An aircraft structure state recognition method based on a self-organizing neural network

Abstract

The application provides an airplane structure state recognition method based on a self-organizing neural network, and belongs to the technical field of airplane structure test data analysis, and comprises the following steps: collecting test data to construct an input vector, obtaining a metric space based on all input types; determining basic parameters of a self-organizing neural network model; determining initial values of clustering center vectors of the self-organizing neural network model; normalizing the input vector and the clustering center vectors; training the self-organizing neural network model based on the normalized input vector and the clustering center vectors, determining a winning clustering center according to a preset winning criterion, adjusting the clustering center vector corresponding to the winning clustering center, and obtaining the self-organizing neural network model when the change amount of the clustering center vector corresponding to the winning clustering center tends to be stable; and performing state recognition on an input vector of an unknown airplane structure state category based on the trained self-organizing neural network model, and obtaining a deformation state of the airplane structure.

Description

Technical Field

[0001] This invention belongs to the field of aircraft structural test data analysis technology, and specifically relates to an aircraft structural state identification method based on a self-organizing neural network (SOM). Background Technology

[0002] In static testing of aircraft, the possible structural changes of an aircraft can be categorized into six types: elastic deformation, linear instability, nonlinear structural instability, material nonlinear instability, structural cracking, and macroscopic structural failure. Elastic deformation is characterized by a linear relationship between strain and load. The structure returns to its original state after the load is removed, with no residual strain locally and no residual deformation overall. Under constrained loads, the aircraft structure primarily exhibits elastic deformation. Linear instability is characterized by a linear relationship between strain or deformation and load in the initial stage. When the load reaches a critical value, the structure suddenly becomes unstable, but the strain value has not reached the material's yield strength. The structure returns to its original state after the load is removed, with no residual deformation. Linear instability typically occurs in areas such as the skin. Nonlinear instability is characterized by a linear relationship between strain or deformation and load in the initial stage. When the load reaches a critical value, the structure becomes unstable, and the strain value has not reached the material's yield strength. However, the structure cannot return to its original state after the load is removed; a lateral load must be applied to restore it to its original state. After restoration, there is no residual deformation within the structure. The characteristics of nonlinear instability in materials are: strain or deformation has a linear relationship with load in the initial stage; when the load reaches a critical value, the structure becomes unstable, and at the same time, local areas within the structure enter plasticity, i.e., plastic deformation occurs. Common forms of structural instability include... Figure 1 As shown, structural cracks typically first appear in high-stress areas such as rivets, notches, and the R-zone. Once a crack appears, the stress measurements near the crack exhibit nonlinear characteristics. After macroscopic failure of the structure, the entire structure or parts of it lose their load-bearing capacity, and the stress field in the surrounding area undergoes drastic changes, often jumping from a maximum value to near zero.

[0003] The existing technologies for data processing in static tests of aircraft structures generally include the following two methods:

[0004] The first method involves forming an experience pattern by reviewing a large amount of data on the characteristics of the above six morphological changes, and then manually analyzing each piece of test data to identify the aircraft's structural state. However, in the face of massive amounts of test data, manual analysis is inefficient and lacks timeliness, which is far from meeting the requirements for processing test data.

[0005] The second method is to identify the structural state by analyzing test data based on empirical patterns formed from reviewed data and writing fixed judgment programs. Although this method is efficient, the program's algorithm is fixed and lacks the ability to learn and evolve, making it difficult to adapt to the complex and ever-changing response characteristics of aircraft structures. Summary of the Invention

[0006] The purpose of this invention is to provide an aircraft structural state recognition method based on a self-organizing neural network to solve or mitigate at least one of the problems in the background art.

[0007] The technical solution of this invention is: a method for aircraft structural state recognition based on a self-organizing neural network, comprising:

[0008] Step S1: Collect experimental data to construct an input vector X for training a self-organizing neural network model, and obtain a metric space based on all input types;

[0009] Step S2: Determine the basic parameters of the self-organizing neural network model;

[0010] Step S3: Determine the initial values ​​of the cluster center vectors for the self-organizing neural network model;

[0011] Step S4: Normalize the input vector and the cluster center vector;

[0012] Step S5: Train the self-organizing neural network model based on the normalized input vector and cluster center vector. Determine the winning cluster center according to the preset winning criterion and adjust the cluster center vector corresponding to the winning cluster center. When the change of the cluster center vector corresponding to the winning cluster center tends to stabilize, the self-organizing neural network model is obtained.

[0013] Step S6: Based on the trained self-organizing neural network model, perform state recognition on the input vector of the unknown aircraft structure state category to obtain the deformation state of the aircraft structure.

[0014] In at least one embodiment of the present invention, in step S3, the initial value of the cluster center vector is selected based on experience as a vector that matches the data type corresponding to the aircraft structural morphology, or is given randomly.

[0015] In at least one embodiment of the present invention, the method for normalizing the input vector and the cluster center vector in step S4 is as follows:

[0016] (i=1,2,...,n)

[0017] In the formula, X̄ is the normalized input vector, ‖X‖ is the magnitude of the input vector, X is the input vector, and x is the input vector. iLet i be the i-th element in the input vector;

[0018] (j=1,2,…,m)

[0019] In the formula, Ōⱼ is the normalized cluster center vector, ‖X‖ is the magnitude of the cluster center vector, and O j Let o be the cluster center vector. j It is the j-th element in the cluster center vector.

[0020] In at least one embodiment of the present invention, in step S5, the preset winning criteria include:

[0021] Using the minimum Euclidean distance as the criterion: ‖X̄ - Ōⱼ‖ = min{ ‖X̄ - Ōⱼ‖}, (j=1,2,...,m);

[0022] The criterion is to maximize the dot product: X̄ᵀ · Ōⱼ = max{ X̄ᵀ · Ōⱼ}, (j=1,2,...,m).

[0023] In at least one embodiment of the present invention, the method for adjusting the cluster center vector corresponding to the winning cluster center in step S5 is as follows:

[0024] Oⱼ(k+1) = Oⱼ(k) + α · ( X̄ - Oⱼ(k) );

[0025] In the formula, Oⱼ(k) is the cluster center vector before the k-th iteration; Oⱼ(k+1) is the cluster center vector after the k-th iteration; and α is the learning rate.

[0026] In at least one embodiment of the present invention, the learning rate α is a decreasing function of the number of training iterations, which gradually decreases as the number of training iterations increases.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] 1) By utilizing the self-organizing and adaptive characteristics of self-organizing neural networks, it is possible to automatically learn and accumulate experience from massive experimental data. The network parameters can be continuously evolved and optimized with training, overcoming the defect that fixed algorithm programs cannot adapt to complex data characteristics.

[0029] 2) By constructing a quantifiable and computable data structure that reflects the characteristics of the classification pattern as input to the network model, the network model can accurately capture the mapping relationship between the aircraft structural response and the load, thereby improving the accuracy of state recognition.

[0030] 3) The normalization process eliminates the influence of different dimensions on the classification results, while supporting both Euclidean distance and dot product as the winning criteria, thus enhancing the applicability and robustness of the method.

[0031] 4) Cluster centers can be initialized based on experience, which significantly improves the network convergence speed; at the same time, random initialization is supported to meet the needs of different application scenarios.

[0032] 5) The method of the present invention can replace manual analysis work, greatly improve the processing efficiency and timeliness of aircraft static test data, and meet the needs of rapid analysis of massive data in practical engineering applications. Attached Figure Description

[0033] To more clearly illustrate the technical solution provided by the present invention, the accompanying drawings will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention.

[0034] Figure 1 This is a schematic diagram of the instability of the aircraft panel structure.

[0035] Figure 2 This is a schematic diagram of the aircraft structural state recognition method based on self-organizing neural networks in this invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions in the embodiments of this invention will be described in more detail below with reference to the accompanying drawings.

[0037] To address the problem of analyzing aircraft static test data in existing technologies, this invention proposes an evolutionary data analysis and classification method using a self-organizing neural network to perform pattern recognition on the response of aircraft structures and loads, thereby replacing manual analysis and improving the efficiency of data analysis.

[0038] like Figure 2 As shown, the aircraft structural state recognition method based on self-organizing neural networks provided by this invention includes the following process:

[0039] Step S1: Collect experimental data to construct input vectors for training the self-organizing neural network model, and obtain the metric space based on all input vectors.

[0040] During static testing of an aircraft, strain sensors installed on the aircraft structure collect strain data during static loading to form a raw dataset. The raw dataset undergoes preprocessing, including noise removal and outlier handling, to construct a quantifiable data structure reflecting clustering pattern characteristics, i.e., the input vector X = (x1, x2, ..., x...). n), where n is the dimension of the input vector (i.e., the number of features). All input vectors X constitute a metric space, and each input vector X is an element in this metric space.

[0041] In this invention, the construction of the data structure of the metric space follows the principles of minimizing the amount of feature data, clearly describing the features, and ensuring uniqueness.

[0042] Step S2: Determine the basic parameters of the self-organizing neural network model.

[0043] The basic parameters of the self-organizing neural network model are determined based on the number of elements in the metric space, data structure characteristics, and the types of cluster centers. The self-organizing neural network model consists of two layers: an input layer and an output layer. The dimension n of the input layer vector equals the dimension of the input samples, and the dimension m of the output layer equals the number of clustering results, i.e., the number of possible state categories of the aircraft structure, including six categories: elastic deformation, linear instability, structural nonlinear instability, material nonlinear instability, structural cracks, and macroscopic structural failure. In this embodiment, the number of cluster center types is set to 6, corresponding to the six structural states of elastic deformation, linear instability, structural nonlinear instability, material nonlinear instability, structural cracks, and macroscopic structural failure, i.e., the output layer dimension m=6.

[0044] Step S3: Determine the initial value of the cluster center vector of the self-organizing neural network model.

[0045] The vector corresponding to the initial cluster centers is denoted as Oⱼ 0 (j=1,2,...,m), whose dimension is the same as the input vector X. The initial value of the cluster center vector can be selected empirically to match the data types corresponding to the six aircraft structural morphologies, in order to improve the iteration convergence speed. Specifically, six typical samples can be selected from known types of experimental data as the initial cluster center vector Oⱼ. 0 (j=1,2,...,6); can also be given randomly. The final result of the cluster center vector is determined through iterative learning.

[0046] Step S4: Normalize the input vector and the cluster center vector.

[0047] To eliminate the influence of different units on the training of neural network models, the input vector and cluster center vector need to be normalized. The normalization formula is as follows:

[0048] (i=1,2,...,n)

[0049] In the formula, X̄ is the normalized input vector, ‖X‖ is the magnitude of the input vector, and x... i Let i be the i-th element in the input vector.

[0050] Similarly, (j=1,2,…,m)

[0051] In the formula, Ōⱼ is the normalized cluster center vector, ‖X‖ is the magnitude of the cluster center vector, and o j It is the j-th element in the cluster center vector.

[0052] Step S5: Train the self-organizing neural network model based on the normalized input vector and the cluster center vector. When the change in the cluster center vector tends to stabilize, the self-organizing neural network model is obtained.

[0053] The input vector X of the known aircraft structural state category is fed into the self-organizing neural network model for training one by one. For each input vector X, the distance (or similarity) between it and each cluster center vector is calculated to determine the winning cluster center.

[0054] In this invention, there are two methods for determining the winning cluster center:

[0055] The first method uses the minimum Euclidean distance as the winning criterion, i.e.: ‖X̄ - Ōⱼ‖ = min{ ‖X̄ - Ōⱼ‖}, (j=1,2,...,m);

[0056] The second method uses the maximum dot product as the winning criterion, i.e.: X̄ᵀ · Ōⱼ = max{ X̄ᵀ · Ōⱼ}, (j=1,2,...,m);

[0057] For each input vector sample, the cluster center vector corresponding to the winning cluster center is adjusted once. The adjustment method is: Oⱼ(k+1) = Oⱼ(k) + α · ( X̄ - Oⱼ(k) );

[0058] In the formula, Oⱼ(k) is the cluster center vector before the k-th iteration; Oⱼ(k+1) is the cluster center vector after the k-th iteration; α is the learning rate, 0<α≤1, and the learning rate is generally a decreasing function of the number of training iterations, that is, it gradually decreases as the number of training iterations increases; k is the current iteration number.

[0059] By repeating the above training process until all cluster centers stabilize, training stops, thus obtaining a well-trained self-organizing neural network model.

[0060] For example, in this embodiment, the initial learning rate α is set to 0.5. As the number of training iterations k increases, the learning rate α decreases according to the rule α(k) = 0.5 / (1+k / 100). For each input sample, only the cluster center vector corresponding to the cluster center of the winning neuron is adjusted, while the remaining cluster centers remain unchanged. Iterative training continues until the change in all cluster center vectors is less than a preset threshold (e.g., 10⁻⁻⁴).4 When the cluster centers are determined to be stable, training stops.

[0061] Step S6: Based on the trained self-organizing neural network model, perform state recognition on the input vector of the unknown aircraft structure state category to obtain the deformation state of the aircraft structure.

[0062] The strain data collected under the unknown aircraft structural state category under the test condition are used to construct an input vector according to the method in step S1. After normalizing the constructed input vector according to the process in step S4, it is input into the trained self-organizing neural network model. The trained self-organizing neural network model can identify the morphological change type of the aircraft structure under the current loading condition, thereby completing the analysis of test data and automatic identification of the aircraft structural state.

[0063] The distance between the input vector and each cluster center is calculated, and the vector is classified into the structural state type corresponding to the nearest cluster center, thereby completing the automatic identification of the aircraft structural morphology change type.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] 1) By utilizing the self-organizing and adaptive characteristics of self-organizing neural networks, it is possible to automatically learn and accumulate experience from massive experimental data. The network parameters can be continuously evolved and optimized with training, overcoming the defect that fixed algorithm programs cannot adapt to complex data characteristics.

[0066] 2) By constructing a quantifiable and computable data structure that reflects the characteristics of the classification pattern as input to the network model, the network model can accurately capture the mapping relationship between the aircraft structural response and the load, thereby improving the accuracy of state recognition.

[0067] 3) The normalization process eliminates the influence of different dimensions on the classification results, while supporting both Euclidean distance and dot product as the winning criteria, thus enhancing the applicability and robustness of the method.

[0068] 4) Cluster centers can be initialized based on experience, which significantly improves the network convergence speed; at the same time, random initialization is supported to meet the needs of different application scenarios.

[0069] 5) The method of the present invention can replace manual analysis work, greatly improve the processing efficiency and timeliness of aircraft static test data, and meet the needs of rapid analysis of massive data in practical engineering applications.

[0070] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for aircraft structural state recognition based on a self-organizing neural network, characterized in that, include: Step S1: Collect experimental data to construct an input vector X for training a self-organizing neural network model, and obtain a metric space based on all input types; Step S2: Determine the basic parameters of the self-organizing neural network model; Step S3: Determine the initial values ​​of the cluster center vectors for the self-organizing neural network model; Step S4: Normalize the input vector and the cluster center vector; Step S5: Train the self-organizing neural network model based on the normalized input vector and cluster center vector. Determine the winning cluster center according to the preset winning criterion and adjust the cluster center vector corresponding to the winning cluster center. When the change of the cluster center vector corresponding to the winning cluster center tends to stabilize, the self-organizing neural network model is obtained. Step S6: Based on the trained self-organizing neural network model, perform state recognition on the input vector of the unknown aircraft structure state category to obtain the deformation state of the aircraft structure.

2. The aircraft structural state recognition method based on self-organizing neural networks as described in claim 1, characterized in that, In step S3, the initial value of the cluster center vector is selected based on experience and is a vector that matches the data type corresponding to the aircraft structure, or it is given randomly.

3. The aircraft structural state recognition method based on self-organizing neural networks as described in claim 1 or 2, characterized in that, In step S4, the input vector and cluster center vector are normalized as follows: ,(i=1,2,...,n) In the formula, X̄ is the normalized input vector, ‖X‖ is the magnitude of the input vector, X is the input vector, and x is the input vector. i Let i be the i-th element in the input vector; ,(j=1,2,…,m) In the formula, Ōⱼ is the normalized cluster center vector, ‖X‖ is the magnitude of the cluster center vector, and O j Let o be the cluster center vector. j It is the j-th element in the cluster center vector.

4. The aircraft structural state recognition method based on self-organizing neural networks as described in claim 3, characterized in that, In step S5, the preset winning criteria include: Using the minimum Euclidean distance as the criterion: ‖X̄ - Ōⱼ‖ = min{‖X̄ - Ōⱼ‖}, (j=1,2,...,m); The criterion is to maximize the dot product: X̄ᵀ · Ōⱼ = max{ X̄ᵀ · Ōⱼ}, (j=1,2,...,m).

5. The aircraft structural state recognition method based on self-organizing neural networks as described in claim 4, characterized in that, In step S5, the method for adjusting the cluster center vector corresponding to the winning cluster center is as follows: Oⱼ(k+1) = Oⱼ(k) + α · ( X̄ - Oⱼ(k) ); In the formula, Oⱼ(k) is the cluster center vector before the k-th iteration; Oⱼ(k+1) is the cluster center vector after the k-th iteration; and α is the learning rate.

6. The aircraft structural state recognition method based on self-organizing neural networks as described in claim 5, characterized in that, The learning rate α is a decreasing function of the number of training iterations, gradually decreasing as the number of training iterations increases.

Images