent algorithm debugging and analysis, and help the computer system learn and understand the relationship between bridge bearing units and the characteristic differences in different regions, thereby improving the accuracy and reliability of the bearing state decision algorithm. For example, in the actual bridge bearing maintenance decision, by analyzing the positive and negative samples of the local sensor network, the computer system can more accurately determine whether the target bearing unit is in an abnormal state, providing a scientific basis for maintenance work.

[0115] As an implementation mode, step S12, taking the target support unit in the bridge support IoT sensor network sample as the starting unit to wander to the surrounding support units until a preset wandering number is met, and obtaining a local sensor network active sample corresponding to the target support unit according to the wandering support units, includes:

[0116] Step S121: Take the target support unit in the bridge support IoT sensor network sample as the starting unit and wander to the surrounding support units until the preset wandering number is met, so as to sample and obtain a local sensor network, shield the characteristics of the target support unit in the sampled local sensor network, and obtain a local sensor network positive sample corresponding to the target support unit.

[0117] The computer system starts from the target support unit and walks to the surrounding support units. This process aims to explore the network structure around the target support unit and obtain local information related to it. The walking process can use a variety of graph traversal algorithms, including depth-first search (DFS) and breadth-first search (BFS). Here, the breadth-first search algorithm is used as an example. It is a hierarchical search strategy that starts from the starting vertex and first visits all vertices closest to the starting vertex, and then visits vertices that are slightly farther away in turn, and so on.

[0118] In specific implementation, the computer system can use a queue to assist in completing the breadth-first search. First, put the target support unit into the queue. Then, take the head element from the queue, visit all the unvisited adjacent vertices of the element (i.e., the surrounding support units), and put these adjacent vertices into the queue. Repeat this process until the preset number of walks is met or the queue is empty.

[0119] After obtaining the local sensor network, the computer system shields the characteristics of the target support unit. The purpose of feature shielding is to eliminate the influence of the characteristics of the target support unit itself on the characteristics of the local sensor network samples, so that the local sensor network positive samples can better reflect the characteristics of the surrounding support units, thereby providing purer local environmental information for subsequent analysis.

[0120] There are many ways to implement feature masking. One way is to set the feature data of the target support unit to a specific value, such as setting all feature values ​​to 0 or a specific constant. Assume that each support unit has three features, namely displacement d, pressure p and temperature t. The original feature data of the target support unit A is (d A , p A , t A )=(2, 3, 4). When performing feature shielding, the computer system modifies its feature data to (0, 0, 0).

[0121] Another method is to directly delete the characteristic data of the target support unit. In this case, when the local sensor network positive samples are subsequently analyzed, the computer system will ignore the characteristic information of the target support unit and only focus on the characteristics of other support units.

[0122] Through such feature shielding operations, the computer system obtains local sensor network positive samples corresponding to the target support unit. Such local sensor network positive samples can more accurately reflect the environmental characteristics around the target support unit, providing a more valuable data basis for subsequent algorithm debugging and analysis. For example, in the subsequent calculation of field deviation error values, feature consistency error values, etc., such local sensor network positive samples can help the computer system more accurately evaluate the difference between the target support unit and the surrounding environment, thereby improving the accuracy and reliability of the support state decision algorithm.

[0123] As an implementation mode, step S20, performing feature mapping on the local sensor network samples to obtain a local two-dimensional representation array, includes:

[0124] Step S21: obtaining a two-dimensional array of sample degrees representing the degrees of each support unit in the local sensor network sample;

[0125] Step S22: using the sample degree two-dimensional array to perform a standardization operation on the connection relationship between the support units in the sample connectivity array of the local sensor network sample, to obtain a sample dependency two-dimensional array for representing the dependency between each support unit in the local sensor network sample;

[0126] Step S23: integrating the sample dependency two-dimensional array with the sample initial feature array of the local sensor network sample to obtain a local two-dimensional representation array.

[0127] In step S21, the computer system determines the degree of each support unit in the local sensor network sample and forms these degree values ​​into a two-dimensional array, namely, a sample degree two-dimensional array, also known as a degree matrix. In graph theory, the degree of a vertex refers to the number of edges connected to the vertex. For the graph structure of the local sensor network sample, each support unit can be regarded as a vertex, and its degree reflects the degree of connection between the support unit and other surrounding support units.

[0128] The computer system calculates the degree value by traversing each support unit in the local sensor network sample and counting the number of edges connected to it. The specific implementation method can be achieved by operating the adjacency matrix of the network. The adjacency matrix is ​​a two-dimensional array used to represent the connection relationship between vertices in the graph. If there is an edge connecting vertex i and vertex j, the element in the i-th row and j-th column of the adjacency matrix is ​​1, otherwise it is 0. For an undirected graph, the adjacency matrix is ​​a symmetric matrix.

[0129] For example, suppose there is a local sensor network sample containing 4 support units, marked as A, B, C, and D, and the connection relationship between them is: A is connected to B and C, B is connected to A and D, C is connected to A, and D is connected to B. Then the adjacency matrix A of the local sensor network sample is: ; The degree of support unit A: degree(A)=1+1+0=2; The degree of support unit B: degree(B)=1+0+1=2; The degree of support unit C: degree(C)=1+0+0=1; The degree of support unit D: degree(D)=0+1+0=1; Arrange these degree values ​​in the order of support units to obtain the sample degree two-dimensional array (degree matrix) D: .

[0130] In step S22, after obtaining the sample degree two-dimensional array, the computer system uses the array to perform a standardization operation on the sample connectivity array (i.e., adjacency matrix) of the local sensor network sample, in order to obtain a sample dependency two-dimensional array that can represent the dependency between each support unit. This standardization operation can eliminate the influence of the difference in the degree of different support units on the connection relationship, so that the dependency between each support unit can be compared on a relatively fair scale.

[0131] One normalization method is to divide each element in the sample connectivity array by the square root of the degree value of the corresponding row and column.

[0132] In step S23, after obtaining the sample dependency two-dimensional array, the computer system integrates it with the sample initial feature array of the local sensor network sample to obtain a local two-dimensional representation array. The sample initial feature array contains the initial feature information of each support unit in the local sensor network sample, such as the physical properties of the support unit, the data collected by the sensor, etc. The purpose of the integration operation is to combine the dependency information between the support units with their initial feature information to more comprehensively describe the characteristics of the local sensor network sample.

[0133] One integration operation method is to perform matrix multiplication on the sample dependency two-dimensional array and the sample initial feature array. Assuming that the sample dependency two-dimensional array is L, whose dimension is n×n (n is the number of support units), and the sample initial feature array is X, whose dimension is n×m (m is the number of features of each support unit), the calculation formula of the local two-dimensional representation array H is: H=LX.

[0134] Through the above steps S21 to S23, the computer system can extract more representative and comprehensive feature information from the local sensor network samples to obtain a local two-dimensional characterization array. This two-dimensional characterization array integrates the connection relationship and initial feature information between the bearing units, providing a richer data basis for subsequent analysis and decision-making. For example, in the maintenance decision of bridge bearings, the local two-dimensional characterization array can be used to evaluate the overall status of the bearing units in the local area, and then determine whether certain bearing units need to be maintained.

[0135] As an implementation mode, step S23, integrating the sample dependency two-dimensional array and the sample initial feature array of the local sensor network sample to obtain a local two-dimensional representation array, includes:

[0136] Step S231: According to the dependency between the support units in the sample dependency two-dimensional array, the array members in the sample initial feature array are repeatedly integrated until the number of integration operations meets the set maximum number, and the integration operation result obtained by the latest integration operation is used as the local two-dimensional characterization array.

[0137] The sample dependency two-dimensional array reflects the dependency relationship between each support unit in the local sensor network sample, which can be understood as a connection relationship, while the sample initial feature array contains the initial feature information of each support unit. By integrating the sample initial feature array according to the dependency in the sample dependency two-dimensional array, the computer system can more fully integrate the relationship information between the support units and their own feature information, thereby obtaining a more representative and comprehensive local two-dimensional representation array. This integration operation helps to explore the hidden feature patterns and laws in the local sensor network samples, providing a richer and more accurate information basis for subsequent analysis and decision-making.

[0138] The computer system integrates the sample initial feature array according to the inter-support unit dependency in the sample dependency two-dimensional array. The specific implementation method is to achieve it through repeated application of matrix multiplication and activation function. The following is a detailed operation process and example description:

[0139] Assume that the sample dependency two-dimensional array is L, whose dimension is n×n (n is the number of support units), and the sample initial feature array is X, whose dimension is n×m (m is the number of features of each support unit). In each integration operation, the computer system first performs matrix multiplication operation on the sample dependency two-dimensional array L and the sample initial feature array X to obtain a new array X', calculated as: X'=LX.

[0140] For example, suppose there is a local sensor network sample containing three support units, the sample dependency two-dimensional array L and the sample initial feature array X are: ; Then the X' obtained by the first integration operation is: ; In order to introduce nonlinear transformation and enhance the expressiveness of the model, the computer system applies an activation function after each matrix multiplication operation. The role of the activation function is to perform nonlinear mapping on the result of matrix multiplication, so that the model can learn more complex feature relationships. Activation functions include ReLU (Rectified Linear Unit), Sigmoid, Tanh, etc. Here we take the ReLU function as an example, which is defined as: ; That is, for the input value x, if x is greater than 0, then output x; if x is less than or equal to 0, then output 0. Apply the activation function to the above-obtained X' to obtain the activated result X'': The computer system will repeat the above matrix multiplication and activation function application operations until the number of integration operations meets the set maximum number. Assuming that the maximum number is set to 3 times, after the first integration operation obtains X'', X'' is used as the new sample initial feature array to continue the next integration operation.

[0141] Second integration operation:

[0142] ;

[0143] ;

[0144] The third integration operation:

[0145] ;

[0146] ;

[0147] At this time, the number of integration operations has reached the set maximum number of 3 times, and the computer system uses the result X'''''' obtained from the latest integration operation as the local two-dimensional representation array.

[0148] Through such repeated integration operations, the computer system can continuously fuse the dependencies and feature information between the bearing units, so that the obtained local two-dimensional representation array can more accurately reflect the characteristics of the local sensor network samples. This integration operation plays an important role in dealing with complex network structures and feature relationships, and can provide more powerful support for subsequent analysis and decision-making. For example, in bridge bearing maintenance decisions, a more accurate local two-dimensional representation array can help to more accurately evaluate the status of the bearing units, thereby improving the accuracy and reliability of maintenance decisions.

[0149] As an implementation mode, the local sensor network sample includes a local sensor network positive sample and a local sensor network negative sample; step S40, according to the consistency score of the sample local one-dimensional characterization array and the support one-dimensional characterization array in the feature domain, obtains the domain deviation error value, including:

[0150] Step S41: performing a dot multiplication operation on the one-dimensional representation array of the support and the flip array of the sample local one-dimensional representation array of the local sensor network active samples to obtain a first integrated representation array;

[0151] Step S42: performing inference according to the first integrated representation array to obtain a first deviation coefficient;

[0152] Step S43: performing a dot multiplication operation on the one-dimensional representation array of the support and the flip array of the sample local one-dimensional representation array of the negative samples of the local sensor network to obtain a second integrated representation array;

[0153] Step S44: performing inference according to the second integrated representation array to obtain a second deviation coefficient;

[0154] Step S45: Determine the domain deviation error value according to the first deviation coefficient and the second deviation coefficient.

[0155] In step S41, the computer system first obtains the support one-dimensional characterization array and the sample local one-dimensional characterization array of the local sensor network active samples. The support one-dimensional characterization array is a one-dimensional array obtained after feature mapping of the target support unit, which contains the feature information of the target support unit. The sample local one-dimensional characterization array of the local sensor network active samples is a one-dimensional array obtained after feature aggregation of the local sensor network active samples around the target support unit, reflecting the characteristics of the environment around the target support unit.

[0156] In order to perform the dot multiplication operation, the computer system needs to flip the local one-dimensional representation array of the positive samples of the local sensor network to obtain its flipped array. Assume that the one-dimensional representation array of the support is , the local one-dimensional representation array of the positive samples of the local sensor network is , whose flip array is The formula for the dot product operation is: ;

[0157] in, represents the first integrated characterization array. For example, assuming the support one-dimensional characterization array =(1, 2, 3), the local one-dimensional representation array of the sample of the local sensor network positive sample =(4, 5, 6), then it flips the array , the first integrated representation array obtained by performing a dot multiplication operation .

[0158] In step S42, after obtaining the first integrated characterization array, the computer system needs to infer based on the array to obtain the first deviation coefficient. When the target support unit corresponds to multiple local sensor network positive samples, the computer system will perform an activation operation for the first integrated characterization array corresponding to each local sensor network positive sample. The role of the activation function is to introduce nonlinear transformation and enhance the expression ability of the model. Activation functions include ReLU (Rectified Linear Unit), Sigmoid, Tanh, etc. Here, the ReLU function is taken as an example, and its definition is: ; That is, for the input value x, if x is greater than 0, then output x; if x is less than or equal to 0, then output 0.

[0159] Assuming that the target support unit corresponds to 3 local sensor network positive samples, the first integrated representation arrays are . After activating these arrays, we get:

[0160] ;

[0161] ; ;

[0162] Then, the computer system calculates the mean of all activation results to obtain the first deviation coefficient. The formula for calculating the mean is: ;in, represents the first deviation coefficient, m represents the number of positive samples in the local sensor network, represents the result after activation of the first integrated representation array corresponding to the i-th local sensor network positive sample. For the above example, the first deviation coefficient for: Step S43 is similar to step S41. The computer system performs a dot multiplication operation on the support one-dimensional representation array and the flip array of the sample local one-dimensional representation array of the local sensor network negative sample to obtain a second integrated representation array. Assume that the sample local one-dimensional representation array of the local sensor network negative sample is , whose flip array is , then the formula for the dot multiplication operation is: .

[0163] in, represents the second integrated characterization array. For example, assuming the support one-dimensional characterization array , the sample local one-dimensional representation array of negative samples in the local sensor network , then the reverse array , the second integrated representation array obtained by performing a dot multiplication operation for: .

[0164] Step S44 is similar to step S42, and inference is performed based on the second integrated characterization array to obtain a second deviation coefficient. When the target support unit corresponds to multiple local sensor network negative samples, the computer system will also perform an activation operation for each second integrated characterization array corresponding to the local sensor network negative sample, and perform mean calculation on all activation results to obtain the second deviation coefficient.

[0165] Assuming that the target support unit corresponds to two local sensor network negative samples, the second integrated representation arrays are . After activating these arrays, we get:

[0166] ; Then, the computer system calculates the mean of all activation results to obtain the second deviation coefficient. The formula for the mean calculation is: ;in, represents the second deviation coefficient, k represents the number of negative samples in the local sensor network, represents the result after activation of the second integrated representation array corresponding to the negative sample of the i-th local sensor network. For the above example, the second deviation coefficient for: .

[0167] In step S45, after obtaining the first deviation coefficient and the second deviation coefficient, the computer system determines the domain deviation error value according to the two coefficients. One way is to calculate the absolute value of the difference between the two, that is, the domain deviation error value E d The calculation formula is: ;in, Indicates the absolute value operation. For the above example, the first deviation coefficient , the second deviation coefficient , then the domain deviation error value E d for: .

[0168] Through the above steps S41 to S45, the computer system can accurately obtain the domain deviation error value according to the consistency score of the local one-dimensional characterization array of the sample and the one-dimensional characterization array of the support in the feature domain. This error value reflects the consistency between the target support unit and the local sensor network sample, which is of great significance for evaluating the status of the target support unit and subsequent maintenance decisions.

[0169] As an embodiment, the method further includes:

[0170] Step S70: using the initialization support state decision algorithm, further performing different feature mapping on the local sensor network samples to obtain a graph structure feature integration array;

[0171] Step S80: restore and map the graph structure feature integration array to obtain a sample feature decoding two-dimensional array;

[0172] Step S90: Determine the support unit deviation error value according to the error between the feature decoding one-dimensional array of the target support unit in the sample feature decoding two-dimensional array and the initial feature array of the target support unit.

[0173] Based on this, step S40, determining a target deviation error value based on the domain deviation error value, includes: determining the target deviation error value according to the domain deviation error value and the support unit deviation error value.

[0174] In step S70, the computer system uses the initialization bearing state decision algorithm to perform additional feature mapping operations on the local sensor network samples to obtain a graph structure feature integration array. This process is intended to extract richer and more representative feature information from the local sensor network samples in order to more comprehensively analyze and understand the state of the bridge bearings.

[0175] Specifically, the initial support state decision algorithm may include multiple different feature mapping functions or modules, each of which focuses on capturing the characteristics of a specific aspect of the local sensor network sample. For example, it may include a module for analyzing the topological relationship between support units and a module for deep mining of the sensor data of each support unit.

[0176] Assume that the local sensor network sample contains 5 support units, marked as A, B, C, D, and E, and each support unit has 3 initial features, such as displacement d, pressure p, and temperature t. The sample initial feature array X is: ; The initialization support state decision algorithm can process the topological structure information of the sample through a module based on a graph convolutional network (GCN). GCN extracts the local and global features of the vertex (support unit) by convolving the adjacency matrix (representing the connection relationship between support units) and the initial feature array of the sample. Assume that after processing by the GCN module, a new feature array X1 is obtained, whose dimension is the same as X, but the value of each element has been extracted and transformed by the GCN feature.

[0177] Then, a module based on a recurrent neural network (RNN) is used to process the time series information of the sensor data (assuming that the sensor data is collected in time sequence). RNN can capture the temporal dependencies in the data and further enrich the feature representation. After processing by the RNN module, another feature array X2 is obtained.

[0178] Finally, the computer system integrates the feature arrays obtained from these different modules, for example, through a splicing operation, to obtain a graph structure feature integration array Z. Assuming that the dimensions of X1 and X2 are both 5×3, the dimension of the spliced ​​graph structure feature integration array Z is 5×6, indicating that each support unit now has 6 features.

[0179] In step S80, after obtaining the graph structure feature integration array, the computer system performs a restoration mapping, i.e., a decoding operation, on it to obtain a sample feature decoding two-dimensional array. The purpose of the restoration mapping is to restore the abstract feature representation obtained after feature mapping and integration back to a form with similar structure and semantics to the original data, so as to better understand and analyze these features.

[0180] One way to restore the mapping is to use a decoder network, such as a decoder corresponding to the encoder (the network used in the feature mapping process). The structure of the decoder network is usually symmetrical with the encoder network, and it attempts to recover the information of the original data from the encoded features. Taking the previously obtained graph structure feature integration array Z as an example, assume that the decoder network is a multilayer perceptron (MLP). The computer system passes Z as input to the MLP, and the MLP gradually restores the features through a series of linear transformations and activation function operations. Assume that the MLP has two hidden layers, the first hidden layer has 8 neurons, the second hidden layer has 4 neurons, and the number of neurons in the output layer is the same as the feature dimension of the original sample initial feature array X, that is, 3.

[0181] The specific calculation process is as follows: Assuming that the weight matrix of the first hidden layer is W1 and the bias vector is b1, the output H1 of the first hidden layer is: H1=σ(ZW1+b1); where σ is the activation function, such as the ReLU function.

[0182] Similarly, the output H2 of the second hidden layer is: H2=σ(H1W2+b2); where W2 and b2 are the weight matrix and bias vector of the second hidden layer.

[0183] Finally, the output of the output layer, that is, the sample feature decoding two-dimensional array Y is: Y=H2W3+b3; where W3 and b3 are the weight matrix and bias vector of the output layer.

[0184] After such decoding operations, the computer system obtains the sample feature decoding two-dimensional array Y, whose dimension is the same as the original sample initial feature array X. The value of each element represents the characteristic value of the corresponding support unit obtained after feature mapping, integration and restoration.

[0185] In step S90, the computer system determines the deviation error value of the support unit by comparing the difference between the feature decoding one-dimensional array of the target support unit in the sample feature decoding two-dimensional array and the initial feature array of the target support unit. This error value reflects the degree of deviation between the feature of the target support unit and the original feature after a series of feature processing, which is of great significance for evaluating the accuracy of feature mapping and restoration mapping and the state of the target support unit.

[0186] Assuming the target support unit is A, its feature decoding one-dimensional array in the sample feature decoding two-dimensional array Y is , the initial feature array in the sample initial feature array X is One way to calculate the error is the mean squared error (MSE), which is calculated as: ; where n is the dimension of the feature, for the above example, n=3.

[0187] For example, if , then the support unit deviation error value is: .

[0188] In the sub-step of step S40, the computer system comprehensively considers the domain deviation error value and the support unit deviation error value to determine the target deviation error value. The target deviation error value is a comprehensive indicator that can more comprehensively reflect the state of the target support unit in the local sensor network and the error situation in the feature processing process.

[0189] One method to determine the target deviation error value is weighted summation, that is, assigning certain weights to the domain deviation error value and the support unit deviation error value, and then adding them together to obtain the target deviation error value. Assume that the domain deviation error value is E d , the deviation error value of the support unit is E mse , and assign them weights α and β (α+β=1), then the target deviation error value E t The calculation formula is: E t =αE d +βE mse ; For example, assuming the domain deviation error value E d =0.8, support unit deviation error value E mse =0.1, weight α=0.7, β=0.3, then the target deviation error value E t For: E t =0.7×0.8+0.3×0.1=0.56+0.03=0.59.

[0190] As an implementation mode, the local sensor network sample includes a local sensor network positive sample and a local sensor network negative sample; step S50, according to the commonality measurement result between the sample local one-dimensional characterization array and the support one-dimensional characterization array, determines the feature consistency error value, including:

[0191] Step S51: determining the positive commonality measurement result between the sample local one-dimensional characterization array of the positive sample of the local sensor network and the support one-dimensional characterization array;

[0192] Step S52: determining the negative commonality measurement result between the sample local one-dimensional characterization array of the negative sample of the local sensor network and the support one-dimensional characterization array;

[0193] Step S53: Determine a feature consistency error value according to a comparison error between the positive commonality measurement result and the negative commonality measurement result.

[0194] In step S51, the computer system first obtains the sample local one-dimensional representation array of the local sensor network active sample and the support one-dimensional representation array. The sample local one-dimensional representation array of the local sensor network active sample here is a one-dimensional array obtained by performing a series of feature mapping and aggregation operations on the local sensor network active sample, which represents a representation of the local sensor network active sample in the feature space. The support one-dimensional representation array is a one-dimensional array obtained after feature mapping of the target support unit, which represents the feature representation of the target support unit.

[0195] In order to determine the positive commonality measurement result between them, the computer system can adopt a variety of implementation methods, such as calculating cosine similarity or Euclidean distance, etc., which are not limited here.

[0196] In step S52, the computer system also first obtains the sample local one-dimensional representation array and the support one-dimensional representation array of the local sensor network negative sample. The sample local one-dimensional representation array of the local sensor network negative sample corresponds to the local sensor network positive sample, which is determined based on the local sensor network positive samples of the remaining target support units different from the target support unit. The method for determining the negative commonality measurement result is similar to the method for determining the positive commonality measurement result, and can also be calculated using cosine similarity or Euclidean distance.

[0197] In step S53, after obtaining the positive commonality measurement results and the negative commonality measurement results, the computer system calculates their contrast errors to determine the feature consistency error value. The contrast error can reflect the degree of difference in features between the positive samples and negative samples of the local sensor network and the target support unit.

[0198] One method of calculating the contrast error is to directly calculate the difference between the two. The feature consistency error value can be determined based on the contrast error. One way is to use the contrast error as the feature consistency error value, that is, the feature consistency error value is equal to E. Another way is to further process the contrast error, such as mapping it through a nonlinear function, to better adapt to specific application scenarios. For example, a Sigmoid function can be used to map the contrast error. The formula of the Sigmoid function is: ; Substitute the contrast error E into the Sigmoid function to obtain the feature consistency error value E f : ; In the above example, substituting E=0.68 into the Sigmoid function yields: Through the above steps, the computer system can determine the feature consistency error value, which can be used to balance the feature domain during feature mapping, play an important role in subsequent algorithm debugging, and help improve the accuracy and performance of the support state decision algorithm.

[0199] As an implementation manner, step S60, generating a comprehensive error value according to the target deviation error value and the feature consistency error value, includes:

[0200] Step S61: obtaining a first weight parameter representing the contribution of the target deviation error value in the debugging process, wherein the first weight parameter gradually increases as the debugging process progresses;

[0201] Step S62: obtaining a second weight parameter representing the contribution of the feature consistency error value in the debugging process, wherein the second weight parameter gradually decreases as the debugging process progresses;

[0202] Step S63: According to the first weight parameter and the second weight parameter, the target deviation error value and the feature consistency error value are merged to obtain a comprehensive error value.

[0203] In step S61, the target deviation error value reflects the consistency between the target support unit and the local sensor network sample and the deviation of the support unit itself, which has different importance for the optimization of the algorithm at different debugging stages. In order to reflect this change in importance, the computer system needs to obtain a first weight parameter representing the contribution of the target deviation error value in the debugging link, and the parameter gradually increases as the debugging process deepens.

[0204] One feasible implementation method is to determine the first weight parameter by setting a function related to the number of debugging times. Assume that the debugging process is performed for a total of N iterations, and the number of debuggings currently completed is n ( ), then the first weight parameter It can be calculated by the following linear increasing function: For example, assume that the entire debugging process is set to perform 100 iterations (ie, N=100). When debugging reaches the 20th iteration (ie, n=20), the first weight parameter is calculated according to the above formula: ; This indicates that when debugging is performed for the 20th time, the contribution of the target deviation error value in the debugging process is 0.2. As the number of debugging increases, for example, when debugging is performed for the 80th time (i.e., n=80), then: At this time, the contribution of the target deviation error value in the debugging process has increased to 0.8, which shows that as the debugging process deepens, the importance of the target deviation error value gradually increases.

[0205] In step S62, the feature consistency error value is used to balance the feature domain during feature mapping, which is more important for algorithm optimization in the early stage of debugging, because it is necessary to pay attention to the consistency of features in the initial stage to ensure the accurate mapping of features by the algorithm. As the debugging process deepens, its importance gradually decreases. The computer system needs to obtain a second weight parameter representing the contribution of the feature consistency error value in the debugging link, and this parameter gradually decreases as the debugging process deepens.

[0206] Similar to obtaining the first weight parameter, the second weight parameter can be determined by setting a function related to the number of debugging times. Assume that the debugging process has a total of N iterations, and the number of debuggings currently completed is n ( ), then the second weight parameter It can be calculated by the following linear decreasing function: For example, assuming that the entire debugging process is set to perform 100 iterations (ie, N=100). When debugging is performed for the 20th time (ie, n=20), the second weight parameter is calculated according to the above formula: ; This shows that when debugging is carried out for the 20th time, the contribution of the feature consistency error value in the debugging process is 0.8. When debugging is carried out for the 80th time (i.e. n=80), then: At this time, the contribution of the feature consistency error value in the debugging process is reduced to 0.2, which shows that as the debugging process deepens, the importance of the feature consistency error value gradually decreases.

[0207] In step S63, after obtaining the first weight parameter and the second weight parameter Finally, the computer system needs to fuse (such as adding or weighted summing) the target deviation error value and the feature consistency error value according to the two weight parameters to obtain a comprehensive error value. The comprehensive error value can comprehensively consider the importance of the target deviation error value and the feature consistency error value at different debugging stages, thereby more effectively guiding the debugging of the algorithm.

[0208] In this way, the computer system merges the target deviation error value and the feature consistency error value according to their contribution in the current debugging stage, and the obtained comprehensive error value can more comprehensively reflect the error situation of the algorithm in the current state. In the subsequent debugging process, the computer system will adjust the initial bearing state decision algorithm according to the comprehensive error value, so that the algorithm is continuously optimized to improve the accuracy and reliability of bridge bearing maintenance decisions. As the debugging process deepens, the first weight parameter gradually increases, and the target deviation error value accounts for an increasingly higher proportion in the comprehensive error value, reflecting a greater focus on the target deviation; while the second weight parameter gradually decreases, and the proportion of the feature consistency error value gradually decreases, which is in line with the changing law of the degree of attention paid to different error factors at different debugging stages.

[0209] As an implementation method, step S61, obtaining a first weight parameter representing the contribution of the target deviation error value in the debugging link, includes: obtaining the critical number of target iterations and the number of debuggings that have been completed in the current debugging link; and determining the ratio of the number of completed debuggings to the critical number of target iterations as the first weight parameter.

[0210] Specifically, the computer system first obtains the target critical number of iterations and the number of debuggings that have been completed in the current debugging phase. The target critical number of iterations is an important parameter pre-set in the entire debugging process. It represents a key limit or an important reference value for the total number of times the entire debugging process is expected to be performed. The setting of this value is usually based on a comprehensive consideration of multiple factors such as the complexity of the problem, the size of the data, and the convergence of the algorithm. For example, in the debugging of the bridge bearing maintenance decision algorithm, based on past experience and preliminary analysis of the algorithm, researchers can set the target critical number of iterations to 1,000. This means that the entire debugging process is expected to be repeated 1,000 times to bring the algorithm to a more ideal state.

[0211] The number of debugging times completed in the current debugging phase refers to the actual number of debugging cycles that the computer system has completed at the current moment. The computer system can accurately obtain this value through an internal counter or recording mechanism. For example, when the computer system is debugging an algorithm, it will automatically add 1 to the variable recording the number of debugging times after each debugging cycle is completed. Assuming that at a certain moment, the computer system has completed 300 debugging cycles, then the number of debugging times completed in the current debugging phase is 300.

[0212] After obtaining the target iteration critical number and the number of debugging completed in the current debugging link, the computer system determines the ratio of the completed debugging number to the target iteration critical number as the first weight parameter. The principle of this determination method is that as the debugging process progresses, the ratio of the completed debugging number to the target iteration critical number will gradually increase, so that the first weight parameter will also gradually increase, reflecting the characteristic that the contribution of the target deviation error value in the debugging link gradually increases as the debugging process deepens.

[0213] The specific calculation method can be expressed as: ;in, represents the first weight parameter, n represents the number of debugging times completed in the current debugging phase, represents the target critical number of iterations. For example, assuming the target critical number of iterations is =1000, when the computer system has completed 200 debuggings (ie n=200), the first weight parameter is calculated according to the above formula: ; This indicates that at this time, the contribution of the target deviation error value in the debugging process is 0.2. As the debugging continues, when the number of completed debugging increases to 600 times (ie, n=600), the first weight parameter becomes: . At this time, the contribution of the target deviation error value in the debugging link is increased to 0.6. In this way, the first weight parameter can dynamically adjust the importance of the target deviation error value in the comprehensive consideration according to the progress of the debugging process, so that in the early stage of debugging, other factors (such as feature consistency error value) may be relatively more concerned, and as the debugging approaches the critical number of target iterations, the role of the target deviation error value gradually becomes prominent, which is more conducive to the optimization and adjustment of the algorithm at different stages, improves the accuracy and convergence of the algorithm, and enables it to better adapt to the needs of the complex task of bridge bearing maintenance decision-making.

[0214] As an implementation mode, the target local sensor network includes an active target local sensor network covering the support unit to be analyzed and a passive target local sensor network not covering the support unit to be analyzed; step S300, according to the consistency score of the target local one-dimensional representation array and the target support one-dimensional representation array in the feature domain, determines the neighborhood deviation coefficient, including:

[0215] Step S310: obtaining a positive consistency score between the target local one-dimensional representation array of the positive target local sensor network and the target support one-dimensional representation array in the feature domain (the calculation method can refer to the calculation method of the first deviation coefficient mentioned above);

[0216] Step S320: Obtain the negative consistency score of the target local one-dimensional representation array and the target support one-dimensional representation array of the negative target local sensor network in the feature domain; Step S330: Determine the neighborhood deviation coefficient target domain deviation error value based on the error between the positive consistency score and the negative consistency score.

[0217] The calculation method of the positive consistency score in step S310 can refer to the calculation method of the first deviation coefficient mentioned above. First, the relevant data of the positive target local sensor network is processed. The positive target local sensor network is a local sensor network covering the support unit to be analyzed. The corresponding target local one-dimensional representation array is obtained through a series of feature mapping and aggregation operations, which can reflect the characteristics of the local sensor network in the feature space. The target support one-dimensional representation array is a one-dimensional array obtained after feature mapping of the support unit to be analyzed, which represents the feature representation of the support unit.

[0218] To obtain a positive consistency score, the computer system uses a calculation similar to the first deviation coefficient.

[0219] The negative consistency score in step S320 may be calculated in a manner similar to the calculation method of the second deviation coefficient.

[0220] Negative consistency scores are calculated in a similar manner to positive consistency scores, using a calculation similar to the second deviation coefficient.

[0221] Step S330 After obtaining the positive consistency score and the negative consistency score, the computer system needs to determine the neighborhood deviation coefficient target domain deviation error value based on the error between the two. This error reflects the degree of difference in characteristics between the local sensor network covering the intended support unit and the local sensor network not covering the unit and the target support unit, thereby measuring the degree of inconsistency between the intended support unit and the surrounding support units, that is, the degree of neighborhood deviation. One way to calculate the error is to directly calculate the difference between the two.

[0222] Through the above steps, the computer system can determine the deviation error value of the neighborhood deviation coefficient target area according to the error between the positive consistency score and the negative consistency score. This value is of great significance for judging whether the bearing unit to be analyzed is a bearing unit to be maintained, and can provide a key basis for the maintenance decision of the bridge bearing.

[0223] As an embodiment, the method further includes:

[0224] Step S301: using the support state decision algorithm after debugging and convergence, further adopting different feature mappings for the target local sensor network to obtain a target graph structure feature integration array;

[0225] Step S302: restore and map the target graph structure feature integration array to obtain a target feature decoding two-dimensional array;

[0226] Step S303: Determine the support unit deviation coefficient according to the error between the target feature decoding two-dimensional array and the initial feature array of the support unit in the target local sensor network.

[0227] Based on this, step S400, determining whether the support unit to be analyzed is a support unit to be maintained according to the neighborhood deviation coefficient, includes: determining whether the support unit to be analyzed is a support unit to be maintained according to the neighborhood deviation coefficient and the support unit deviation coefficient.

[0228] In step S301, the support state decision algorithm that has been debugged and converged is first called. After previous training and debugging, the algorithm has the ability to effectively process and analyze bridge support-related data. The target local sensor network contains local network information related to the support unit to be analyzed, and the computer system passes it as input data to the debugged and converged support state decision algorithm.

[0229] In order to obtain the target graph structure feature integration array, the computer system will adopt different feature mapping methods for the target local sensor network. Feature mapping is a way to convert raw data into more representative and discriminative features, which can mine the potential information in the data for subsequent analysis and processing. For example, the computer system can use a graph convolutional neural network (GCN) for feature mapping. GCN is a deep learning method designed specifically for graph structure data. It learns the representation of vertices by aggregating and updating the features of vertices and their neighbor vertices in the graph.

[0230] Assume that the target local sensor network can be represented as a graph G=(V, E), where V is the vertex set (i.e., support units) and E is the edge set (indicating the connection relationship between support units). There is an initial eigenvector The basic operation of GCN can be expressed as: ;in, is the vertex feature matrix of layer l, is the weight matrix of the lth layer, is the adjacency matrix A plus the self-connection (I is the identity matrix), yes is the degree matrix, and σ is the activation function. Through the operation of multiple layers of GCN, the computer system can obtain the target graph structure feature integration array, which contains the graph structure information after deep feature learning.

[0231] For example, suppose there is a simple target local sensor network, which contains 5 support units (vertices), whose initial eigenvectors are x1=(1, 2), x2=(3, 4), x3=(5, 6), x4=(7, 8), x5=(9, 10), and the adjacency matrix A represents the connection relationship between these vertices. After the first layer of GCN calculation (assuming the weight matrix and activation function σ has been determined), we can get the vertex feature matrix H of the first layer 1 . Continue to perform calculations on the second and third layers, and finally obtain the target graph structure feature integration array.

[0232] In step S302, after obtaining the target graph structure feature integration array, the computer system needs to perform a restoration mapping operation on it to obtain the target feature decoding two-dimensional array. Restoration mapping (decoding) is an operation corresponding to feature mapping (encoding). Its purpose is to restore the abstract feature representation obtained after feature mapping to a form with similar structure and semantics to the original data, so as to perform subsequent error calculation and analysis.

[0233] One way to restore the mapping is to use a decoder network. Assuming that the target graph structure feature integration array is Z, the input of the decoder network is Z. The decoder network gradually restores the features through a series of deconvolution operations and activation function operations. For example, a layer of the decoder network can be expressed as: ;in, is the output feature of the decoder layer l, is the input feature of the decoder layer l, is the weight matrix of the decoder layer l, and σ is the activation function. After the operation of the multi-layer decoder network, the computer system finally obtains the target feature decoding two-dimensional array. For example, suppose the target graph structure feature integration array Z is a feature matrix after multi-layer GCN encoding, and its dimension is n×m (n represents the number of vertices, and m represents the feature dimension). The decoder network starts from Z and obtains a new feature matrix through the first layer of deconvolution operation. , whose dimensions may vary depending on the design of the decoder. Continue with the deconvolution operation of the subsequent layers, and finally obtain the target feature decoding 2D array, whose dimensions may be close to the original initial feature array dimensions for subsequent error calculation.

[0234] In step S303, after obtaining the target feature decoding two-dimensional array, the computer system needs to calculate the error between it and the initial feature array of the support unit in the target local sensor network, so as to determine the support unit deviation coefficient. This error reflects the difference between the feature information and the original feature after feature mapping and restoration mapping, so as to measure the feature deviation of the support unit.

[0235] One way to calculate the error is the mean squared error (MSE). Assume that the target feature decoding two-dimensional array is , where i represents the index of the vertex (support unit) and j represents the index of the feature; the initial feature array of the support unit in the target local sensor network is The calculation formula of mean square error is: ; where n is the number of vertices (support elements) and m is the dimension of the feature.

[0236] For example, suppose there is a target local sensor network consisting of 3 support units, each with 2 features. Target feature decoded 2D array , the initial feature array of the support unit in the target local sensor network According to the mean square error formula, we can get: After obtaining the mean square error, the computer system can use it as the support unit deviation coefficient, or further normalize it so that it takes values ​​within a certain range, so as to make comprehensive comparison and analysis with other coefficients.

[0237] In step S410, after obtaining the neighborhood deviation coefficient and the support unit deviation coefficient, the computer system needs to determine whether the support unit to be analyzed is a support unit to be maintained based on these two coefficients. A simple and direct method is to add the two coefficients and then determine whether the result is greater than the maintenance threshold. If so, the support unit to be analyzed is determined to be a support unit to be maintained.

[0238] For example, suppose the maintenance threshold is set to T=0.8. For a support unit to be analyzed, its neighborhood deviation coefficient is 0.5 and the support unit deviation coefficient is 0.4, then the two are added to get 0.5+0.4=0.9. Since 0.9>0.8, the computer system determines that the support unit to be analyzed is a support unit to be maintained.

[0239] Another method is to assign different weights to the neighborhood deviation coefficient and the support unit deviation coefficient, and then perform a weighted sum and compare it with the maintenance threshold. Assuming that the weight of the neighborhood deviation coefficient is w1=0.6 and the weight of the support unit deviation coefficient is w2=0.4, the calculation formula for the weighted sum is: ; For example, for a support unit to be analyzed, its neighborhood deviation coefficient is 0.4, and the support unit deviation coefficient is 0.6, then the weighted summation result is: S=0.6×0.4+0.4×0.6=0.48; assuming that the maintenance threshold is still T=0.8, since 0.48<0.8, the computer system determines that the support unit to be analyzed is not a support unit to be maintained.

[0240] As an implementation manner, step S410, determining whether the support unit to be analyzed is a support unit to be maintained according to the neighborhood deviation coefficient and the support unit deviation coefficient, includes:

[0241] Step S411: determining a maximum neighborhood deviation coefficient and a minimum neighborhood deviation coefficient among neighborhood deviation coefficients of each bearing unit in the target bridge bearing IoT sensor network;

[0242] Step S412: using the maximum neighborhood deviation coefficient and the minimum neighborhood deviation coefficient, normalizing the neighborhood deviation coefficient of the support element to be analyzed to obtain a neighborhood deviation normalization value;

[0243] Step S413: determining a maximum support unit deviation coefficient and a minimum support unit deviation coefficient among the support unit deviation coefficients of each support unit in the target bridge support IoT sensor network;

[0244] Step S414: using the maximum support unit deviation coefficient and the minimum support unit deviation coefficient, standardizing the support unit deviation coefficient of the support unit to be analyzed to obtain the support unit deviation standardization coefficient;

[0245] Step S415: The neighborhood deviation normalization value and the support unit deviation normalization coefficient are fused to obtain a fusion result of the support unit to be analyzed. When the fusion result meets a set threshold, the support unit to be analyzed is determined to be a support unit to be maintained.

[0246] In step S411, the computer system first traverses the neighborhood deviation coefficients of all bearing units in the target bridge bearing IoT sensor network. The neighborhood deviation coefficient reflects the degree of inconsistency between each bearing unit and the surrounding bearing units. In order to find the maximum neighborhood deviation coefficient and the minimum neighborhood deviation coefficient, the computer system can use a simple comparison algorithm.

[0247] Specifically, the computer system first assigns the neighborhood deviation coefficient of the first support unit to the initial values ​​of the maximum neighborhood deviation coefficient and the minimum neighborhood deviation coefficient. Then, starting from the second support unit, the neighborhood deviation coefficient of each support unit is compared with the current maximum neighborhood deviation coefficient and the minimum neighborhood deviation coefficient. If the neighborhood deviation coefficient of a support unit is greater than the current maximum neighborhood deviation coefficient, the maximum neighborhood deviation coefficient is updated to the neighborhood deviation coefficient of the support unit; if the neighborhood deviation coefficient of a support unit is less than the current minimum neighborhood deviation coefficient, the minimum neighborhood deviation coefficient is updated to the neighborhood deviation coefficient of the support unit.

[0248] For example, suppose there are 5 support units in the target bridge support IoT sensor network, and their neighborhood deviation coefficients are 0.2, 0.5, 0.1, 0.8, and 0.3, respectively. The computer system first assigns the neighborhood deviation coefficient of the first support unit, 0.2, to the maximum neighborhood deviation coefficient and the minimum neighborhood deviation coefficient. Then, the neighborhood deviation coefficient of the second support unit, 0.5, is compared. Since 0.5 is greater than 0.2, the maximum neighborhood deviation coefficient is updated to 0.5. Next, the neighborhood deviation coefficient of the third support unit, 0.1, is compared. Since 0.1 is less than 0.2, the minimum neighborhood deviation coefficient is updated to 0.1. The neighborhood deviation coefficient of the fourth support unit, 0.8, is compared. Since 0.8 is greater than 0.5, the maximum neighborhood deviation coefficient is updated to 0.8. Finally, the neighborhood deviation coefficient of the fifth support unit, 0.3, is compared. The maximum neighborhood deviation coefficient and the minimum neighborhood deviation coefficient remain unchanged. Therefore, in this example, the maximum neighborhood deviation coefficient is 0.8 and the minimum neighborhood deviation coefficient is 0.1.

[0249] In step S412, the standardization operation is to convert the neighborhood deviation coefficient of the support unit to be analyzed into a unified and comparable range for subsequent analysis and processing. The computer system can adopt a linear standardization implementation method, and its formula is: Standardized value = original value minimum value / maximum value minimum value, where the original value is the neighborhood deviation coefficient of the support unit to be analyzed, the minimum value is the minimum neighborhood deviation coefficient determined in step S411, and the maximum value is the maximum neighborhood deviation coefficient determined in step S411.

[0250] In step S413, similar to determining the maximum neighborhood deviation coefficient and the minimum neighborhood deviation coefficient, the computer system needs to traverse the support unit deviation coefficients of all support units in the target bridge support IoT sensor network to find the maximum support unit deviation coefficient and the minimum support unit deviation coefficient. The same comparison algorithm can be used to achieve this. For example, assume that there are 5 support units in the target bridge support IoT sensor network, and their support unit deviation coefficients are 0.3, 0.6, 0.2, 0.9, and 0.4, respectively. The computer system compares and updates in the same way as determining the maximum neighborhood deviation coefficient and the minimum neighborhood deviation coefficient. First, the support unit deviation coefficient of the first support unit, 0.3, is assigned to the initial value of the maximum support unit deviation coefficient and the minimum support unit deviation coefficient. Then, the support unit deviation coefficients of other support units are compared in turn, and finally the maximum support unit deviation coefficient is 0.9 and the minimum support unit deviation coefficient is 0.2.

[0251] In step S414, the linear normalization method is also used to normalize the support unit deviation coefficient of the support unit to be analyzed, and the formula is the same as the normalization formula in step S412.

[0252] In step S415, the fusion operation is to comprehensively consider the influence of the neighborhood deviation normalization value and the support unit deviation normalization coefficient on the state of the support unit to be analyzed. One fusion method is weighted summation, that is, different weights are assigned to the neighborhood deviation normalization value and the support unit deviation normalization coefficient, and then they are weighted and added to obtain the fusion result. Another simple method is direct addition.

[0253] If the direct addition method is used, the fusion result is the sum of the neighborhood deviation standardization value and the support unit deviation standardization coefficient. For example, if the neighborhood deviation standardization value is 0.43 and the support unit deviation standardization coefficient is 0.43, the fusion result is 0.43+0.43=0.86. Assuming that the threshold is set to 0.8, since 0.86>0.8, the computer system determines that the support unit to be analyzed is a support unit to be maintained.

[0254] Through the above steps S411 to S415, the computer system can comprehensively consider the neighborhood deviation coefficient and the bearing unit deviation coefficient, accurately judge whether the bearing unit to be analyzed is a bearing unit to be maintained, and provide a scientific basis for maintenance decisions of bridge bearings.

[0255] Figure 3 Schematic diagram of the structure of the computer system provided in the embodiment of the present application. Figure 3 As shown, the above-mentioned computer system 10 may include: a processor 1001, a network interface 1004 and a memory 1005. In addition, the above-mentioned computer system 1000 may also include: a user interface 1003, and at least one communication bus 1002. Among them, the communication bus 1002 is used to realize the connection and communication between these components. Among them, the user interface 1003 may include a display screen (Display), a keyboard (Keyboard), and the user interface 1003 may optionally include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 may be a high-speed RAM memory, or it may be a non-volatile memory (non-volatile memory), such as at least one disk storage. The memory 1005 may optionally also be at least one storage device away from the aforementioned processor 1001. As Figure 3 As shown, the memory 1005 as a computer-readable storage medium may include an operating system, a network communication module, a user interface module, and a device control application. Figure 3 In the computer system 1000 shown, the network interface 1004 can provide network communication functions; the user interface 1003 is mainly used to provide an input interface; and the processor 1001 can be used to call the device control application stored in the memory 1005 to implement the method provided in the above embodiment.

Claims

1. A bridge bearing maintenance decision-making method combining the Internet of Things and big data, characterized in that: The method comprises: Acquire a support unit to be analyzed and a target local sensor network corresponding to the support unit to be analyzed in the target bridge support IoT sensor network; Using the support state decision algorithm after debugging and convergence, feature mapping is performed on the target local sensor network to obtain a target local two-dimensional representation array, and feature mapping is performed on the support unit to be analyzed to obtain a target support one-dimensional representation array; Aggregating the target local two-dimensional representation array to obtain a target local one-dimensional representation array, and determining a neighborhood deviation coefficient according to a consistency score between the target local one-dimensional representation array and the target support one-dimensional representation array in a feature domain; Determining whether the support unit to be analyzed is a support unit to be maintained according to the neighborhood deviation coefficient; The debugging steps of the support state decision algorithm include: Acquire a target bearing unit and a local sensor network sample corresponding to the target bearing unit from the bridge bearing IoT sensor network sample; Using an initialization support state decision algorithm, feature mapping is performed on the local sensor network samples to obtain a local two-dimensional representation array, and feature mapping is performed on the target support unit to obtain a support one-dimensional representation array; Aggregating the local two-dimensional representation array to obtain a sample local one-dimensional representation array; Obtaining a domain deviation error value according to a consistency score between the local one-dimensional characterization array of the sample and the one-dimensional characterization array of the support in a feature domain, and determining a target deviation error value based on the domain deviation error value; Determining a feature consistency error value according to a commonality measurement result between the local one-dimensional characterization array of the sample and the one-dimensional characterization array of the support; A comprehensive error value is generated according to the target deviation error value and the characteristic consistency error value, and the initialization support state decision algorithm is repeatedly debugged using the comprehensive error value until a preset debugging end condition is met, thereby obtaining the support state decision algorithm.

2. The method according to claim 1, characterized in that The local sensor network samples include local sensor network positive samples and local sensor network negative samples; the step of obtaining a target support unit from the bridge support IoT sensor network samples and the local sensor network samples corresponding to the target support unit includes: Each support unit in the bridge support IoT sensor network sample walks in sequence, and the target support unit is randomly selected from each wandering support unit; Taking the target support unit in the bridge support IoT sensor network sample as the starting unit, wandering to the surrounding support units until a preset wandering number is met, and obtaining a local sensor network active sample corresponding to the target support unit according to the wandering support units; Determine a local sensor network negative sample corresponding to the target support unit according to local sensor network positive samples of other target support units different from the target support unit; The method of taking the target support unit in the bridge support IoT sensor network sample as the starting unit and wandering to the surrounding support units until a preset wandering number is met, and obtaining a local sensor network active sample corresponding to the target support unit according to the wandering support units, includes: The target support unit in the bridge support IoT sensor network sample is taken as the starting unit to wander towards the surrounding support units until a preset wandering number is met, thereby sampling and obtaining a local sensor network. The characteristics of the target support unit in the sampled local sensor network are shielded to obtain a positive sample of the local sensor network corresponding to the target support unit.

3. The method according to claim 1, characterized in that The step of performing feature mapping on the local sensor network samples to obtain a local two-dimensional representation array includes: Obtaining a two-dimensional array of sample degrees representing the degrees of each support unit in the local sensor network sample; Using the sample degree two-dimensional array to perform a normalization operation on the connection relationship between the support units in the sample connectivity array of the local sensor network sample, to obtain a sample dependency two-dimensional array for representing the dependency between the support units in the local sensor network sample; An integration operation is performed on the sample dependency two-dimensional array and the sample initial feature array of the local sensor network samples to obtain a local two-dimensional representation array.

4. The method according to claim 3, characterized in that The step of integrating the sample dependency two-dimensional array with the sample initial feature array of the local sensor network sample to obtain a local two-dimensional representation array includes: According to the dependency between the support units in the sample dependency two-dimensional array, the array members in the sample initial feature array are repeatedly integrated until the number of integration operations meets the set maximum number, and the integration operation result obtained by the latest integration operation is used as the local two-dimensional characterization array.

5. The method according to claim 1, characterized in that The local sensor network samples include local sensor network positive samples and local sensor network negative samples; the domain deviation error value is obtained according to the consistency score between the local one-dimensional characterization array of the sample and the one-dimensional characterization array of the support in the feature domain, including: Performing a dot multiplication operation on the one-dimensional representation array of the support and the flip array of the sample local one-dimensional representation array of the active samples of the local sensor network to obtain a first integrated representation array; Performing reasoning based on the first integrated representation array to obtain a first deviation coefficient; Performing a dot multiplication operation on the one-dimensional representation array of the support and the flip array of the sample local one-dimensional representation array of the negative samples of the local sensor network to obtain a second integrated representation array; Performing inference according to the second integrated representation array to obtain a second deviation coefficient; Determining a domain deviation error value according to the first deviation coefficient and the second deviation coefficient; The method further comprises: Using an initialization support state decision algorithm, further different feature mappings are performed on the local sensor network samples to obtain a graph structure feature integration array; Performing restoration mapping on the graph structure feature integration array to obtain a sample feature decoding two-dimensional array; Determine a support unit deviation error value according to an error between a feature decoding one-dimensional array of the target support unit in the sample feature decoding two-dimensional array and an initial feature array of the target support unit; The determining of the target deviation error value based on the domain deviation error value comprises: A target deviation error value is determined based on the field deviation error value and the support unit deviation error value.

6. The method according to claim 1, characterized in that The local sensor network samples include local sensor network positive samples and local sensor network negative samples; the characteristic consistency error value is determined according to the commonality measurement result between the local one-dimensional characterization array of the samples and the one-dimensional characterization array of the support, including: Determine a positive commonality measurement result between a sample local one-dimensional characterization array of a positive sample of the local sensor network and the support one-dimensional characterization array; Determine a negative commonality measurement result between a sample local one-dimensional characterization array of negative samples of the local sensor network and the support one-dimensional characterization array; Determining a feature consistency error value according to a comparison error between the positive commonality measurement result and the negative commonality measurement result; The generating a comprehensive error value according to the target deviation error value and the feature consistency error value comprises: Acquire a first weight parameter representing the contribution of the target deviation error value in the debugging process, wherein the first weight parameter gradually increases as the debugging process progresses; Obtaining a second weight parameter representing the contribution of the feature consistency error value in the debugging process, wherein the second weight parameter gradually decreases as the debugging process progresses; The target deviation error value and the feature consistency error value are fused according to the first weight parameter and the second weight parameter to obtain a comprehensive error value.

7. The method according to claim 6, characterized in that The obtaining of a first weight parameter representing a contribution of the target deviation error value in the debugging phase includes: Get the critical number of target iterations and the number of debugging completed in the current debugging phase; The ratio of the completed debugging times to the target critical number of iterations is determined as a first weight parameter.

8. The method according to claim 1, characterized in that The target local sensor network includes a positive target local sensor network covering the support unit to be analyzed and a negative target local sensor network not covering the support unit to be analyzed; the neighborhood deviation coefficient is determined according to the consistency score between the target local one-dimensional characterization array and the target support one-dimensional characterization array in the feature domain, including: Obtaining a positive consistency score between the target local one-dimensional representation array of the positive target local sensor network and the target support one-dimensional representation array in a feature domain; Obtaining a negative consistency score between the target local one-dimensional representation array of the negative target local sensor network and the target support one-dimensional representation array in a feature domain; A neighborhood deviation coefficient target domain deviation error value is determined based on an error between the positive consistency score and the negative consistency score.

9. The method according to claim 1, characterized in that: The method further comprises: Using the support state decision algorithm after debugging and convergence, further adopting different feature mappings for the target local sensor network to obtain a target graph structure feature integration array; Performing restoration mapping on the target graph structure feature integration array to obtain a target feature decoding two-dimensional array; Determining a support unit deviation coefficient based on an error between the target feature decoding two-dimensional array and an initial feature array of a support unit in the target local sensor network; The step of determining whether the support unit to be analyzed is a support unit to be maintained according to the neighborhood deviation coefficient includes: Determining whether the support unit to be analyzed is a support unit to be maintained according to the neighborhood deviation coefficient and the support unit deviation coefficient; The step of determining whether the support unit to be analyzed is a support unit to be maintained according to the neighborhood deviation coefficient and the support unit deviation coefficient includes: Determining a maximum neighborhood deviation coefficient and a minimum neighborhood deviation coefficient among neighborhood deviation coefficients of each bearing unit in the target bridge bearing IoT sensor network; Using the maximum neighborhood deviation coefficient and the minimum neighborhood deviation coefficient, a neighborhood deviation coefficient of the support element to be analyzed is normalized to obtain a neighborhood deviation normalization value; Determining a maximum support unit deviation coefficient and a minimum support unit deviation coefficient among the support unit deviation coefficients of each support unit in the target bridge support IoT sensor network; Using the maximum support unit deviation coefficient and the minimum support unit deviation coefficient, a support unit deviation coefficient of the support unit to be analyzed is normalized to obtain a support unit deviation normalization coefficient; The neighborhood deviation normalization value and the support unit deviation normalization coefficient are fused to obtain a fusion result of the support unit to be analyzed, and when the fusion result meets a set threshold, the support unit to be analyzed is determined to be a support unit to be maintained.

10. A computer system, characterized in that: include: processor; and a memory, wherein the memory stores a computer-readable code, and when the computer-readable code is executed by the processor, the processor executes the method as claimed in any one of claims 1 to 9.

Images