Concrete structure damage identification method based on multi-source data fusion

Abstract

The invention relates to a concrete structure damage identification method based on multi-source data fusion. The method comprises the following steps: respectively obtaining damage prediction results corresponding to different data sources; and based on a Dempster rule, performing decision-level data fusion on different damage prediction results to obtain a damage identification result. Compared with the prior art, the method has the advantages that the D-S evidence theory is used, decision-level data fusion is carried out on data information of different sources, uncertain parts and parts possibly having conflicts in the information are processed, so that a more excellent and stable damage identification effect is achieved, the problem of conflicts among different data sources can be solved, and the accuracy of damage identification is improved. And the damage identification efficiency and accuracy are improved.

Description

Technical Field

[0001] The invention relates to the technical field of building structure state detection, and in particular to a concrete structure damage identification method based on multi-source data fusion. Background Art

[0002] Due to material aging, load effects, etc., reinforced concrete structures may suffer various types of damage and material performance degradation during service, which will lead to a decrease in the structural bearing capacity and use function. In some cases, the internal steel bars of the structure will be exposed and rusted, and in severe cases, the structure may even fail completely or even collapse. Therefore, it is necessary to conduct damage detection on concrete structures in order to obtain the performance status of concrete structures in a timely manner.

[0003] Traditional concrete structure damage detection is mainly based on visual inspection (observing directly visible damage such as concrete cracks, alkali, spalling, exposed steel bars, etc.), structural displacement measurement, concrete strength detection based on concrete rebound tester, reinforcement detection based on steel bar positioner, structural tilt detection, remodeling and bearing capacity verification, etc., but these methods are time-consuming and labor-intensive, and their accuracy is questionable. In recent years, the application of computer vision technology and image processing algorithms based on deep learning has attracted more and more attention from scholars. Its ability to extract, measure and quantify the size of cracks on the surface of concrete structures makes it helpful to determine the damage status of buildings, but this method mostly relies on a single damage identification data source, which is difficult to guarantee the accuracy and reliability of detection and identification. If considering the fusion analysis of multiple different damage identification data sources, there are problems in dealing with uncertainty and conflict. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a concrete structure damage identification method based on multi-source data fusion, which can solve the conflict problem between different data sources and improve the efficiency and accuracy of damage identification.

[0005] The purpose of the present invention can be achieved by the following technical solution: A method for concrete structure damage identification based on multi-source data fusion, comprising the following steps:

[0006] S1. Obtain damage prediction results corresponding to different data sources respectively;

[0007] S2. Based on Dempster's law, different damage prediction results are fused at the decision level to obtain damage identification results.

[0008] Furthermore, the different data sources in step S1 include image information, sound information and acceleration information;

[0009] Furthermore, the damage prediction result in step S1 includes no damage, slight damage, moderate damage or severe damage.

[0010] Furthermore, the step S2 comprises the following steps:

[0011] S21, predetermining basic confidence allocation values ​​corresponding to damage predictions from different data sources;

[0012] S22. Dempster's synthesis rule is used to solve and calculate the basic confidence distribution value after fusion, that is, the damage identification result is obtained.

[0013] Furthermore, in step S21, if the data source is image information, a basic reliability allocation value corresponding to a no-damage prediction result, a basic reliability allocation value corresponding to a slight-damage prediction result, a basic reliability allocation value corresponding to a moderate-damage prediction result, and a basic reliability allocation value corresponding to a severe-damage prediction result are defined respectively.

[0014] Furthermore, in step S21, if the data source is sound information, a basic confidence allocation value corresponding to a no-damage prediction result, a basic confidence allocation value corresponding to a slight-damage prediction result, a basic confidence allocation value corresponding to a moderate-damage prediction result, and a basic confidence allocation value corresponding to a severe-damage prediction result are defined respectively.

[0015] Furthermore, in step S21, if the data source is acceleration information, a basic confidence allocation value corresponding to a no-damage prediction result, a basic confidence allocation value corresponding to a slight-damage prediction result, a basic confidence allocation value corresponding to a moderate-damage prediction result, and a basic confidence allocation value corresponding to a severe-damage prediction result are defined respectively.

[0016] Furthermore, the step S22 specifically includes:

[0017] S221, calculating a normalization coefficient;

[0018] S222. Use the normalization coefficient to solve the basic confidence allocation value corresponding to each fused damage prediction result, where the damage prediction result with the highest basic confidence allocation value after fusion is the damage identification result.

[0019] Furthermore, in step S221, since any two damage states cannot occur at the same time, the intersections between them are all empty sets, and the normalization coefficient is obtained as:

[0020]

[0021] Among them, K3 is the normalization coefficient, A1, A2, and A3 are different damage states, I represents the prediction result of no damage, II represents the prediction result of slight damage, III represents the prediction result of moderate damage, and IV represents the prediction result of severe damage. m1 represents the basic reliability allocation value when the data source is image information, m2 represents the basic reliability allocation value when the data source is sound information, and m3 represents the basic reliability allocation value when the data source is acceleration information.

[0022] Furthermore, the basic confidence distribution values ​​corresponding to the fused damage prediction results in step S222 are:

[0023]

[0024] Among them, m 123 (I) is the basic confidence distribution value corresponding to the fused damage-free prediction result, m 123 (II) is the basic confidence distribution value corresponding to the fused minor injury prediction result, m 123 (III) is the basic confidence distribution value corresponding to the fused moderate damage prediction result, m 123 (IV) is the basic confidence distribution value corresponding to the fused severe injury prediction result.

[0025] Compared with the prior art, the present invention has the following advantages:

[0026] Aiming at the conflicting problems existing in various sources of damage identification data, the present invention is based on Dempster's law and combines evidence from different sources through certain rules, so as to achieve the goals of processing multi-source information, processing uncertainty and conflict, flexibility and theoretical completeness. Thus, decision-level data fusion is performed on damage prediction results from different data sources, and the uncertain and possibly conflicting parts of such information are processed, so as to effectively resolve the conflict problems between different data sources and greatly improve the efficiency and accuracy of damage identification.

[0027] When the present invention applies the decision-level data fusion method based on DS evidence theory to different data sources, the basic confidence distribution values ​​corresponding to four possible damage prediction results of a node from three different identification data sources are predetermined, and then the normalization coefficient is calculated and used to solve the basic confidence distribution value after calculation and fusion, which can improve the accuracy of damage judgment and can accurately and efficiently identify and detect the damage status of any node in the concrete structure. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0029] Figure 2 Schematic diagram of the application process of the embodiment. DETAILED DESCRIPTION

[0030] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0031] Example

[0032] To facilitate the understanding of this scheme, the following explanation is first given for the conceptual basis of this scheme:

[0033] In order to further improve the accuracy, efficiency and reliability of concrete structure damage identification, this scheme intends to use data fusion methods to extract more comprehensive damage information based on image-based damage information and damage information based on generalized acoustic features (including sound and acceleration features). Data fusion usually includes three levels of fusion operations: data-level fusion, feature-level fusion and decision-level fusion. Since it is difficult to achieve data-level and feature-level fusion of image information and acoustic features at the first two levels, this scheme uses the third level-decision-level fusion method for research and discussion.

[0034] The main goal of decision-level fusion is to combine information from multiple sources (in this case, acceleration information, sound information, and image information) to make more accurate and effective decisions. Fusion methods at this level include rule-based systems, Dempster-Shafer (DS) evidence theory, Bayesian networks, and other advanced statistical and artificial intelligence methods such as convolutional neural networks. These methods are able to process large amounts of information from a variety of different data sources and make fast and accurate decisions.

[0035] This scheme uses the DS evidence theory to perform decision-level data fusion on information from three different sources, and to process the uncertain and potentially conflicting parts of the information, thereby achieving a better and more stable damage identification effect.

[0036] Among them, DS Evidence Theory, full name Dempster-Shafer Evidence Theory, is also known as Belief Function Theory. It was proposed by Arthur Dempster and Glenn Shafer in the 1970s and is a mathematical framework for understanding and processing uncertain information. DS Evidence Theory mainly consists of two parts: Belief Function and Evidence Fusion.

[0037] In DS evidence theory, after data fusion, all possible results or events (in this solution, the node damage degree, which is still divided into four categories from 1 to 4) form a set, which is called the identification framework:

[0038] θ={θ1,θ2,…,θ n},

[0039] Where n is the total number of hypotheses (n=4 in this scheme). For each subset (each damage state), a confidence value can be assigned, ranging from 0 to 1, indicating the degree of confidence in the subset. These confidence values ​​together constitute a confidence function. The confidence function of proposition A (for example, the damage state of a node is level 2 - slight damage) under the identification framework Θ is defined as follows: For any subset A in the identification framework Θ, the confidence function Bel(A) is equal to the sum of the basic confidence distribution m(B) of all subsets B of A (B is included in A), which can be written in mathematical language as:

[0040]

[0041] Among them, mapping m: 2 Θ →[0,1] satisfies

[0042]

[0043] m(A) indicates the degree of support for proposition A by the evidence currently obtained, also known as the basic credibility distribution. If the basic credibility distribution m(A) of a proposition is greater than 0, it is called the focal element of the credibility function Bel(A). The credibility function Bel expresses the degree of trust in the proposition (considering the cause and effect).

[0044] In addition to the belief function Bel, there is another function that measures the correctness of a proposition, called the plausibility function (Pl). Suppose the mapping Pl:2 Θ →[0,1] satisfies

[0045]

[0046] Then Pl is called the plausibility function, and the plausibility Pl(A) of A represents the degree to which A is not doubted, which includes the basic credibility distribution of all propositions that are compatible with A.

[0047] When there are two or more credibility functions (in this solution, there are multiple sources of data for identifying damage), the Dempster rule can be used to fuse them to obtain a new credibility function, thereby achieving decision-level data fusion. The following briefly introduces the specific application and calculation route of the Dempster rule when there are two or more credibility functions.

[0048] 1. Dempster's Law for the Fusion of Dual Belief Functions

[0049] Assume that there are two recognition functions Bel1 and Bel2 in a recognition framework Θ, which have corresponding basic credibility assignments m1 and m2, respectively, and their focal elements are A1, A2, …, Ai and B1,B2,…,B j ,but

[0050]

[0051] It is the Dempster's law synthesized by two credibility functions, and K represents the degree of conflict between different evidences. If the evidences behind the corresponding propositions of m1 and m2 are completely conflicting, then K is 0 and the synthesis function does not exist; if the evidences behind the corresponding propositions of m1 and m2 completely support each other, then K is 1. They are the upper and lower bounds of K, and in practical applications, K will be between 0 and 1. Similarly, when there are three or more credibility functions, the synthesized Dempster's law can be generalized in a similar form.

[0052] 2. Dempster's Law for Fusion of Multiple Belief Functions

[0053] Assume that there are n belief functions Bel1, Bel2, ..., Bel in a recognition framework Θ n , which have corresponding basic credibility distributions m1,m2,…,m n ,like exists and its basic credibility is assigned to m, and for Both Then there is

[0054]

[0055] in

[0056] It can be seen that the synthesis formula of multiple credibility functions is obtained by nesting and repeating the synthesis formula of dual credibility functions many times, and the synthesis operation It has commutative and associative laws, that is, This new belief function will reflect the common information of all original belief functions, thus providing more comprehensive, reliable and robust results.

[0057] One of the main advantages of DS evidence theory is that it can handle incomplete and conflicting information. Compared with other methods of dealing with uncertainty, such as probability theory or fuzzy set theory, DS evidence theory can provide a richer and more sophisticated way to characterize uncertainty. Therefore, it has been widely used in many fields, including artificial intelligence, machine learning, information fusion, decision analysis, etc. This program will apply DS evidence theory to the fusion of multi-source data of concrete structure damage and explore the concrete structure damage identification results of multi-source data fusion.

[0058] This proposal proposes a concrete structure damage identification method based on multi-source data fusion, such as Figure 1 As shown, the following steps are included:

[0059] S1. Obtain damage prediction results corresponding to different data sources respectively;

[0060] S2. Based on Dempster's law, different damage prediction results are fused at the decision level to obtain damage identification results.

[0061] This embodiment applies the above solution, such as Figure 2 As shown in the figure, based on the DS theory, different credibility functions (recognition methods based on images, sounds, and acceleration) are synthesized into a new credibility function based on the corresponding evidence information through the Dempster law, thereby improving the recognition accuracy and robustness. The specific contents include:

[0062] 1. Define the recognition framework

[0063] When applying the decision-level data fusion method based on DS evidence theory to different data sources, it is necessary to predetermine the basic confidence allocation values ​​m corresponding to the four possible damage conditions of a node for three different identification methods. i (). In this embodiment, the average error of semantic segmentation in identifying cracks is 10.45%. According to the definition of damage level, there are no visible cracks at nodes with no damage (I), the crack width at nodes with slight damage (II) is less than 1mm, the crack width at nodes with moderate damage (III) is greater than 1mm, and the crack width at nodes with severe damage (IV) is greater than 5mm due to obvious peeling or exposed steel bars. If the average crack width of a node (data source 1) falls within damage level A, the basic reliability allocation value corresponding to level A can be set to 0.8955, and the basic reliability allocation values ​​of the other three levels are all (1-0.8955) / 3=0.035. For example, after semantic segmentation of cracks at a node and calculation of the average crack width, the result is 1.2 mm, which means it is a moderate damage, m1(II) = 0.8955, m1(I) = m1(III) = m1(IV) = 0.035; for the damage classification result based on sound information (data source 2), according to the corresponding confusion matrix and the prediction accuracy obtained by the test, the corresponding basic confidence allocation value can be set as m2(III) = 0.9805, m2(I) = m2(II) = m2(IV) = 0.0065. Similarly, for the damage classification result based on acceleration information (data source 3), the corresponding basic confidence allocation value is m3(III) = 0.9472, m3(I) = m3(II) = m3(IV) = 0.0176.

[0064] In summary, for a damaged node S, if the damage results predicted by the data from images, sounds, and acceleration are slight damage, moderate damage, and severe damage, respectively, the following basic credibility distribution value table can be constructed based on the above assumptions:

[0065] Table 1 Basic credit distribution table for a damaged node S

[0066] <![CDATA[m1()]]> <![CDATA[m2()]]> <![CDATA[m3()]]> No damage (I) 0.035 0.0065 0.0176 Minor injury (II) 0.8955 0.0065 0.0176 Moderate injury (III) 0.035 0.9805 0.9472 Severe injury (IV) 0.035 0.0065 0.0176

[0067] Then its recognition framework is Θ = {I, II, III, IV}; where subscript 1 represents the data from the image, subscript 2 represents the sound data, and subscript 3 represents the acceleration data. According to the Dempster law theory of multi-credibility function fusion and the corresponding calculation formula, the basic credibility distribution value m of the damage state that integrates the three types of data can be calculated. 123 .

[0068] 2. Use Dempster's synthesis rule to find the fusion reliability function

[0069] First, we need to find the normalization coefficient K3. Since any two damage states in the identification framework cannot occur at the same time, the intersection between them is an empty set. K3 can be found as follows:

[0070]

[0071] Using K3, we can find the basic credibility distribution value after fusion; since any two elements in the recognition framework do not intersect, the only case where the intersection of three elements is element A is when all three elements are A. We can get:

[0072]

[0073] Finally, the above calculation process obtains the basic credibility distribution value m after fusion 123 (), that is, the prediction result for this node, shows that the basic confidence distribution value of damage state III (moderate damage) is the highest, which is 0.8454. Therefore, for this node, the decision-level data fusion prediction result is damage state III. Combining the above data with Table 1, the following table can be obtained:

[0074] Table 2 Basic credibility distribution table of the damage of a node S after data fusion

[0075] <![CDATA[m1()]]> <![CDATA[m2()]]> <![CDATA[m3()]]> <![CDATA[m 123 ()]]> No damage (I) 0.035 0.0065 0.0176 0.005604 Minor injury (II) 0.8955 0.0065 0.0176 0.1433 Moderate injury (III) 0.035 0.9805 0.9472 0.8454 Severe injury (IV) 0.035 0.0065 0.0176 0.005604

[0076] By adopting this scheme, the damage status of any node can realize decision-level data fusion through the above path, thereby improving the accuracy of damage judgment; when there is a conflict between the three classification methods, the data fusion method can make the best decision and avoid erroneous classification results as much as possible.

[0077] The above content takes a certain node as an example to explore the whole process of decision-level data fusion based on Dempster's law and show the final fusion result. The effectiveness of this solution has not been verified. This embodiment performs data fusion operations on more nodes and compares the final recognition accuracy to objectively present the effect of data fusion.

[0078] When comparing the recognition accuracy, the 7 nodes of the concrete structure before reinforcement were processed by data fusion based on DS evidence theory and three sources of prediction data, and their corresponding damage states were determined. The results are shown in Table 3. The table only contains nodes with different results for different recognition methods, because it is meaningless to perform decision-level data fusion on nodes with exactly the same recognition results for the three recognition methods. It should be noted that since the judgment of the node damage degree by acoustic features is based on the signal generated by a single knock, the prediction results based on acoustic features may be different for the same node (due to the instability of a single knock or the influence of external noise), resulting in the same node appearing multiple times in the table.

[0079] Table 3 Comparison of damage state prediction results based on data fusion based on DS evidence theory

[0080]

[0081] As can be seen from Table 3, at these "ambiguous" nodes, the accuracy of data fusion reached 100%, which proves the effectiveness of decision-level data fusion and solves the problem of conflicting evidence when having different data sources.

[0082] In summary, this scheme aims at the conflicting problems existing in various sources of damage identification data, and studies the decision-level data fusion method based on the Dempster-Shafer evidence theory. It merges evidence from different sources through certain rules, and can achieve the goals of processing multi-source information, handling uncertainty and conflict, flexibility and theoretical completeness. These characteristics enable the fusion method based on the DS evidence rule to achieve a high accuracy rate on the basis of being able to solve the conflict problem between evidence, which is conducive to improving the efficiency and accuracy of concrete structure damage identification.

Claims

1. A concrete structure damage identification method based on multi-source data fusion, characterized in that: The following steps are involved: S1. Obtain damage prediction results corresponding to different data sources respectively; S2. Based on Dempster's law, different damage prediction results are fused at the decision level to obtain damage identification results.

2. The method for concrete structure damage identification based on multi-source data fusion according to claim 1 is characterized in that: The different data sources in step S1 include image information, sound information and acceleration information.

3. The method for concrete structure damage identification based on multi-source data fusion according to claim 2 is characterized in that: The damage prediction result in step S1 includes no damage, slight damage, moderate damage or severe damage.

4. The method for concrete structure damage identification based on multi-source data fusion according to claim 3 is characterized in that: The step S2 comprises the following steps: S21, predetermining basic confidence allocation values ​​corresponding to damage predictions from different data sources; S22. Dempster's synthesis rule is used to solve and calculate the basic confidence distribution value after fusion, that is, the damage identification result is obtained.

5. The method for concrete structure damage identification based on multi-source data fusion according to claim 3 is characterized in that: In step S21, if the data source is image information, the basic reliability allocation value corresponding to the no-damage prediction result, the basic reliability allocation value corresponding to the slight-damage prediction result, the basic reliability allocation value corresponding to the moderate-damage prediction result, and the basic reliability allocation value corresponding to the severe-damage prediction result are defined respectively.

6. The method for concrete structure damage identification based on multi-source data fusion according to claim 3 is characterized in that: In step S21, if the data source is sound information, the basic confidence allocation value corresponding to the no-damage prediction result, the basic confidence allocation value corresponding to the slight-damage prediction result, the basic confidence allocation value corresponding to the moderate-damage prediction result, and the basic confidence allocation value corresponding to the severe-damage prediction result are defined respectively.

7. The method for concrete structure damage identification based on multi-source data fusion according to claim 3 is characterized in that: In step S21, if the data source is acceleration information, the basic confidence allocation value corresponding to the no-damage prediction result, the basic confidence allocation value corresponding to the slight-damage prediction result, the basic confidence allocation value corresponding to the moderate-damage prediction result, and the basic confidence allocation value corresponding to the severe-damage prediction result are defined respectively.

8. The method for concrete structure damage identification based on multi-source data fusion according to claim 3 is characterized in that: The step S22 specifically includes: S221, calculating a normalization coefficient; S222. Use the normalization coefficient to solve the basic confidence allocation value corresponding to each fused damage prediction result, where the damage prediction result with the highest basic confidence allocation value after fusion is the damage identification result.

9. The method for concrete structure damage identification based on multi-source data fusion according to claim 8 is characterized in that: In step S221, since any two damage states cannot occur at the same time, the intersections between them are all empty sets, and the normalization coefficient is calculated as: Among them, K3 is the normalization coefficient, A1, A2, and A3 are different damage states, I represents the prediction result of no damage, II represents the prediction result of slight damage, III represents the prediction result of moderate damage, and IV represents the prediction result of severe damage. m1 represents the basic reliability allocation value when the data source is image information, m2 represents the basic reliability allocation value when the data source is sound information, and m3 represents the basic reliability allocation value when the data source is acceleration information.

10. The method for concrete structure damage identification based on multi-source data fusion according to claim 9 is characterized in that: The basic confidence distribution value corresponding to each fused damage prediction result in step S222 is: Among them, m 123 (I) is the basic confidence distribution value corresponding to the fused damage-free prediction result, m 123 (II) is the basic confidence distribution value corresponding to the fused minor injury prediction result, m 123 (III) is the basic confidence distribution value corresponding to the fused moderate damage prediction result, m 123 (IV) is the basic confidence distribution value corresponding to the fused severe injury prediction result.

Images