Mechanical structure crack identification method based on boundary element field variable data driving

Abstract

The invention discloses a mechanical structure crack identification method based on boundary element field variable data driving, which comprises the following steps of: firstly, establishing a forward analysis model of crack analysis by utilizing a boundary element numerical analysis technology, and calculating physical field variables of mechanical structures in different types, different crack parameters and different positions; performing post-processing to obtain a related numerical analysis field variable cloud picture; then arranging related sensors in different areas to collect structure images and physical quantity images of the areas, combining numerical analysis image data with image data collected by the sensors, performing image processing through technologies such as data fusion, normalization and data enhancement, and inputting the processed data into a crack identification hybrid network deep learning model; according to the method, a Fisher vector based on digital-analog driving and a deep convolutional neural network fusion crack identification model are established, crack parameters in a mechanical structure are accurately identified, an accurate and reliable prediction method is provided for crack identification and crack fault evolution of the mechanical structure, and technical support is provided for intelligent detection of mechanical equipment.

Description

Technical Field

[0001] The present invention relates to the field of mechanical structure crack damage identification, and in particular to a mechanical structure crack identification method based on boundary element field variable data drive, belonging to the field of mechanical structure fault diagnosis. Background Art

[0002] With the acceleration of industrialization, a large number of mechanical structures and infrastructure, such as bridges, tunnels, pipelines, etc., have been in service for a long time in high-intensity and harsh environments, and are prone to fatigue damage and microcracks. If these microcracks are not identified and handled in time, they may lead to structural failure, accidents, and even serious economic losses and casualties. Therefore, timely identification and assessment of cracks can not only effectively prevent potential accidents, but also implement targeted maintenance and repair, thereby extending the service life of equipment, reducing replacement costs, and ensuring the safe operation of equipment. It has important theoretical research value and extensive practical application value.

[0003] Existing crack identification technologies mainly include methods based on manual identification, model-driven, and sensor data-driven. Among them, the method based on manual identification requires engineers to have rich experience, the operation process is intuitive and easy to understand, and engineers can make flexible judgments based on on-site conditions; however, this method relies on the professional skills and experience of technicians, and inexperienced personnel may find it difficult to accurately identify cracks, which in turn affects the identification results. The model-driven method mainly uses numerical models to inversely identify cracks, which can provide higher accuracy, especially in the prediction of complex structures and crack morphology. It can accurately simulate the expansion behavior and stress field of cracks, and can also flexibly adjust the model according to different materials, geometric shapes and loading conditions. It is suitable for a variety of application scenarios, but the identification results are easily affected by the calculation accuracy of the numerical model. The sensor data-driven method mainly uses the sensor data (such as strain, displacement, vibration, temperature, acoustic emission, etc.) collected in real time or regularly, analyzes and extracts features of the data through data processing and machine learning technology, establishes a mapping relationship between data and crack parameters, extracts features related to cracks, and then judges the existence of cracks, crack size, location and other information; the advantage of this method is that the collected data is used to train machine learning models (such as support vector machines, decision trees, neural networks, etc.) to identify crack features, which can realize real-time monitoring of structural crack parameters, etc., without the need for destructive testing of the structure, reducing maintenance costs and risks; the disadvantage is that it is highly dependent on sensors, has high requirements for sensor layout and accuracy, and the location, type and sensitivity of sensors directly affect the detection effect. Improper location or insufficient performance of sensors will reduce the accuracy of crack identification, and sensors are difficult to be arranged in dangerous areas such as high temperature and high pressure, inside the structure, and at the edge. Data in dangerous areas is difficult to obtain, and the model built in this way will affect the recognition accuracy of cracks. The image-based crack recognition method mainly collects image data when the structure is in service, and mines texture features related to cracks from the image data to identify defect cracks. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention integrates data-driven and model-driven technologies to identify the crack parameters of mechanical structures. First, a forward analysis model of crack analysis is established using boundary element numerical analysis technology, and the physical field variables under different crack parameters in the dangerous area are calculated to obtain the relevant numerical analysis field variable cloud map; then, relevant sensors are arranged in the safe area to collect image data of the structure when it is in service, and the numerical analysis image data is combined with the sensor collected image data. After image processing using data fusion, normalization, data enhancement and other technologies, the image is input into a hybrid network structure, and a crack recognition technology based on the fusion of Fisher vector driven by digital analog and deep convolutional neural network is built. The texture features related to cracks are mined from the image data to identify crack defects, and the crack parameters in the mechanical structure are accurately identified. Compared with traditional crack recognition technology, the method of the present invention can obtain images of various areas (high temperature and high pressure, structural edges and interiors and other harsh working conditions) through numerical analysis and experimental testing, forming a relatively complete crack recognition image data sample set, and realizing parameter recognition of various types of cracks; boundary element model driven technology can accurately calculate the physical field variables in crack analysis, and can provide a more accurate data set for crack recognition; the image fusion method based on Fisher vector and deep convolutional neural network can simultaneously retain the image's detailed information (local features) and high-level semantic information (global features), thereby improving the comprehensiveness of the feature information represented by the image.

[0005] The present invention discloses a mechanical structure crack identification method based on boundary element field variable data drive, comprising the following steps:

[0006] S1: Establish a physical model for crack analysis of the inspected mechanical structure;

[0007] S2: Construct mathematical model for crack analysis;

[0008] S3: Generation of crack identification image dataset and its preprocessing;

[0009] S4: Build a deep learning model for crack identification.

[0010] Furthermore, step S1 is specifically as follows:

[0011] S11: Determine the structural features and geometry of the mechanical structure being inspected

[0012] Measure and record the geometric shape and size information of the mechanical structure being tested in detail, and determine the basic mechanical performance parameters such as elastic modulus, Poisson's ratio, density, etc. of the mechanical structure material being tested;

[0013] S12: Get the geometric information of the crack

[0014] Use ultrasonic detection technology to obtain geometric information of cracks in the mechanical structure being tested, including parameters such as shape, size, and position, as the basis for subsequent modeling and analysis;

[0015] S13: Establishing physical model for crack analysis

[0016] According to the geometric characteristics and material features of the mechanical structure to be inspected, a physical model of the mechanical structure to be inspected is established, and cracks are added to the physical model according to the geometric information of the cracks. Parameters of different types of cracks are set to obtain crack analysis physical models of different types such as crack parameters and positions. The crack analysis physical model is corrected through measurement and detection to ensure that the established physical model can truly reflect the crack situation and avoid large deviations between the physical model and the actual crack model.

[0017] Furthermore, step S3 is specifically as follows:

[0018] S31: Generate image data using mathematical models

[0019] According to the material and boundary conditions of the inspected mechanical structure and its service conditions, the stress and strain data of the surface of the inspected mechanical structure under different working conditions and different crack types are calculated by the boundary element numerical analysis method, and the cloud map of related physical quantities is obtained after post-processing.

[0020] S32: Data Collection

[0021] First, at the key position of the mechanical structure to be inspected, image data is collected by a high-definition camera, or stress field distribution data on the surface of the structure is collected using stress strain sensor equipment. The data is displayed in the form of an image through a visualization method to indirectly reflect the stress concentration in the crack area.

[0022] S33: Preprocessing the generated crack image data

[0023] The generated crack image data is preprocessed using bicubic interpolation technology to reduce the storage volume of image data and the training time of the prediction model.

[0024] Furthermore, step S4 is specifically as follows:

[0025] S41: Construct Fisher vector:

[0026] S42: Building a convolutional neural network model

[0027] Further, step S42 is specifically as follows:

[0028] S421. Construct convolutional layer;

[0029] S422. Construct ReLU activation function;

[0030] S423. Construct a pooling layer;

[0031] S424. Construct a fully connected layer;

[0032] S425. Construct Softmax layer;

[0033] S426. Parameter tuning of the built crack identification model is performed to optimize the performance of the CNN model.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] (1) This method uses the boundary element method to visualize and analyze crack characteristics, which can provide high-precision crack identification results. It can calculate the physical field variables under different crack parameters in any area and provide a more accurate data set for crack identification.

[0036] (2) A mechanical structure crack recognition model that performs classification methods based on multi-layer information encoded by CNN and Fisher vector provides an effective image fusion solution by combining the advantages of traditional image processing technology and deep learning. It has many advantages such as improving image quality and enhancing detail capture capabilities.

[0037] (3) The deep learning model built can describe the characteristics of crack images more comprehensively and accurately, thereby improving the accuracy of crack image recognition, especially in complex backgrounds or in the presence of interference. The combination of Fisher and CNN can also enable the model to learn more representative and universal features, reduce dependence on specific data sets, improve the generalization ability of the model, and thus improve the accuracy of convolutional neural networks in texture classification.

[0038] (4) The present invention combines data-driven and model-driven technologies to identify crack parameters of mechanical structures. By establishing a forward model for crack analysis through boundary element numerical analysis technology, a cloud map of physical field variables in any area can be generated, and fused with the image data collected by the sensor to further enrich the image data set and ensure the integrity of the sample data; after data fusion, normalization and enhancement, a crack recognition model based on digital-analog drive, Fisher vector and deep convolutional neural network is constructed, which can realize accurate recognition of crack parameters according to the texture features of the image. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 The overall flow chart of mechanical structure crack identification based on boundary element field variable data drive;

[0040] Figure 2 Build a flowchart for a hybrid model based on Fisher vectors and convolutional neural networks. DETAILED DESCRIPTION

[0041] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0042] like Figure 1-2 As shown, a mechanical structure crack identification method based on boundary element field variable data driven includes the following steps:

[0043] S1: Establish a physical model for crack analysis of the mechanical structure being tested

[0044] S11: Determine the structural features and geometry of the mechanical structure being inspected

[0045] Measure and record the geometric shape and size information of the mechanical structure being tested in detail, and determine the basic mechanical performance parameters such as elastic modulus, Poisson's ratio, density, etc. of the mechanical structure material being tested.

[0046] S12: Get the geometric information of the crack

[0047] Ultrasonic detection technology is used to obtain geometric information of cracks in the mechanical structure being inspected, including parameters such as shape, size and position, which serves as the basis for subsequent modeling and analysis.

[0048] S13: Establishing physical model for crack analysis

[0049] According to the geometric characteristics and material features of the mechanical structure to be tested, a physical model of the mechanical structure to be tested is established, and cracks are added to the physical model according to the geometric information of the cracks, and parameters of different types of cracks are set to obtain physical models of crack analysis with different crack parameters, positions, etc. The physical model of crack analysis is corrected by means of measurement and detection to ensure that the established physical model can truly reflect the crack situation and avoid large deviations between the physical model and the actual crack model.

[0050] S2: Constructing a mathematical model for crack analysis

[0051] In order to realize the crack identification technology based on field variable drive, it is first necessary to collect the physical field of crack analysis under different working conditions. The present invention uses boundary element crack analysis technology to calculate the structure with cracks, and uses the constitutive equation representing the stress-strain relationship to construct the dual boundary integral equation of crack analysis. The constitutive equation is

[0052] σ kl =D klrs ε rs =D klrs u r,s (1)

[0053] Among them, ε rs =(u r,s +u s,r ) / 2 is the strain tensor standard; σij represents the stress tensor, u r Denotes the displacement vector. klrs is the constitutive tensor, and its specific expression is:

[0054]

[0055] Wherein, μ(x) represents the shear modulus. In the present invention, μ(x) is defined as a function of the coordinate x, δ ij represents the Kronecker symbol; v represents the Poisson's ratio.

[0056] Use b k The stress balance equation can be expressed as

[0057] σ kl,l +b k =0 (3)

[0058] At a certain point, the equivalent surface force can be calculated using the Cauchy formula

[0059] t i =σ ij n j (4)

[0060] Take Kelvin's fundamental solution as the weight function U ik , and integrating the weak form of the equation part by part, the displacement boundary integral equation can be obtained from the equation:

[0061]

[0062] In equation (5), Γ n is the unbroken boundary, Γ c + is the upper crack surface, Γ c - is the lower crack surface, c lk (y) is the free term coefficient, which is related to the position of the source point y. For smooth boundaries, when l = k, c lk (y) = 0.5, and l≠k when c lk (y) = 0. Considering the plane strain problem, the displacement basic solution is Basic solution to dough mixing as follows:

[0063]

[0064] In equations (6) and (7), v is Poisson's ratio, G is the shear modulus, and r represents the distance between the source point and the field point. It should be noted that The following displacement boundary integral equation can be obtained:

[0065]

[0066] In equation (8), if the crack is free of surface forces, then t j (x + )+t j (x - )=0 and the crack opening displacement on the upper crack surface is u j (x + )-u j (x - ). When the source point y + Located on the upper crack surface When , the displacement boundary integral equation can be modified as follows:

[0067]

[0068] When the source point y - Located below the crack When , the traction boundary integral equation is used:

[0069]

[0070] In equation (10), the displacement basic solution D lkj (y,x), and the basic solution of stress S lkj (y,x) is as follows:

[0071]

[0072] Where D lkj (y + ,x + )=D lkj (y - ,x - ), S lkj (y + ,x + )=-S lkj (y - ,x - ), n l (y - )=-n l (y + ). The traction boundary integral equation on the lower crack surface can be transformed into the traction boundary integral equation on the upper crack surface, that is:

[0073]

[0074] Therefore, it is necessary to input the geometric information of the upper crack surface and obtain the geometric information of the lower crack surface through the upper crack surface.

[0075] The boundary integral equation is discretized using M units, which can be expressed as:

[0076]

[0077]

[0078] In equations (14)-(16), M is the number of elements on the uncracked boundary, M c is the number of elements on the upper crack surface.

[0079] In any unit m, the displacement and traction at point x can be expressed as:

[0080]

[0081] Substituting equation (17) into equations (14)-(16),

[0082]

[0083] In the four types of integrals (Eq. (18) to (25)), J(ξ) is the Jacobian mapping from Cartesian coordinates to local coordinates. α (ξ) represents the element shape function, at the crack tip element In other units N α (ξ)=N pα . When the source point y is located on the unit m, equations (18) and (20) have weak singularity, the integrals in equations (19), (21), (22) and (23) have strong singularity, and the integrals in equations (24) and (15) have hypersingularity, and cannot be calculated by conventional Gaussian integrals before regularization. For weak singular integrals, the present invention uses polar coordinate transformation to eliminate the influence of weak singular integrals on the solution of the equation; for strong singular integrals and hypersingular integrals, the present invention uses local coordinate approximate expansion method to solve, assuming that the coordinates of the source point are y and the coordinates of the field point are x, then the local coordinates of the source point in the integral unit are ξ. A first-order Taylor expansion is used near the source point.

[0084]

[0085] where ξ-ξ0=δ,

[0086] r 2 =(x i -y i ) 2 =A 2 δ 2 +2Cδ 3 +O(δ 4 ) (27)

[0087] In formula (25),

[0088]

[0089] Using equations (25) and (26), we can get

[0090]

[0091] Using the local approximation formula in equation (28), we can get

[0092]

[0093] Where sgn(x) represents the sign function.

[0094] Using formula (29), the Jacobian matrix at the field point can be obtained as follows:

[0095]

[0096] in

[0097]

[0098] The Jacobian J1(ξ), J2(ξ) and the shape function N α (ξ) is expanded by Taylor around the origin, and the following formula can be obtained:

[0099]

[0100] Using the Taylor expansion formula of equation (26)-(36) about the distance r, we substitute it into the strong singular integral term and the hypersingular integral term in equation (18)-(25). After simplification, the strong singular integral term and the hypersingular integral term can be transformed into conventional Gaussian integral and analytical integral, and a more accurate integral result can be obtained by using the direct solution method.

[0101] Finally, by solving the above boundary integral equation, we can obtain the physical field variables under different crack types. By adjusting and optimizing the boundary element crack analysis numerical model through analytical calculation, experimental verification and other methods, the crack analysis results are closer to engineering, providing data support for the identification of structural cracks.

[0102] S3: Generation of crack identification image dataset and its preprocessing

[0103] S31: Generate image data using mathematical models

[0104] According to the material and boundary conditions of the inspected mechanical structure and its service conditions, the stress and strain data of the surface of the inspected mechanical structure under different working conditions and different crack types are calculated by the boundary element numerical analysis method, and the cloud map of related physical quantities is obtained after post-processing.

[0105] S32: Data Collection

[0106] First, at the key position of the mechanical structure to be inspected, image data is collected by a high-definition camera, or stress field distribution data on the surface of the structure is collected using stress strain sensor equipment. The data is displayed in the form of an image through a visualization method to indirectly reflect the stress concentration in the crack area.

[0107] S33: Preprocessing the generated crack image data

[0108] The generated crack image data is preprocessed using bicubic interpolation technology to reduce the storage volume of image data and the training time of the prediction model. The bicubic interpolation technology process is as follows:

[0109] Estimate the pixel value of the image at a specific location by determining the interpolation point and the surrounding 16 pixels. Find the 16 neighboring points around it. These 16 points extend two units above, below, left, and right of the integer pixel grid where (x, y) is located, forming a 4×4 matrix. Assuming that the integer point on the grid is (i, j), the target interpolation point (x, y) is located in the center area of ​​the 16 pixels around (i, j). The core of bicubic interpolation is to use a cubic polynomial for interpolation in each direction, so the weights in the horizontal and vertical directions need to be calculated separately. When calculating weights, cubic interpolation polynomials are often used, and the weight formula can be expressed as:

[0110]

[0111] Among them, m is the distance between the interpolation point and the neighboring point, and n is a parameter (usually -0.5, called the cubic convolution kernel).

[0112] First, in the horizontal direction, for each row of 4 points, the calculated horizontal weight is multiplied by the corresponding pixel value and then added to obtain a temporary result for the row. Then in the vertical direction, the vertical weight is multiplied by these temporary results and then added to finally obtain the pixel value of the interpolation point (x, y).

[0113] After the crack image is processed by bicubic interpolation as above, the image quality is high and the problem of discontinuous grayscale value of the image is overcome, which can speed up the training speed of the state assessment model. In addition, in order to speed up the model network connection speed and the solution speed of gradient reduction, the present invention performs normalization processing on the collected and generated images, using the following formula:

[0114]

[0115] After compression and normalization, the image is input into the convolutional feature extraction model as a training sample to provide high-quality input data for the crack recognition algorithm. After network iterative operations, the relationship between it and the crack state is established to achieve the evaluation of its state.

[0116] S4: Building a deep learning model for crack identification

[0117] The present invention takes convolutional neural network (CNN) as the core, and uses Fisher vector to encode the image features extracted by CNN to construct a mechanical structure crack recognition model that uses CNN and Fisher vector to encode multi-layer information to perform a classification method. The model can more comprehensively and accurately describe the features of crack images, thereby improving the accuracy of crack image recognition, especially in complex backgrounds or in the presence of interference. The combination of the two can also enable the model to learn more representative and universal features, reduce dependence on specific data sets, improve the generalization ability of the model, and thus improve the accuracy of convolutional neural networks in texture classification.

[0118] The structural scheme of the method of the present invention is provided in detail below:

[0119] S41: Construct Fisher vector:

[0120] Let X = {x t ,t=1...T|x t ∈R D} represents a sample of T observations. Assume that the generation process of X can be obtained through a probability density function u with parameter λ λ Modeling. Then, the observations in X can be characterized by the following gradient vector.

[0121]

[0122] The gradient vector given by equation (39) can be classified using any classification algorithm. The Fisher information classification matrix F λ It can be defined as:

[0123]

[0124] Where E represents the mathematical expectation of X. Fisher kernel (FK) is used to evaluate the similarity between two samples X and Y. The kernel is defined as follows:

[0125]

[0126] because is positive semidefinite, so F λ Also. Using Cholesky decomposition K FK (X,Y) can be rewritten as:

[0127]

[0128] in

[0129]

[0130] vector is the Fisher vector, whose dimension is Therefore, performing classification using a linear kernel machine with Fisher vectors as feature vectors is equivalent to using K FK The classification is performed using a nonlinear kernel machine as the kernel.

[0131] Based on the advantages of Fisher vectors in extracting texture features of image data, the present invention uses CNN and multi-layer information encoded by Fisher vectors to perform classification. First, the crack image is used as the input of the pre-trained CNN, and the output of any number of convolutional blocks is collected. All selected layers contain feature information about the image. However, the features extracted from the later layers contain higher-level information than the features extracted from the earlier layers. The specific contents are as follows: Let represents the output set of the nth convolutional block, where D n and the scalar product T n =W n ·H n are the number of channels and the number of pixels for each channel respectively. n is a local feature, X n is the set of local features extracted from the nth convolutional block.

[0132] Let n and m represent two convolutional blocks, where D n ≥D m . Use X n→m Indicated by X n The local features are mapped to X m The set of local feature spaces is:

[0133]

[0134] Where f n,m : is a dimensionality reduction function. If two convolution blocks have the same number of channels, that is, D n =D m , then the function f n,m is the identity function. Otherwise, if D n >D m , then f n,m Each component in Given by:

[0135]

[0136] in is x t The dth component of .

[0137]

[0138] K=D n-i -D n-k ·S (47)

[0139] In addition to using average pooling and maximum pooling, the present invention also introduces the principal component analysis method to calculate f n,m .

[0140] Given k+1 convolutional blocks n,n-1,...,nk, through the pooling set X n ,X n-1 ,...,X n-k Construct a combination set X of local features. Such a fusion is given by equation (44), where k=2.

[0141]

[0142] Use the local feature combination set X = {x t ,t=1,2,...,T|x t ∈R D}Calculate the Fisher vector, where and D = D n-k Assume that each local feature x t By u λ The distribution is generated independently. In this case, equation (43) becomes:

[0143]

[0144] Select u λ As a Gaussian mixture model composed of K Gaussian distributions. In this model, u λ It is expressed as:

[0145]

[0146] Where λ = {w i ,μ i ,Σ i ,i=1,...,K} and w i ,μ i ,Σ i Respectively represent and Gauss u i The associated weights, means, and covariance matrices. γ i (x t ) represents the observation x t By Gauss u i Probability of generation:

[0147]

[0148] Since any distribution can be approximated with arbitrary precision by the weighted sum of Gaussian functions with diagonal covariance, the present invention assumes that the covariance matrix is ​​a diagonal matrix, and the solution formula (43) for the Fisher vector can be simplified as:

[0149]

[0150] Equations (52), (53) and (54) are used to calculate each component of the Fisher vector, which is any set of local features extracted from the CNN convolutional layer, and then the crack identification information is extracted from the last fully connected layer to realize the identification calculation of the structural crack parameters.

[0151] S42: Building a convolutional neural network model

[0152] The convolutional neural network (CNN) architecture for extracting image features based on Fisher vectors used in the present invention is composed of a convolutional layer, a maximum pooling layer, and a fully connected layer using a Softmax regression model. This type of neural network can be decomposed into two main parts. The first part is used to extract features from the image, and is composed of a convolutional layer, a pooling layer, an activation layer, and a normalization layer. The second part is composed of a fully connected layer, and its purpose is classification.

[0153] Given a network layer L i,j , this network layer is a function that can transform a convolutional block B j Written as a composition of layers, taking a tensor as input and outputting another tensor:

[0154]

[0155] CNN architectures for image classification are usually constructed by combining convolutional blocks, i.e., N = B1·B2····(X). Typically, two convolutional blocks B j and B k consists of the same layer sequence, that is, if L i,j is a pooling layer, then L i,k It will also be a pooling layer. The feature extraction part of CNN can be seen as a set of filters, where each channel of each convolutional layer is a specific filter, extracting mid-level local features from the early convolutional layers of the pre-trained backbone and combining them with the high-level local features of the last layer, which can be encoded by Fisher Vector (FV) pooling (model construction process as shown in Figure 2 The following is a detailed description of the construction scheme of the convolutional neural network model:

[0156] S421. Constructing convolutional layers

[0157] The main function of the convolution layer is to serve as a powerful feature extraction module to generate high-quality local feature descriptors and provide input for the subsequent Fisher vector. Each convolution kernel scans the feature area of ​​the input matrix to capture local patterns of the image, such as edges, corners, textures, etc.

[0158]

[0159] Where (i, j) is the output position of the feature map. k×k represents the convolution kernel size. i+m-1,j+n-1 is the pixel value of the corresponding position in the feature map, w mn is the weight of the convolution kernel. b is the bias term.

[0160] By applying the ReLU function after convolution, the network can not only be trained efficiently, but also better extract and process complex image features.

[0161] S422. Construct ReLU activation function

[0162] The derivatives of the Sigmoid and Tanh functions are close to zero, which prevents the error from propagating further and hinders parameter updates, resulting in the problem of gradient vanishing. The present invention uses ReLu as the activation function of the convolutional neural network neurons to increase the nonlinear characteristics of the network:

[0163]

[0164] S423. Constructing pooling layer

[0165] The pooling layer aggregates the features of the local area, and the generated downsampled features can be used as the basic input for Fisher vector modeling. This local aggregation enables the Fisher vector to better describe the global statistical distribution of local features. When the convolutional features are input into the Fisher vector model, the pooling layer compresses the number of descriptors to facilitate subsequent statistical modeling. The maximum value in the region is used to represent the features of the region. Assuming the pooling window size is p×p, the pooling layer outputs y ij The formula is:

[0166]

[0167] Among them, x (i-1)p+m,(j-1)p+n is the input feature map value in the pooling window, and (i, j) is the output position after pooling. The pooling operation can reduce the dimension and computational complexity, enhance the robustness of features, highlight key features, and reduce redundant information.

[0168] S424. Constructing a fully connected layer

[0169] After the convolution and pooling operations, the fully connected layer summarizes the feature information provided by the pooling layer and the convolution layer, and the feature map will pass through one or more fully connected layers for high-level feature combination. It integrates all information in the spatial dimension to form a global feature representation. This global feature can be used as input and combined with the Fisher vector. The last layer is combined with the Softmax layer to output a probability distribution. Assuming the input is vector x (flattened features), the weight matrix is ​​W, and the bias is b, the calculation formula for the output y is:

[0170] y=W·x+b (59)

[0171] S425. Construct Softmax layer

[0172] The last layer is the Softmax regression layer, which serves as the output layer and converts the output of the fully connected layer into a probability distribution.

[0173] Assume that the output of the fully connected layer is a vector z = [z1, z2, ..., z n ], y of the Softmax layer = [y1,y2,...,y n ]The calculation formula for the output is:

[0174]

[0175] Among them, y i Represents the probability of category i. The Softmax function ensures that all output values ​​are between [0,1] and the sum of all category probabilities is 1.

[0176] S426. Parameter tuning of the built crack identification model is performed to optimize the performance of the CNN model.

[0177] The acquired image data set is preprocessed in S33, which can accelerate the convergence speed of the model and improve the accuracy of the model; then the data set is divided into a training set and a test set and input into the model to identify the crack information of the mechanical structure and compare and analyze it with the detection results. The parameters of the model are adjusted by adjusting the learning rate of the model and adding a regularization norm penalty term to the loss function to prevent the overfitting of the model; in addition, the robustness of the model is evaluated by adding noise to the input signal (using Gaussian noise with a mean of zero and a variance equal to the image mean divided by the signal-to-noise ratio, and adding noise after the input image is resized and normalized) to further improve the robustness and generalization ability of the model until a crack recognition model that meets the requirements is obtained.

[0178] In summary, the multi-layer information fusion model constructed by the present invention based on CNN and Fisher vector encoding to execute the classification method can extract the information related to the crack characteristics contained in the image, establish a mapping relationship between the data and the crack parameters, output the characteristic information of the crack state, and realize the accurate identification of the structural crack information.

[0179] The above are preferred embodiments of the present invention. A person skilled in the art may make various changes or improvements on this basis. Without departing from the general concept of the present invention, these changes or improvements should fall within the scope of protection required by the present invention.

Claims

1. A mechanical structure crack identification method based on boundary element field variable data driven, characterized in that: The steps include: S1: Establish a physical model for crack analysis of the inspected mechanical structure; S2: Construct mathematical model for crack analysis; S3: Generation of crack identification image dataset and its preprocessing; S4: Build a deep learning model for crack identification.

2. A mechanical structure crack identification method based on boundary element field variable data drive according to claim 1, characterized in that: Step S1 is specifically as follows: S11: Determine the structural features and geometry of the mechanical structure being inspected Measure and record the geometric shape and size information of the tested mechanical structure in detail, and determine the basic mechanical performance parameters of the tested mechanical structure material, such as elastic modulus, Poisson's ratio and density; S12: Get the geometric information of the crack Use ultrasonic detection technology to obtain geometric information of cracks in the mechanical structure being tested, including shape, size and location parameters, as the basis for subsequent modeling and analysis; S13: Establishing physical model for crack analysis According to the geometric characteristics and material features of the mechanical structure to be inspected, a physical model of the mechanical structure to be inspected is established, and cracks are added to the physical model according to the geometric information of the cracks. Parameters of different types of cracks are set to obtain crack analysis physical models of different crack parameters and positions. The crack analysis physical model is corrected through measurement and detection methods to ensure that the established physical model can truly reflect the crack situation and avoid large deviations between the physical model and the actual crack model.

3. A mechanical structure crack identification method based on boundary element field variable data drive according to claim 1, characterized in that: Step S3 is specifically as follows: S31: Generate image data using mathematical models According to the material, boundary conditions and service conditions of the mechanical structure being tested, the stress and strain data of the surface of the mechanical structure being tested under different working conditions and different crack types are calculated by the boundary element numerical analysis method, and the cloud map of relevant physical quantities is obtained after post-processing; S32: Data Collection First, at the key position of the mechanical structure being tested, image data is collected by a high-definition camera, or stress field distribution data on the surface of the structure is collected by using stress and strain sensor equipment. The data is displayed in the form of images through a visualization method to indirectly reflect the stress concentration in the crack area. S33: Preprocessing the generated crack image data The generated crack image data is preprocessed using bicubic interpolation technology to reduce the storage volume of image data and the training time of the prediction model.

4. A mechanical structure crack identification method based on boundary element field variable data drive according to claim 1, characterized in that: Step S4 is specifically as follows: S41: Construct Fisher vector: S42: Build a convolutional neural network model.

5. A mechanical structure crack identification method based on boundary element field variable data drive as claimed in claim 4, characterized in that: Step S42 is specifically as follows: S421. Construct convolutional layer; S422. Construct ReLU activation function; S423. Construct a pooling layer; S424. Construct a fully connected layer; S425. Construct Softmax layer; S426. Parameter tuning of the built crack identification model is performed to optimize the performance of the CNN model.

Images