Research on an Improved SOM Model for Damage Identiﬁcation of Concrete Structures

: In order to solve the problem of intelligent detection of damage of modern concrete structures under complex constraints, an improved self-organizing mapping (SOM) neural network model algorithm was proposed to construct an accurate identiﬁcation model of concrete structure damage. Based on the structure and algorithm of the SOM network model, the whole process of the core construction of the concrete structure damage identiﬁcation network model is summarized. Combined with the damage texture characteristics of concrete structures, through the self-developed 3D laser scanning system, an improved method based on a small number of samples to effectively improve the effectiveness of network input samples is proposed. Based on the principle of network topology map analysis and its image characteristics, a SOM model improvement method that can effectively improve the accuracy of the network identiﬁcation model is studied. In addition, based on the reactive powder concrete bending fatigue loading test, the feasibility and accuracy of the improved method are veriﬁed. The results show that the improved SOM concrete structure damage identiﬁcation model can effectively identify unknown neuron categories in a limited sample space, and the identiﬁcation accuracy of the SOM network model is improved by 4.69%. The proposed improved SOM model method fully combines the network topology and its unique image features and can accurately identify structural damage. This research contributes to the realization of high-precision intelligent health monitoring of damage to modern concrete structures. In addition, it is of great signiﬁcance for the timely detection, identiﬁcation and localization of early damage to structures. an improved method for constructing input samples of SOM network model based on the gray level co-occurrence matrix and digital feature screening is proposed. Based on the principle of network topology map analysis and the characteristics of grayscale images of topology maps, an improved SOM topology map analysis algorithm is proposed. The improved SOM algorithm model was applied to the bending fatigue test of reactive powder concrete. Based on the improvement of the recognition accuracy and the test effect, the validity of the proposed improved algorithm model is veriﬁed.


Introduction
Structural damage detection research is one of the most critical research contents in Structural Health Monitoring (SHM) [1][2][3][4]. As the relevant technology for structural damage detection, pattern recognition processes various forms of structural damage information to carry out structural damage analysis and is an important part of information science and artificial intelligence. Selecting an intelligent detection method suitable for practical engineering, combining damage indicators with feature-level and decision-level data, thereby simplifying calculation and inference time, and realizing efficient and automated intelligent evaluation are key issues that need further research [5][6][7][8].
The neural network has the learning ability to deal with nonlinear problems, strong fault tolerance and robustness [9,10]. Damage identification based on the neural network is based on the physical parameters or dynamic parameters of the structure in different states of health. The parameters sensitive to structural damage are selected as the input of the neural network [11]. The neural network is trained with a large number of damage cases in numerical simulations. Finally, the mature network is trained to realize automatic damage recognition based on the real structural response [12,13]. Scholars choose various pattern recognition techniques for in-depth research on structural damage recognition, such as an improved method for constructing input samples of SOM network model based on the gray level co-occurrence matrix and digital feature screening is proposed. Based on the principle of network topology map analysis and the characteristics of grayscale images of topology maps, an improved SOM topology map analysis algorithm is proposed. The improved SOM algorithm model was applied to the bending fatigue test of reactive powder concrete. Based on the improvement of the recognition accuracy and the test effect, the validity of the proposed improved algorithm model is verified.

Self-Organizing Map
In order to improve the SOM model, the network structure and network algorithm are analyzed.

Network Structure
The self-organizing map is an unsupervised feed-forward neural network model, in which neurons compete and cooperate with each other to identify pattern sets. The structure of the SOM model is shown in Figure 1.
images and an improved method for constructing input samples of SOM n based on the gray level co-occurrence matrix and digital feature screenin Based on the principle of network topology map analysis and the characte scale images of topology maps, an improved SOM topology map analys proposed. The improved SOM algorithm model was applied to the bendi of reactive powder concrete. Based on the improvement of the recognition the test effect, the validity of the proposed improved algorithm model is v

Self-Organizing Map
In order to improve the SOM model, the network structure and netw are analyzed.

Network Structure
The self-organizing map is an unsupervised feed-forward neural netw which neurons compete and cooperate with each other to identify pattern ture of the SOM model is shown in Figure 1. The input layer consists of n self-organizing neurons (x1, x2, x3, …, xn− petition layer consists of a 2D planar (s × p) array of n input vector maps model identifies pattern categories for a given data set by continuously adj nection weights of low-dimensional to high-dimensional network nodes.

Network Algorithm
As the core of model construction, the network algorithm is the key to ization and the mapping characteristics of the model. The SOM neural netw includes network initialization, input vector setting, etc.

I.
Initialize. Generally, the weight vector will be given any value in th represented by i W . The learning rate is  .
II. Set input vector input. The input vector is the network model trainin  The input layer consists of n self-organizing neurons (x 1 , x 2 , x 3 , . . . , x n−1 , x n ). The competition layer consists of a 2D planar (s × p) array of n input vector maps. The network model identifies pattern categories for a given data set by continuously adjusting the connection weights of low-dimensional to high-dimensional network nodes.

Network Algorithm
As the core of model construction, the network algorithm is the key to the selforganization and the mapping characteristics of the model. The SOM neural network algorithm includes network initialization, input vector setting, etc.

I.
Initialize. Generally, the weight vector will be given any value in the interval [0, 1], represented by W i . The learning rate is η.
II. Set input vector input. The input vector is the network model training sample: III. Derive Euclidean Distance. W ij represents the weight between the input layer neuron i, and the mapping layer neuron j. Derive the Euclidean distance between the input vector and the weight vector to get the specific position of the neuron. The Euclidean distance is calculated as: IV. Label the winning neuron. The winning neuron position is the position of the neuron with the minimum Euclidean distance between the input vector and the weight vector. The input vector is denoted by X , the winning neuron is denoted by c, Then its calculation formula is: V. Adjust weights. Correct the input neuron and the neuron connection weights in the neighborhood according to Equation (3): Among, η(t) is the learning rate at t, η(t) ∈ [0, 1], η(t) gradually decreases with time, Inversely proportional to t, its expression is: VI. Calculate the output value O k : Among, f represents the function that takes the smallest Euclidean distance. Determine whether the output results meet the requirements. If the result meets the classification requirements, output the category; if the result does not meet the category requirements, return to step (2) to continue learning until the judgment result is met. Output and end learning.

Improved SOM Damage Identification Method
Based on the core steps in the construction of the damage identification network model, the research on the improvement method of the SOM network model is carried out.

Construction of Damage Identification Model
In order to establish the damage identification model of concrete structure, the structure and algorithm characteristics of the SOM neural network are analyzed according to the performance requirements of the model. Its core steps include the selection of input samples, the setting of network parameters, the judgment of winning neurons and the analysis of topological graphs. Figure 2 is the overall process diagram of the construction method of the damage identification algorithm model for concrete structures.
its calculation formula is: V. Adjust weights. Correct the input neuron and the neuron connection weights in the neighborhood according to Equation (3): ) (t  gradually decreases with time, Inversely proportional to t, its expression is: Among, f represents the function that takes the smallest Euclidean distance. Determine whether the output results meet the requirements. If the result meets the classification requirements, output the category; if the result does not meet the category requirements, return to step (2) to continue learning until the judgment result is met. Output and end learning.

Improved SOM Damage Identification Method
Based on the core steps in the construction of the damage identification network model, the research on the improvement method of the SOM network model is carried out.

Construction of Damage Identification Model
In order to establish the damage identification model of concrete structure, the structure and algorithm characteristics of the SOM neural network are analyzed according to the performance requirements of the model. Its core steps include the selection of input samples, the setting of network parameters, the judgment of winning neurons and the analysis of topological graphs. Figure 2 is the overall process diagram of the construction method of the damage identification algorithm model for concrete structures.   Figure 2 shows the whole process of structural damage identification based on the SOM neural network, which is the core framework system of the research. Analysis of Figure 2 shows that the selection of input samples is the first step in the construction of the network model, which directly determines the network structure and is an important factor affecting the efficiency of network operation. As the last key step of the network model, the analysis of the topology structure directly determines the specific category of each neuron and is a key factor affecting the accuracy of network recognition. Therefore, in order to improve the recognition performance of the SOM network for concrete structure damage, the research will mainly focus on these two parts to improve the SOM neural network model.

a.
Selection of input samples In order to reduce the interference of complex factors such as environment and humans, a method based on machine vision is proposed to obtain input samples. The damage signal is collected based on the vision sensor, and the initial sample is extracted by the feature extraction algorithm. In order to reduce the requirement for the number of input samples, the input samples that can effectively characterize the damage characteristics are automatically screened based on statistical theory. The improved SOM model and its input sample selection process is shown in Figure 3.
SOM neural network, which is the core framework system of the research. An Figure 2 shows that the selection of input samples is the first step in the constru the network model, which directly determines the network structure and is an im factor affecting the efficiency of network operation. As the last key step of the model, the analysis of the topology structure directly determines the specific cat each neuron and is a key factor affecting the accuracy of network recognition. Th in order to improve the recognition performance of the SOM network for structure damage, the research will mainly focus on these two parts to improve neural network model.

a. Selection of input samples
In order to reduce the interference of complex factors such as environm humans, a method based on machine vision is proposed to obtain input samp damage signal is collected based on the vision sensor, and the initial sample is e by the feature extraction algorithm. In order to reduce the requirement for the nu input samples, the input samples that can effectively characterize the characteristics are automatically screened based on statistical theory. The improv model and its input sample selection process is shown in Figure 3. An input sample for constructing a damage identification network model b 3D laser scanning technology is proposed, as shown in Figure 3. First, a 3D ima specimen under the loading system is acquired by adding 1D transmission equip a 2D laser sensor. Then, the initial samples are extracted by constructing the gray occurrence matrix (GLCM) of structural damage. Finally, in order to further imp effectiveness of the damaged samples, based on the digital feature screening (D feature parameters are selected as the input samples of the network model.

b. Analysis of topology map
In order to accurately identify the damage category information containe topology map image, according to the characteristics of the topology map i network model optimization method, the topology grayscale (TOP-G) algor proposed. Figure 4 shows the flow of the TOP-G algorithm. An input sample for constructing a damage identification network model based on 3D laser scanning technology is proposed, as shown in Figure 3. First, a 3D image of the specimen under the loading system is acquired by adding 1D transmission equipment to a 2D laser sensor. Then, the initial samples are extracted by constructing the gray level co-occurrence matrix (GLCM) of structural damage. Finally, in order to further improve the effectiveness of the damaged samples, based on the digital feature screening (DFS), the feature parameters are selected as the input samples of the network model.

b.
Analysis of topology map In order to accurately identify the damage category information contained in the topology map image, according to the characteristics of the topology map image, a network model optimization method, the topology grayscale (TOP-G) algorithm, is proposed. Figure 4 shows the flow of the TOP-G algorithm.

•
The first step is to determine the grayscale of the topology map: First, determine the number L of connection polygons between neurons. The gray level of the topology map is determined according to the number of L, and the gray value range of the pixels in the topology map should be [0, L]. Thus, it is judged that the gray level of the image is g = L = 2 n , and it is deduced that n = log 2 L, g = 2 [n] , where [n] represents the value of n is the smallest integer that exceeds the value of n. Then the obtained g = 2 [n] is the gray level of the topology map.

•
The second step is to grayscale the topological distance map: According to the gray level of the topological map, the topological distance color image is converted into a grayscale image, which is called a topological grayscale map.

•
The third step is to create a sliding window: Suppose the number of neurons is m, create a sliding window, label L 1 -L m and assign grayscale values g 1 -g m to each neighborhood polygon in the topological grayscale map.

•
The fourth step is to discriminate the category of neurons: The gray values gi of all neighboring neurons of the unknown neuron i are extracted, compared and sorted. Determine whether g i is the largest gray value in the neighborhood. If so, the neuron connecting the neighborhood polygon is a class, and the output i belongs to this class; If not, rejudge until the attribution category of all unknown neurons is determined, and the result is output. Appl

Experiments and Results Analysis
In order to verify that the improved SOM neural network model can effectively improve the recognition accuracy of the network model, based on the self-developed 3D laser scanning system, a network model for the recognition of bending fatigue damage of reactive powder concrete was established.

Selection of Input Samples of RPC Bending Fatigue Damage Identification Model
There is no obvious change in the appearance of the specimen before loading in the RPC bending fatigue test. When the loading force reaches 70% to 80% of the ultimate bending strength, initial cracks appear in the mid-span accompanied by the sound of steel fibers being pulled out, and damage images are obtained during this process. It was observed that the flexural strength value did not decrease with the occurrence of mid-span cracks in the specimen until the steel fibers in the crack section were completely pulled out, and the specimen lost its bearing capacity and declared failure. With the development of the experimental phenomenon, a three-dimensional model of the concrete specimen was obtained. Figure 5 shows the entire process of acquiring images during 3D damaged specimen loading, and Figure 6 is the obtained three-dimensional model of microcrack damage.
The gray values gi of all neighboring neurons of the unknown neuron i are extracted, compared and sorted. Determine whether gi is the largest gray value in the neighborhood. If so, the neuron connecting the neighborhood polygon is a class, and the output i belongs to this class; If not, rejudge until the attribution category of all unknown neurons is determined, and the result is output.

Experiments and Results Analysis
In order to verify that the improved SOM neural network model can effectively improve the recognition accuracy of the network model, based on the self-developed 3D laser scanning system, a network model for the recognition of bending fatigue damage of reactive powder concrete was established.

Selection of Input Samples of RPC Bending Fatigue Damage Identification Model
There is no obvious change in the appearance of the specimen before loading in the RPC bending fatigue test. When the loading force reaches 70% to 80% of the ultimate bending strength, initial cracks appear in the mid-span accompanied by the sound of steel fibers being pulled out, and damage images are obtained during this process. It was observed that the flexural strength value did not decrease with the occurrence of mid-span cracks in the specimen until the steel fibers in the crack section were completely pulled out, and the specimen lost its bearing capacity and declared failure. With the development of the experimental phenomenon, a three-dimensional model of the concrete specimen was obtained. Figure 5 shows the entire process of acquiring images during 3D damaged specimen loading, and Figure 6 is the obtained three-dimensional model of microcrack damage.  In order to reduce the amount of data and improve the recognition effic dant information is removed based on a 3D point cloud projection algorithm filtering. Based on the GLCM, the damage model input samples are extracte In order to reduce the amount of data and improve the recognition efficiency, redundant information is removed based on a 3D point cloud projection algorithm and median filtering. Based on the GLCM, the damage model input samples are extracted. The image gray level g = 128 is constructed, the generation step size is d = 1, and the generation direction θ takes the gray level co-occurrence matrix of 0 • , 45 • , 90 • and 135 • . The 14 feature parameters such as angular second moment and correlation are extracted. In order to improve the quality of the input samples, the P 1 (angular Second Moment), P 2 (entropy), P 3 (inertia moment), P 4 (correlation), P 5 (inverse difference moment), and P 6 (variance) are screened out as standard samples based on the DFS method. Table 1 shows the damage texture properties represented by the input sample.  In order to reduce the amount of data and improve the recognition efficiency dant information is removed based on a 3D point cloud projection algorithm and filtering. Based on the GLCM, the damage model input samples are extracted. Th gray level g = 128 is constructed, the generation step size is d = 1, and the gener rection θ takes the gray level co-occurrence matrix of 0°, 45°, 90° and 135°. The 1 parameters such as angular second moment and correlation are extracted. In ord prove the quality of the input samples, the P1 (angular Second Moment), P2 (ent (inertia moment), P4 (correlation), P5 (inverse difference moment), and P6 (varia screened out as standard samples based on the DFS method. Table 1 shows the texture properties represented by the input sample.     Figure 7, the selected six feature parameters are used as the SOM network input vector [P 1 , P 2 , P 3 , P 4 , P 5 , P 6 ], the SOM network competition layer is set to 8 × 8 = 64 neurons and the network model output is four categories of damage.

As shown in
In the parameter setting of the SOM network model for concrete structure damage, the number of training steps directly affects the network clustering performance. In order to improve the clustering efficiency, the optimal number of training steps is obtained. After determining the structure of the network model, select different steps for training and observe the performance changes of the network model. Using the step increment as a variable, analyze the clustering results of the network model. The statistics are shown in Table 2 for the clustering results under different training steps. Selecting the training steps with the fewest steps can not only satisfy the sample classification, but also ensure the clustering speed. When the number of training steps is set to 10, 50, 100, 200, 500 and 1000, the classification effect of the network model is shown in Table 1. When the number of training steps is 10, the damage diagnosis model is initially established, and the damage is divided into two categories; as the number of training steps increases, when the number of training steps is 50 and 100, the recognition accuracy is further improved, and the injuries are divided into three categories; when the number of training steps reaches 200, the four injury types are completely distinguished; continue to increase the number of training steps to 500 and 1000, and the damage classification results are the same, which is not practical. Therefore, 200 training steps were chosen as the optimal value for the damage identification model.

Determining the Winning Neurons of RPC Bending Fatigue Damage Model
In order to further verify the accuracy of acquiring neurons when the number of training steps is 200, the topology map of the winning neuron positions of the damage type is output, as shown in Figure 8. 10  55  37  37  50  43  37  37  100  43  1  37  200  49  1  16  500  49  1  16  1000  49  1  16 When the number of training steps is set to 10, 50, 100, 200, 500 fication effect of the network model is shown in Table 1. When the steps is 10, the damage diagnosis model is initially established, and th into two categories; as the number of training steps increases, when th steps is 50 and 100, the recognition accuracy is further improved, an vided into three categories; when the number of training steps reache types are completely distinguished; continue to increase the number 500 and 1000, and the damage classification results are the same, wh Therefore, 200 training steps were chosen as the optimal value for th tion model.

Determining the Winning Neurons of RPC Bending Fatigue Damage M
In order to further verify the accuracy of acquiring neurons w training steps is 200, the topology map of the winning neuron posi type is output, as shown in Figure 8.     Table 1 and Figure 8, it is inferred that honeycombs, holes, sags and cracks correspond to winning neurons numbered 49, 1, 16 and 64, respectively. From the obtained topological positions and the number of winning neurons, the basis for the cluster analysis of the network model is basically obtained. However, further analysis of the network model is required to obtain the specific damage type for each neuron.

Neuron Topology Analysis for RPC Bending Fatigue Damage SOM Network Model
In order to obtain the damage category information corresponding to each neuron, the clustering results of the network model were analyzed. Obtain the topological structure distance map of the structure damage identification network model, as shown in Figure 9. The small gray squares in the figure represent neurons and the straight lines between them represent straight-line connections between neurons. The distance between neurons is obtained by the Euclidean distance formula. The hexagons connect the neurons, with the color depth representing the distance between neurons. The colors are from dark to light, indicating that the distance between neurons is from far to near. It can be inferred that the neurons with a light color have high similarity, and the difference between them is low; while the neurons with a dark color have low similarity, and the difference between them is large.
to obtain the specific damage type for each neuron.

Neuron Topology Analysis for RPC Bending Fatigue Damage SOM Net
In order to obtain the damage category information correspondi the clustering results of the network model were analyzed. Obtain the ture distance map of the structure damage identification network mode ure 9. The small gray squares in the figure represent neurons and th tween them represent straight-line connections between neurons. The neurons is obtained by the Euclidean distance formula. The hexagon rons, with the color depth representing the distance between neurons. T dark to light, indicating that the distance between neurons is from fa inferred that the neurons with a light color have high similarity, and tween them is low; while the neurons with a dark color have low sim ference between them is large. Based on the analysis of the traditional comparison method, amon rons of the damage identification network model, the damage types o spond to the damage types of the standard samples. For example, for 43, 44, 50, 51, 52, 57, 58 and 59, their damage types may correspond to th number 49, which corresponds to the honeycomb damage type. How between the hole and sag damage states, and is far away from neuron sponding to other unknown damage types. In order to clearly presen corresponding to each neuron, a corresponding relationship table be types and sample classification numbers is constructed as shown in Ta   Table 3. Correspondence between damage types and samples.  Based on the analysis of the traditional comparison method, among the 64 input neurons of the damage identification network model, the damage types of 61 neurons correspond to the damage types of the standard samples. For example, for neurons 36,41,42,43,44,50,51,52,57, 58 and 59, their damage types may correspond to the winning neuron number 49, which corresponds to the honeycomb damage type. However, neuron 37 is between the hole and sag damage states, and is far away from neurons 53 and 60, corresponding to other unknown damage types. In order to clearly present the damage type corresponding to each neuron, a corresponding relationship table between the damage types and sample classification numbers is constructed as shown in Table 3. By analyzing Table 2, it can be seen that the model can obtain the damage classification of almost all neurons, and the recognition accuracy rate is as high as 95.31%. However, there are still cases where unknown neurons cannot be associated with their type of injury. In order to further improve the recognition accuracy of the network model, the TOP-G algorithm is used to determine the type of unknown neuron damage. The analysis process is shown in Figure 10.  First, determine the gray level of the topology map, L = 161, and the gray value range is [0, 161], then the gray level of the topology map image is g = 2 [8] ; Then, at this grayscale, the topological distance color map is converted to a grayscale image; Create a sliding window, mark the polygon as L 1 -L 161 , assign the gray value g 1 -g 161 ; Finally, the gray values are sorted from large to small, and the judgment is made according to the sorting result. Neurons 53 and 60 are connected by a neighborhood polygon number 147, with a grayscale value of 205, which can be seen as a class. The gray values of the eight neighborhood polygons are sorted by gray value, and the neighborhood polygon No.124 has the largest gray value, which connects the neuron number 45 and the neuron number 53. Therefore, it is determined that neurons 53 and 60 correspond to crack damage. Neuron 37 corresponds to six neighborhood polygons, and their gray levels are sorted. The gray value of the polygon in the neighborhood of No. 80 is the largest, which is connected to the neuron of No. 24, corresponding to sag damage. Therefore, it was judged that neuron No. 37 corresponds to sag damage. Therefore, all unknown neuron damage categories are determined based on the topology grayscale algorithm, which further improves the network identification accuracy. Compared with the traditional SOM model, the identification accuracy was improved by 4.69%.

Testing of Improved Algorithm Models
In order to further verify the detection effect of the improved SOM neural network model, the classification results of the detection samples were obtained. The classification labels and sample numbers of winning neurons corresponding to honeycombs, holes, sags and cracks are shown in Table 4. Figure 11 shows the classification results of the detected samples. are sorted from large to small, and the judgment is made according to th Neurons 53 and 60 are connected by a neighborhood polygon number 14 scale value of 205, which can be seen as a class. The gray values of the eigh polygons are sorted by gray value, and the neighborhood polygon No.124 gray value, which connects the neuron number 45 and the neuron numbe it is determined that neurons 53 and 60 correspond to crack damage. N sponds to six neighborhood polygons, and their gray levels are sorted. Th the polygon in the neighborhood of No. 80 is the largest, which is connecte of No. 24, corresponding to sag damage. Therefore, it was judged that neu responds to sag damage. Therefore, all unknown neuron damage categ mined based on the topology grayscale algorithm, which further improv identification accuracy. Compared with the traditional SOM model, the id curacy was improved by 4.69%.

Testing of Improved Algorithm Models
In order to further verify the detection effect of the improved SOM model, the classification results of the detection samples were obtained. Th labels and sample numbers of winning neurons corresponding to honeycom and cracks are shown in Table 4. Figure 11 shows the classification results samples.  According to the analysis of Table 3 and Figure 11, the winning neur ing to the test sample is consistent with the actual damage category, and According to the analysis of Table 3 and Figure 11, the winning neuron corresponding to the test sample is consistent with the actual damage category, and the test sample corresponds to the actual sample type. The damage type corresponding to each sample can be detected based on the improved SOM neural network model.

Discussion
The core content of this paper is the improved of SOM algorithm model in structural damage identification. The ultimate goal of structural health monitoring research is to detect damage as early as possible in order to provide appropriate measures to avoid disaster. It is worth noting that the research object of this paper is micro-damage, and the size of the damage is usually less than 1 mm, which mainly depends on the accuracy of the image acquisition device (laser ranging sensor).
Therefore, the significance of this research is not limited to providing an improved SOM neural network model with a higher recognition accuracy based on a small number of samples. Research can help to effectively identify and even detect and locate damage information in the budding stage of damage, which is of great significance for the timely detection of early structural damage.

Conclusions
Taking the four core steps of constructing the SOM concrete structure damage identification network model as the main line, the network SOM algorithm improvement research is carried out and the following conclusions are obtained:

•
Combined with the self-developed 3D laser scanning system and GLCM theory, the input sample selection method of the SOM network is improved; • Based on the principle of the network topology map analysis and its image characteristics, the concept of the topology grayscale map and the TOP-G algorithm method, and process for the SOM topology map analysis are proposed for the first time; • Based on the active powder concrete bending fatigue loading test, the damage (cracks, sags, honeycombs and holes) identification research of the improved SOM algorithm model was carried out.