Crack slip type identification method based on neural network and pre-training principle

Abstract

The invention discloses a crack slip type identification method based on a neural network and a pre-training principle, and the method comprises the following steps: building a numerical calculation model according to the geometric features and material parameters of a real object structure; adding a crack source, and setting the size, position and type of the crack; adding calculation result output positions, wherein the output positions respectively correspond to the arrangement positions of the sensors in the real object structure; calculating an elastic wave field excited by crack slippage / expansion by using a numerical model, outputting an elastic wave time domain signal curve at a set position, and picking up the amplitude of the signal curve at each position; the type of the crack is changed randomly, other parameters are kept unchanged, signal amplitudes corresponding to different crack types are calculated, and a crack type characteristic parameter-signal amplitude data set is established; designing the number of neurons of an input layer, a middle layer and an output layer based on the feedforward neural network model; training a neural network model; and evaluating a training result to obtain a crack slippage type identification model.

Description

Technical Field

[0001] The invention relates to a crack slip type identification method, in particular to a crack slip type identification method based on a neural network and pre-training principles. Background Art

[0002] Crack identification technology is a key technology for evaluating structural integrity and predicting the remaining service life of structures, and can provide important data support for structural safety analysis. At present, traditional crack identification methods mainly identify parameters such as the crack extension stage, location and length. In order to further improve the accuracy of crack identification and identify key features such as the slip type and direction of cracks, some scholars have developed the moment tensor theory. The crack inversion method based on the moment tensor theory is an excellent crack dynamic extension identification method. This method can accurately calculate and identify the detailed characteristics of the crack through the elastic wave signal excited by the crack slip, and clarify which type of crack (shear or tensile) the crack belongs to, thereby providing detailed data support for the analysis of the crack extension rate and the stress / strain distribution characteristics of the crack area, so as to realize the judgment of the structural damage condition and the accurate assessment of the remaining service life.

[0003] The existing crack identification technology based on moment tensor theory solves the moment tensor of the crack through the spatial distribution characteristics of the acoustic emission signal, and then calculates the crack surface slip direction and normal direction of the corresponding crack through the eigenvalue and eigenvector of the moment tensor. Specifically, the existing method arranges a sensor array in the structure, records the elastic wave signal excited by the same crack surface slip at different positions in the structure, and extracts the first arrival amplitude of each signal. Then, the moment tensor solution formula containing the moment tensor elements to be calculated and the known displacement amplitude is written according to the simplified Green's function. In addition, the material parameters (wave velocity) and the direction cosines of the crack source-sensor connection line contained in the formula can be obtained through preliminary testing or crack source auxiliary positioning technology. Therefore, the solution formula only contains unknown moment tensor elements. According to the number of unknown elements, the appropriate number of sensor arrays are selected to receive signals, and the moment tensor corresponding to the crack slip source can be solved, and then the crack size, slip direction and strike characteristics can be obtained through tensor eigenvalue and eigenvector analysis.

[0004] The disadvantage of existing crack identification techniques based on moment tensor theory is that the methods are not very applicable. The reason for this disadvantage is that crack identification methods based on moment tensor theory rely on accurate Green's functions, that is, explicit expressions of elastic wave fields excited by unit loads. Currently, only simple models (infinite elastic bodies and thin plate structures) have explicit expressions of Green's functions. For structures with complex shapes and materials, it is very difficult to solve the explicit expression of Green's functions, resulting in existing crack identification methods based on moment tensor theory being applicable only to simple structural forms. Summary of the invention

[0005] Purpose of the invention: The purpose of the present invention is to propose a crack slip type identification method based on neural network and pre-training principle, which directly fits the mapping relationship between elastic wave field characteristics and crack slip equivalent moment tensor through data-driven principle, so as to solve the problem of crack surface slip equivalent moment tensor solution and type identification in complex structures.

[0006] Technical solution: The present invention comprises the following steps:

[0007] S1: Establish a numerical calculation model based on the geometric characteristics and material parameters of the real object structure;

[0008] S2: In the numerical model, add crack sources and set the size, location and type of cracks;

[0009] S3: In the numerical model, add the output positions of the calculation results, and the output positions correspond to the arrangement positions of the sensors in the real object structure;

[0010] S4: using the numerical model, calculate the elastic wave field excited by crack slip / extension, output the elastic wave time domain signal curve at the position set in S3, and pick up the amplitude of the signal curve at each position;

[0011] S5: repeat steps S2-S4, randomly change the crack type, keep other parameters unchanged, calculate the signal amplitude corresponding to different crack types, and establish a data set of crack type characteristic parameters-signal amplitude;

[0012] S6: Design the number of neurons in the input layer, intermediate layer, and output layer;

[0013] S7: dividing the crack type-signal amplitude data set obtained in S5 into training, testing and validation data sets, and training the neural network model;

[0014] S8: Evaluate the training results and obtain the crack slip type identification model.

[0015] The establishment of the numerical calculation model is specifically as follows:

[0016] Measure the geometry and size of real object structures and establish numerical calculation models;

[0017] Optimize the model and delete the minor structures that do not affect the propagation of elastic waves;

[0018] Delete the decoration and shape-preserving structure on the structure;

[0019] Obtain the material parameters of the object structure and assign the parameters to the corresponding numerical calculation model.

[0020] The adding of crack sources specifically comprises the following steps:

[0021] Determine the possible location of cracks based on the stress-strain distribution characteristics of the real object structure;

[0022] The slip type of the crack is determined by randomly selecting the slip vector and normal vector of the crack surface;

[0023] The crack size is selected as a constant value according to the initial crack characteristics of the test object.

[0024] The added calculation result output location is specifically:

[0025] According to the position of the sensor on the real object structure, the signal output position is set for the numerical model, and the output position is exactly the same as the sensor position;

[0026] The output displacement component is set according to the sensitive direction of the sensor on the real object structure.

[0027] The picking up the amplitude of the signal curve at each position specifically includes:

[0028] Use numerical methods to calculate the displacement curve along the sensitive direction of the sensor at the signal output position;

[0029] Extract the displacement amplitude of the signal according to the tangent slope of the displacement curve data;

[0030] Arrange the displacement amplitudes of all signal output displacements into column vectors in order;

[0031] The equivalent moment tensor of the crack is calculated according to the set crack slip vector and normal vector, and the moment tensor is decomposed to obtain the proportions of the isotropic component, the double couple component and the supplementary linear vector dipole component. The three proportions are combined into a column vector, which is a characteristic parameter of the crack type.

[0032] The displacement amplitude column vector and the proportion of the three basic components of the moment tensor are combined together to form a training data pair, where the displacement amplitude column vector is the model input parameter and the proportion of the three basic components of the moment tensor is the model output parameter.

[0033] The method of establishing a crack type characteristic parameter-signal amplitude data set specifically includes: randomly selecting and setting a crack surface slip vector and a normal vector; and storing each set of displacement amplitude column vector-proportion column vector data pairs to form a data set.

[0034] The number of neurons in the input layer, the intermediate layer and the output layer is designed based on a feedforward neural network model.

[0035] The number of neurons in the design input layer, intermediate layer and output layer specifically includes:

[0036] Use a two-layer feedforward neural network, design the number of input layer neurons according to the number of sensors, and the number of output layer neurons is equal to the number of sensors / number of signal output locations;

[0037] For a two-layer feedforward neural network, the number of neurons in the hidden layer is no less than 3;

[0038] For the two-layer feedforward neural network, the proportions of the isotropic, double-couple and complementary linear vector dipole components of the crack moment tensor decomposition are output respectively, corresponding to the three elements of the proportion column vector in turn.

[0039] The training neural network model is specifically:

[0040] Decomposing the displacement amplitude column vector-moment tensor into basic component proportion column vector data sets is divided into training data sets, test data sets and validation data sets according to the proportions;

[0041] Train a feed-forward neural network using the training dataset;

[0042] The performance of the trained feedforward neural network is tested using the test data set, and the performance of the neural network is verified using the validation data set.

[0043] The evaluation training results are specifically:

[0044] Analyze the data correlation coefficient R value during training, testing and validation;

[0045] If the model prediction accuracy meets the requirements, the model parameters are output and a usable prediction model is obtained; if the model prediction accuracy does not meet the requirements, return to step S6, train, test and evaluate the neural network model by modifying the number of neural units in the middle layer, and analyze the prediction accuracy.

[0046] Beneficial effects: The present invention uses a numerical model to simulate the elastic wave field excited by crack surface slip, establishes an elastic wave field-moment tensor training data set, and uses the data set to train a neural network model to obtain a crack identification method that can be applied to specific structures; the method does not rely on the explicit expression of the Green's function, but directly fits the mapping relationship between the elastic wave field characteristics and the crack slip equivalent moment tensor through the data-driven principle, thereby solving the problem of solving and type identification of the crack surface slip equivalent moment tensor in complex structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Schematic diagram of the model shape, crack source position and sensor arrangement in an embodiment of the present invention;

[0048] Figure 2 A schematic diagram of random slip type selection of cracks in an embodiment of the present invention;

[0049] Figure 3A schematic diagram of a feedforward neural network structure in an embodiment of the present invention;

[0050] Figure 4 The training, testing and verification results of the neural network in the embodiment of the present invention are shown in FIG. 1 , wherein (a) is the training result; (b) is the verification result; (c) is the test result; and (d) is all the results. DETAILED DESCRIPTION

[0051] The present invention will be further described below in conjunction with the accompanying drawings.

[0052] like Figure 1 As shown, the crack slip type identification method based on neural network and pre-training principle of the present invention comprises the following steps:

[0053] S1: Establish a numerical calculation model based on the geometric characteristics and material parameters of the real object structure; specifically:

[0054] S11: Measure the geometry and size of the real object structure and establish a numerical calculation model;

[0055] S12: Optimize the model as necessary, delete the secondary structures that do not affect the propagation of elastic waves, delete the skin on the surface of the structure or make it equivalent to the main structure;

[0056] S13: Delete the decoration and shape-preserving structure on the structure;

[0057] S14: Obtain material parameters of the object structure and assign the parameters to a corresponding numerical calculation model.

[0058] S2: In the numerical model, a crack source is added, and the size, location and type of the crack are set; wherein adding a crack source specifically includes the following steps:

[0059] S21: Determine the possible location of crack occurrence based on the stress-strain distribution characteristics of the real object structure;

[0060] S22: The slip type of the crack is determined by randomly selecting the slip vector and normal vector of the crack surface;

[0061] S23: The crack size selects a constant value according to the initial crack characteristics of the test object.

[0062] S3: In the numerical model, add the calculation result output positions, and the output positions correspond to the arrangement positions of the sensors in the real object structure; wherein, add the calculation result output positions, specifically:

[0063] S31: Set the signal output position for the numerical model according to the position of the sensor on the real object structure. The output position should be exactly the same as the sensor position;

[0064] S32: Setting the output displacement component according to the sensitive direction of the sensor on the real object structure.

[0065] S4: Using the numerical model, calculate the elastic wave field excited by crack slip / extension, output the elastic wave time domain signal curve at the position set by S3, and pick up the amplitude of the signal curve at each position; specifically including:

[0066] S41: using a numerical method to calculate a displacement curve along the sensitive direction of the sensor at the signal output position;

[0067] S42: extracting the displacement amplitude of the signal according to the tangent slope of the displacement curve data;

[0068] S43: Arrange the displacement amplitudes of all signal output displacements in order into a column vector;

[0069] S44: Calculate the equivalent moment tensor of the crack according to the set crack slip vector and normal vector, decompose the moment tensor to obtain the proportions of three basic components, namely, the isotropic component (ISO), the double couple component (DC), and the complementary linear vector dipole component (CLVD), and combine the three proportions into a column vector. The proportion column vector is the crack type characteristic parameter, and the crack slip type can be determined according to the column vector;

[0070] S45: Combine the displacement amplitude column vector and the three basic component proportions of the moment tensor to form a training data pair, wherein the displacement amplitude column vector is a model input parameter, and the three basic component proportions of the moment tensor are model output parameters.

[0071] S5: Repeat steps S2-S4, randomly change the crack type, keep other parameters unchanged, calculate the signal amplitude corresponding to different crack types, and establish a data set of crack type characteristic parameters-signal amplitude; specifically including:

[0072] S51: Randomly select and set the crack surface slip vector and normal vector, and repeat steps S2-S4.

[0073] S52: storing each set of displacement amplitude column vector-proportion column vector data pairs to form a data set.

[0074] S6: Based on the feedforward neural network model, design the number of neurons in the input layer, intermediate layer and output layer; specifically include:

[0075] S61: Use a two-layer feedforward neural network and design the number of input layer neurons according to the number of sensors. The number of output layer neurons is equal to the number of sensors / number of signal output locations;

[0076] S62: For a two-layer feedforward neural network, the number of hidden layers is set to 1, and the number of neurons contained in the hidden layer is not less than 3;

[0077] S63: For a two-layer feedforward neural network, the output layer contains 3 neurons, which respectively output the proportions of the isotropic (ISO), double couple (DC) and complementary linear vector dipole (CLVD) components of the crack moment tensor decomposition, which correspond to the three elements of the proportion column vector in turn.

[0078] S7: Divide the crack type-signal amplitude data set obtained in S5 into training, testing and validation data sets, and train the neural network model; specifically:

[0079] S71: decomposing the displacement amplitude column vector-moment tensor obtained in S5 into a basic component ratio column vector data set and dividing it into a training data set, a test data set and a validation data set according to a certain ratio;

[0080] S72: Use the training data set to train a feedforward neural network using the Levenberg-Marquardt algorithm;

[0081] S73: Use the test data set to test the performance of the trained feedforward neural network, and use the validation data set to verify the performance of the neural network.

[0082] S8: Evaluate the training results and obtain the crack slip type identification model for the target structure in S1, specifically:

[0083] S81: Analyze the data correlation coefficient R value during training, testing and validation. The closer the R value is to 1, the stronger the correlation between the predicted value and the true value, and the higher the prediction accuracy of the model;

[0084] S82: If the model prediction accuracy meets the requirements, the model parameters are output and a usable prediction model is obtained; if the model prediction accuracy does not meet the requirements, it can return to step S6, modify the number of neural units in the middle layer, and repeat S6-S8 to train, test and evaluate the neural network model, and analyze the prediction accuracy.

[0085] The present invention can improve the recognition accuracy of crack slip types in structures, overcome the problem that the existing crack moment tensor inversion method cannot be applied to complex morphological structures, and simplify the solution process of crack identification. The crack moment tensor solution method based on neural network and pre-training principle can be applied to any structure, and can quantitatively and accurately analyze the crack slip types in the structure, providing accurate data support for structural life assessment based on fracture mechanics theory.

[0086] Example

[0087] This embodiment is aimed at Figure 1The cubic object structure shown establishes a crack slip type identification method. The object structure is a three-dimensional solid structure. The embodiment will establish a crack slip type identification method for the object structure based on the neural network model and pre-training principle according to the aforementioned steps S1-S8. Specifically, it includes:

[0088] S1: A numerical calculation model is established according to the geometric characteristics and material parameters of the real object structure; the modeling method includes the following steps:

[0089] S11: measuring the geometric dimensions of the three-dimensional solid structure, the dimension of which is 1 m, and establishing a numerical calculation model;

[0090] S12: simplify the model as necessary. The model used in the embodiment has a simple structure and does not include a connection or skin structure, so no further simplification is required and this step is skipped.

[0091] S13: Deleting the decoration and shape-preserving structure on the structure. The model used in the embodiment is the most simplified model structure, so there is no need to delete the relevant structure, and this step is skipped;

[0092] S14: Obtain material parameters of the object structure. In the embodiment, the material of the three-dimensional solid structure is rock, and the elastic modulus, Poisson's ratio and density are 54 GPa, 0.23 and 2300 kg / m3 respectively. Assign the material parameters to the numerical model and proceed to the subsequent steps.

[0093] S2: In the numerical model, add a crack source and set the size, location and type of the crack; adding a crack source includes the following steps:

[0094] S21: According to the stress conditions of the three-dimensional solid structure, it can be determined that the location where the crack of the three-dimensional solid structure may occur is located inside the solid structure, with a depth of 0.5 meters from the upper surface;

[0095] S22: The crack slip type is determined by randomly selecting the crack surface slip vector and normal vector. Figure 2 As shown, the crack slip type is determined by the crack surface slip vector u and the normal vector v, and the crack slip type is set by randomly selecting the angle between vectors u and v (between 0 and 90°);

[0096] S23: The crack size is selected based on the geometric characteristics of the cracks that may occur in the structure. The relevant data can be obtained by consulting relevant literature. For this example, the initial crack size of the rock is selected as 0.01.

[0097] S3: In the numerical model, add the calculation result output positions, and the output positions correspond to the arrangement positions of the sensors in the real object structure; wherein the sensor position selection includes the following steps:

[0098] S31: According to the sensor position arrangement form on the real object structure, in this example, the sensors are installed on the upper surface of the structure. The number of sensors is 6, of which 5 sensors are arranged on the surrounding ring at the same angle interval, and the remaining 1 sensor is arranged at the center point, such as Figure 1 As shown;

[0099] S32: According to the propagation direction of the elastic wave excited by the crack, the sensitive direction of the sensor is set to be perpendicular to the upper surface.

[0100] S4: using the numerical model, calculating the elastic wave field excited by crack slip / extension, outputting the elastic wave time domain signal curve at the position set in S3, and picking up the amplitude of the signal curve at each position; specifically comprising the following steps:

[0101] S41: Calculate the elastic wave field caused by crack slip using a numerical calculation method, and output the elastic wave waveform at the sensor position;

[0102] S42: extracting the amplitude of the first arrival wave of the signal according to the tangent slope;

[0103] S43: Arrange the displacement amplitudes at all signal output positions in order into a column vector;

[0104] S44: Calculate the moment tensor of the corresponding crack according to the crack surface slip vector and normal vector set in step S2. Solve the eigenvalue of the moment tensor, and arrange the eigenvalues ​​along the diagonal to form a principal moment tensor, decompose the principal moment tensor into the proportions of three basic components (ISO, DC and CLVD components), and arrange the three proportion values ​​in order as a column vector;

[0105] S45: Combine the displacement amplitude column vector and the three basic component proportions of the moment tensor to form a training data set, wherein the displacement amplitude column vector is an input parameter and the three basic component proportions of the moment tensor are output parameters.

[0106] S5: repeating steps S2-S4, randomly changing the crack type, keeping other parameters unchanged, calculating the signal amplitudes corresponding to different crack types, and establishing a data set of crack type characteristic parameters-signal amplitudes; specifically comprising the following steps:

[0107] S51: randomly change the angle between the crack surface slip vector and the normal vector, and repeat steps S2-S4;

[0108] S52: For each calculation, each set of displacement amplitude column vector-basic component proportion column vector data pairs is stored, and the calculation is repeated 3000 times to obtain a data set of amplitude column vector-three basic component proportions.

[0109] S6: Based on Figure 3The feedforward neural network model shown in the figure designs the number of neurons in the input layer, intermediate layer, and output layer; including the following steps:

[0110] S61: Use a two-layer feedforward neural network model to establish a neural network structure. The number of neurons in the output layer is equal to the number of sensors. In this example, the number of sensors is 6, so the input layer has 6 neurons, corresponding to the 6 elements of the displacement amplitude column vector;

[0111] S62: In this example, the number of hidden layers of the neural network is 1, and the layer has 3 neurons;

[0112] S63: There are three model output parameters, which respectively output the proportion of isotropic (ISO), double couple (DC) and complementary linear vector dipole (CLVD) components of the crack moment tensor decomposition.

[0113] S7: dividing the crack type-signal amplitude data set obtained in S5 into training, testing and validation data sets, and training a neural network model; comprising the following steps:

[0114] S71: Divide the data set containing 3000 sets of data obtained in step S5 into a training data set, a test data set, and a validation data set, with each data set accounting for 70%, 15%, and 15% respectively;

[0115] S72: Train the neural network using the training data set, using the Levenberg-Marquardt algorithm;

[0116] S73: Use the training data set to train the feedforward neural network, then use the test data set to test the model's classification accuracy for cracks, and use the validation data set to verify the classification accuracy. The corresponding fitting results are as follows: Figure 4 As shown in (a) to (d).

[0117] S8: Evaluate training results

[0118] Depend on Figure 4 The fitting results show that in the trained neural network, the correlation coefficient between the true value and the predicted value is close to 1, and the model can better fit the mapping relationship between the displacement amplitude column vector and the three basic component column vectors. Therefore, the neural network meets the prediction accuracy requirements and can be used for crack slip type identification of real structures.

Claims

1. A crack slip type identification method based on neural network and pre-training principle, characterized in that: The following steps are involved: S1: Establish a numerical calculation model based on the geometric characteristics and material parameters of the real object structure; S2: In the numerical model, add crack sources and set the size, location and type of cracks; S3: In the numerical model, add the output positions of the calculation results, and the output positions correspond to the arrangement positions of the sensors in the real object structure; S4: using the numerical model, calculate the elastic wave field excited by crack slip / extension, output the elastic wave time domain signal curve at the position set in S3, and pick up the amplitude of the signal curve at each position; S5: repeat steps S2-S4, randomly change the crack type, keep other parameters unchanged, calculate the signal amplitude corresponding to different crack types, and establish a data set of crack type characteristic parameters-signal amplitude; S6: Design the number of neurons in the input layer, intermediate layer, and output layer; S7: dividing the crack type-signal amplitude data set obtained in S5 into training, testing and validation data sets, and training the neural network model; S8: Evaluate the training results and obtain the crack slip type identification model.

2. According to claim 1, a crack slip type identification method based on neural network and pre-training principle is characterized in that: The establishment of the numerical calculation model is specifically as follows: Measure the geometry and size of real object structures and establish numerical calculation models; Optimize the model and delete the minor structures that do not affect the propagation of elastic waves; Delete the decoration and shape-preserving structure on the structure; Obtain the material parameters of the object structure and assign the parameters to the corresponding numerical calculation model.

3. According to claim 2, a crack slip type identification method based on neural network and pre-training principle is characterized in that: The adding of crack sources specifically comprises the following steps: Determine the possible location of cracks based on the stress-strain distribution characteristics of the real object structure; The slip type of the crack is determined by randomly selecting the slip vector and normal vector of the crack surface; The crack size is selected as a constant value according to the initial crack characteristics of the test object.

4. According to claim 3, a crack slip type identification method based on neural network and pre-training principle is characterized in that: The added calculation result output location is specifically: According to the position of the sensor on the real object structure, the signal output position is set for the numerical model, and the output position is exactly the same as the sensor position; The output displacement component is set according to the sensitive direction of the sensor on the real object structure.

5. The method for identifying crack slip types based on neural network and pre-training principle according to claim 4 is characterized in that: The picking up the amplitude of the signal curve at each position specifically includes: Use numerical methods to calculate the displacement curve along the sensitive direction of the sensor at the signal output position; Extract the displacement amplitude of the signal according to the tangent slope of the displacement curve data; Arrange the displacement amplitudes of all signal output displacements into column vectors in order; The equivalent moment tensor of the crack is calculated according to the set crack slip vector and normal vector, and the moment tensor is decomposed to obtain the proportions of the isotropic component, the double couple component and the supplementary linear vector dipole component. The three proportions are combined into a column vector, which is a characteristic parameter of the crack type. The displacement amplitude column vector and the proportion of the three basic components of the moment tensor are combined together to form a training data pair, where the displacement amplitude column vector is the model input parameter and the proportion of the three basic components of the moment tensor is the model output parameter.

6. The method for identifying crack slip types based on neural network and pre-training principle according to claim 5 is characterized in that: The method of establishing a crack type characteristic parameter-signal amplitude data set specifically includes: randomly selecting and setting a crack surface slip vector and a normal vector; and storing each set of displacement amplitude column vector-proportion column vector data pairs to form a data set.

7. The method for identifying crack slip types based on neural network and pre-training principle according to claim 6 is characterized in that: The number of neurons in the input layer, the intermediate layer and the output layer is designed based on a feedforward neural network model.

8. The method for identifying crack slip types based on neural network and pre-training principle according to claim 7 is characterized in that: The number of neurons in the design input layer, intermediate layer and output layer specifically includes: Use a two-layer feedforward neural network, design the number of input layer neurons according to the number of sensors, and the number of output layer neurons is equal to the number of sensors / number of signal output locations; For a two-layer feedforward neural network, the number of neurons in the hidden layer is no less than 3; For the two-layer feedforward neural network, the proportions of the isotropic, double-couple and complementary linear vector dipole components of the crack moment tensor decomposition are output respectively, corresponding to the three elements of the proportion column vector in turn.

9. The method for identifying crack slip types based on neural network and pre-training principle according to claim 8 is characterized in that: The training neural network model is specifically: Decomposing the displacement amplitude column vector-moment tensor into basic component proportion column vector data sets is divided into training data sets, test data sets and validation data sets according to the proportions; Train a feed-forward neural network using the training dataset; The performance of the trained feedforward neural network is tested using the test data set, and the performance of the neural network is verified using the validation data set.

10. The method for identifying crack slip types based on neural network and pre-training principle according to claim 9, characterized in that: The evaluation training results are specifically: Analyze the data correlation coefficient R value during training, testing and validation; If the model prediction accuracy meets the requirements, the model parameters are output and a usable prediction model is obtained; if the model prediction accuracy does not meet the requirements, return to step S6, train, test and evaluate the neural network model by modifying the number of neural units in the middle layer, and analyze the prediction accuracy.

Images