Explosive cylindrical part three-dimensional residual stress inversion method based on deep learning

Abstract

The invention discloses an explosive cylindrical part three-dimensional residual stress inversion method based on deep learning. The method comprises the following steps: establishing a data set of stripping strain under each temperature function; a full-connection neural network inversion model suitable for the explosive cylindrical part is constructed and formed; and inputting experimental data into the full-connection neural network inversion model, and outputting the three-dimensional residual stress distribution of the experimental sample. The method has the remarkable advantage that an efficient and accurate method is provided to invert and obtain the three-dimensional residual stress distribution in complex samples such as explosive cylindrical parts and the like. By combining finite element simulation and a deep learning technology, the method can utilize external measurable delamination strain data to reversely deduce internal residual stress distribution which is difficult to directly measure, so that the defect of single data acquisition dimension of a traditional residual stress measurement technology is overcome.

Description

Technical Field

[0001] The present invention relates to the field of detection technology, and more specifically to a three-dimensional residual stress inversion method of an explosive cylindrical component based on deep learning. Background Art

[0002] In industrial production, residual stress is considered to be the most important factor causing workpiece deformation, which will affect the service life and material properties of the workpiece product. PBX (polymer bonded explosive) material is a highly filled composite material formed by hot pressing. Hot pressing causes the interaction of crystals inside PBX, which produces complex residual stress inside PBX. Internal residual stress has a significant impact on the mechanical properties of the material and its destructive behavior. Therefore, understanding the precise distribution of residual stress inside PBX material plays a very important role in improving the mechanical properties of the material and optimizing the material production formula.

[0003] At present, although the existing residual stress test methods can measure the local residual stress of materials to a certain extent, they are insufficient in the dimension of measurement data, and it is difficult to obtain the three-dimensional residual stress distribution inside the material. For example, the X-ray diffraction method mainly measures surface stress and has limited depth; although the hole drilling method and the blind hole method can measure the residual stress at a certain depth, the data dimension is single and it is difficult to fully reflect the complex internal stress field. In addition, the experimental process of these methods is complicated, time-consuming and destructive to the samples, which limits their universality and effectiveness in practical applications. Summary of the invention

[0004] In view of the shortcomings of existing residual stress detection methods in obtaining data of a single dimension, the present invention proposes a three-dimensional residual stress inversion method for explosive cylindrical parts based on deep learning. The strain signal generated by stripping and removing materials is collected through optical fiber, and deep learning is used to obtain the residual stress distribution characteristics, so as to infer the initial three-dimensional distribution of residual stress.

[0005] One aspect of an embodiment of this specification provides a three-dimensional residual stress inversion method for explosive cylindrical parts based on deep learning, comprising:

[0006] Establish a data set of delamination strain under various temperature functions;

[0007] Construct a fully connected neural network inversion model suitable for explosive cylinders;

[0008] The experimental data are input into the fully connected neural network inversion model, and the three-dimensional residual stress distribution of the residual stress of the experimental sample is output.

[0009] In some embodiments, the establishment of a data set of delamination strain under each temperature function includes: performing finite element simulation based on the residual stress generation mechanism, and establishing a temperature distribution function according to the distribution of the maximum heat flux during the cooling process; inputting the obtained temperature distribution function into the finite element model, and performing a delamination experiment on a cylindrical sample to obtain a data set containing different temperature function parameters and their delamination strains, and establishing the delamination strain under each temperature function.

[0010] In some embodiments, the finite element model is established by constructing a geometric model of the sample using finite element analysis software, and simulating the basic distribution characteristics of the residual stress of the sample based on the residual stress generation mechanism of the sample.

[0011] In some embodiments, the temperature distribution function is generated by performing finite element simulation based on the residual stress generation mechanism, establishing a temperature distribution function (T(x, y, z)) as a temperature field according to the maximum heat flux distribution form during the cooling process of the explosive cylinder, and directly presetting the residual stress by presetting the temperature field.

[0012] In some embodiments, a peeling experiment is performed on a cylindrical sample, including: simulating an actual peeling experiment process, applying different temperature distribution functions (T(x, y, z)) to the sample, recording strain data on the sample surface after each peeling, and using it as a data set for model training.

[0013] In some embodiments, the establishment of a fully connected neural network inversion model includes: designing a fully connected neural network structure, using a ReLU activation function to perform nonlinear processing on the network so that it can process the input temperature distribution function (T(x, y, z)) and predict the residual stress distribution of the sample; using the parameters in the temperature distribution function as the input vector, and the delamination strain after finite element simulation as the output vector, each neuron in the fully connected layer is connected to the neurons in the previous layer to form a fully connected structure FCNN, and establishing a mapping relationship between the parameters in the temperature distribution function and the final delamination strain through the fully connected network;

[0014] In some embodiments, the fully connected neural network inversion model training and optimization includes:

[0015] The loss function is constructed using mean square error to evaluate the difference between the predicted results and the actual measured values. The Adam optimizer is used to adjust the network parameters and the learning rate is adaptively adjusted.

[0016] Use the back-propagation algorithm to update the weights and biases of the neural network; the input data passes through the neural network from the input layer to the hidden layer and then to the output layer, and each layer performs weighted summation and activation function calculation;

[0017] Calculate the gradient of the loss function at the output layer, propagate back to each layer, and calculate the gradient of each parameter; update the weights and biases of each layer based on the gradient and learning rate;

[0018] Through hyperparameter search, the optimal hyperparameters are found in each converged model to determine the optimal network structure and learning parameters.

[0019] In some embodiments, the step of inputting the experimental data into a fully connected neural network inversion model and outputting the three-dimensional residual stress distribution of the residual stress of the experimental sample includes: experimental data input: inputting the temperature-strain data obtained from the experiment into the trained fully connected neural network inversion model to predict the three-dimensional stress field distribution of the sample.

[0020] The beneficial effects of the present invention are:

[0021] A significant advantage of the present invention is that it provides an efficient and accurate method to invert the three-dimensional residual stress distribution inside complex samples such as explosive cylinders. By combining finite element simulation and deep learning technology, this method can use external measurable delamination strain data to invert the internal residual stress distribution that is difficult to measure directly, thereby overcoming the shortcomings of traditional residual stress measurement technology in obtaining a single data dimension. This method not only improves the accuracy and efficiency of residual stress analysis, but also expands the application scope of residual stress measurement, providing a new technical means for stress analysis and safety assessment of key engineering materials such as explosive cylinders. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] This specification will be further described in the form of exemplary embodiments, which will be described in detail by the accompanying drawings. These embodiments are not restrictive, and in these embodiments, the same number represents the same structure, wherein:

[0023] Figure 1 , Figure 2 The strain monitoring position in the embodiment is shown;

[0024] Figure 3 A flow chart of the method of the present invention is shown;

[0025] Figure 4 The fitting effect of the temperature distribution function on the temperature distribution field is demonstrated.

[0026] Figure 5 The residual stress distribution diagram inside the explosive cylinder obtained based on the inversion results. DETAILED DESCRIPTION

[0027] In order to more clearly illustrate the technical solutions of the embodiments of this specification, the following is a brief introduction to the drawings required for the description of the embodiments. Obviously, the drawings described below are only some examples or embodiments of this specification. For ordinary technicians in this field, this specification can also be applied to other similar scenarios based on these drawings without creative work. Unless it is obvious from the language environment or otherwise explained, the same reference numerals in the figures represent the same structure or operation.

[0028] It should be understood that the "system", "device", "unit" and / or "module" used in this specification is a method for distinguishing different components, elements, parts, parts or assemblies at different levels. However, if other words can achieve the same purpose, the words can be replaced by other expressions.

[0029] As shown in this specification and claims, unless the context clearly indicates an exception, the words "a", "an", "an" and / or "the" do not refer to the singular and may also include the plural. Generally speaking, the terms "comprise" and "include" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0030] Flowcharts are used in this specification to illustrate the operations performed by the system according to the embodiments of this specification. It should be understood that the preceding or following operations are not necessarily performed precisely in order. Instead, the steps may be processed in reverse order or simultaneously. At the same time, other operations may also be added to these processes, or one or more operations may be removed from these processes.

[0031] See also Figure 1-5 , a deep learning-based three-dimensional residual stress inversion method for explosive cylindrical parts, including:

[0032] Dataset preparation:

[0033] Establish a data set of delamination strain under various temperature functions; perform finite element simulation based on the residual stress generation mechanism, and establish a temperature distribution function according to the distribution of the maximum heat flux during the cooling process; input the obtained temperature distribution function into the finite element model, and conduct delamination experiments on cylindrical samples to obtain the parameters of different temperature functions and their corresponding delamination strains, and establish a delamination strain data set under different temperature functions.

[0034] The finite element model is established by constructing a geometric model of the sample through finite element analysis software (such as ANSYS or ABAQUS), and simulating the basic distribution characteristics of the residual stress according to the residual stress generation mechanism of the sample.

[0035] In some embodiments, the temperature distribution function is generated by performing finite element simulation based on the residual stress generation mechanism, establishing a temperature distribution function (T(x, y, z)) as a temperature field according to the maximum heat flux distribution form during the cooling process of the explosive cylinder, and directly presetting the residual stress by presetting the temperature field.

[0036] In some embodiments, a peeling experiment is performed on a cylindrical sample, including: simulating an actual peeling experiment process, applying different temperature distribution functions (T(x, y, z)) to the sample, and recording the strain data (ε1, ε2, ε3, ..., εn) of the sample after each peeling, and using this as a data set for model training. The characteristic parameters in different temperature functions are used as input vectors of the data set, and the strain data of the sample after peeling is used as an output vector.

[0037] Inversion model construction:

[0038] Construct a fully connected neural network inversion model suitable for explosive cylinders;

[0039] In some embodiments, the establishment of a fully connected neural network inversion model includes: designing a fully connected neural network structure, using the ReLU activation function to perform nonlinear processing on the network so that it can process the input temperature distribution function (T(x, y, z)) and predict the residual stress distribution of the sample; using the parameters in the temperature distribution function as the input vector, and the delamination strain after finite element simulation as the output vector, each neuron in the fully connected layer is connected to the neurons in the previous layer to form a fully connected structure FCNN, and through the fully connected network, a mapping relationship between the parameters in the temperature distribution function and the final delamination strain is established. Specifically:

[0040] FCNN structure: A fully connected neural network (FCNN) structure consisting of an input layer, a hidden layer, and an output layer is designed, and the ReLU activation function is used to perform nonlinear processing on the network so that the structure can process the characteristic parameter vector of the input temperature distribution function and predict the residual stress distribution of the sample based on the input temperature distribution function.

[0041] Loss function: The loss function is constructed using mean square error (MSE), and the data loss evaluates the difference between the predicted result and the actual measured value. The calculation formula of the loss function MSE is:

[0042]

[0043] where ε i is the i-th delamination strain in the data set, is the i-th predicted value obtained by the model calculation,

[0044] Optimizer selection: Use the optimizer Adam to adjust network parameters and adaptively adjust the learning rate to improve training efficiency and effect to minimize the total loss function. The specific optimization formula is as follows

[0045]

[0046] where θ t is the learning rate of the model during the iteration process, and are the first-order and second-order moment estimates of the adjusted vector gradient, respectively.

[0047] Training and optimization process

[0048] The training and optimization of the fully connected neural network inversion model includes:

[0049] The loss function is constructed using mean square error to evaluate the difference between the predicted results and the actual measured values. The Adam optimizer is used to adjust the network parameters and the learning rate is adaptively adjusted.

[0050] The weights (w) and biases (b) of the neural network are updated using the back-propagation algorithm. The input data passes through the neural network from the input layer to the hidden layer and then to the output layer, with weighted summation and activation function calculations performed at each layer.

[0051] Calculate the gradient of the loss function at the output layer, propagate back to each layer, and calculate the gradient of each parameter; update the weights and biases of each layer based on the gradient and learning rate;

[0052] Through hyperparameter search, the optimal hyperparameters are found in each converged model to determine the optimal network structure and learning parameters.

[0053] Experimental data application and model evaluation:

[0054] The method of inputting the experimental data into the fully connected neural network inversion model and outputting the three-dimensional residual stress distribution of the experimental sample includes: experimental data input: inputting the temperature-strain data obtained from the experiment into the trained fully connected neural network inversion model to predict the three-dimensional stress field distribution of the sample.

[0055] Model evaluation: Use additional test data sets to evaluate the prediction accuracy and generalization ability of the model, and verify the effectiveness of the model by comparing the stress distribution predicted by the model with the results of experiments or finite element analysis.

[0056] Through the above steps, a data-driven residual stress inversion model can be constructed, which can accurately predict the residual stress distribution of complex samples and provide a powerful tool for stress analysis in engineering applications.

[0057] The present invention proposes a three-dimensional residual stress inversion method for explosive cylindrical parts based on deep learning. This method combines finite element simulation and deep learning technology to achieve the three-dimensional residual stress distribution inside the PBX material by inverting the external delamination strain data. Specifically, the geometric model and residual stress distribution characteristics of the sample are established by finite element analysis software, and the strain data on the sample surface is obtained by simulating the delamination experiment. Then, the fully connected neural network (FCNN) in deep learning is used to process the data, and the trained model is used to predict the three-dimensional residual stress distribution of the sample. This method not only overcomes the shortcomings of the traditional test method of single data dimension, but also improves the accuracy and efficiency of residual stress analysis, expands the application scope of residual stress measurement, and provides a new technical means for stress analysis and safety assessment of PBX materials and other key engineering materials.

[0058] The following is a specific explanation using the PBX column size as an example:

[0059] 1. Establish a residual stress delamination dataset for explosive cylinders using a preset temperature field

[0060] In this embodiment, the size of the PBX column used is Φ110*120. According to its axisymmetric characteristics, a two-dimensional axisymmetric model is used in abaqus to simplify the calculation. At the same time, in this embodiment, after simulating the residual stress generation process by finite element method, the distribution of its maximum heat flux is used as the distribution function of the preset temperature field. In order to ensure that the set residual stress distribution function can ensure the self-balance of the stress distribution, and ensure that the setting of the distribution function conforms to the physical process, and at the same time ensure that the function setting strictly conforms to the physical reality, we have adopted an innovative method: first preset a temperature distribution field based on physical experimental data and theoretical analysis, and then use the finite element method to simulate it, so as to balance and determine the final stress distribution.

[0061] The method of presetting the initial residual stress by the self-balance of the model after the preset temperature distribution has the following theoretical basis, which can be derived from the thermoelastic equation in the quasi-static temperature change process in thermoelastic mechanics:

[0062]

[0063] Where Δu i ——Laplacian operator of displacement field; v——Poisson's ratio of material; e——volume strain, i.e. α——thermal expansion coefficient; θ i ——given temperature field.

[0064] It can be seen from the above thermoelastic equation that the displacement to be determined consists of two parts, one is the special solution of the equation determined by the existence of the temperature field, and the other is the general solution of the corresponding homogeneous equation, which is not directly related to the distribution of the temperature field. Therefore, when the given temperature field is quasi-static, the stress field inside the material can be expressed as a function of a given constant temperature field. Therefore, the initial residual stress distribution can be preset by presetting a constant temperature distribution field.

[0065] In this embodiment, the preset temperature function used is:

[0066]

[0067] where a i , x i ,y i is the characteristic parameter of the temperature distribution function T(x, y, z). In this embodiment, the neutral points used by the Gaussian function are (0, 0), (0, 60), (0, 120), (57.5, 60), where according to the symmetry of the model, the parameters used by the center points (0, 0) and (0, 120) are the same. Figure 4 The fitting effect of the function on the temperature distribution field is shown. According to the Gaussian function characteristics describing the temperature distribution in this embodiment, the characteristic functions that need to be parameterized are a0, a1, a2, a3, s x1 , a x2 ,s x3 ,s y1 ,s y2 , a y3 .

[0068] According to the established temperature distribution function: change the characteristic parameters in the Gaussian function, use DOE to generate 2000 temperature distribution functions with different characteristic parameters, and use abaqus to perform delamination experiments on them to obtain the delamination strain corresponding to the same position as the experimental measurement position. The different characteristic parameters obtained and their corresponding delamination strains are respectively corresponded as input vectors and output vectors to construct the entire training set.

[0069] The actual data, numerical modeling data and generated data collected above are made into a data set and divided into a training set, a validation set and a test set, with a division ratio of 3:1:1.

[0070] 2. Input the data set into the deep neural network model, and obtain the optimal model for predicting three-dimensional residual stress distribution through the loss function and optimizer.

[0071] Wherein, the network structure of the model includes an input layer, a hidden layer and an output layer, and the structure of the input layer and the hidden layer is consistent, that is, several neurons containing activation functions are arranged as the same layer, and each neuron in the layer receives the input of each neuron in the upper layer (the input layer receives data input) and performs independent calculations, and the same form of calculation is passed to the next layer until the output layer. The activation function of the output layer neuron is determined according to the prediction target, so it is different from the input layer and the hidden layer. The activation function used in the present embodiment is a Sigmoid function.

[0072] The specific training steps of the model are as follows

[0073] S21. Loss calculation and back propagation: Each neuron receives input from the previous layer, adds a bias term after weighted summation, and then transforms through the activation function. For the (j)th neuron in the (l)th layer, its output is It can be expressed as

[0074]

[0075] in, is the weight connecting the i-th neuron in the l-1th layer and the j-th neuron in the lth layer, is the bias of the jth neuron in the lth layer, σ is the activation function ReLU, is the output of the previous layer.

[0076] Once the data is forward propagated through all the layers of the network, the output layer will produce the predicted peeled layer results. These predicted results are used together with the actual label values ​​to calculate the loss function, which is calculated as follows:

[0077]

[0078] where ε i is the i-th delamination strain in the data set, The i-th predicted value obtained by the model calculation.

[0079] The gradient of each weight is calculated through the loss function, and then back propagation is performed to obtain the weight to be updated. The specific calculation formula is:

[0080]

[0081] is the gradient of the loss with respect to the output of this layer, is the gradient of the layer output with respect to the weights.

[0082] S22. Neural network parameter optimization: Through the built-in optimizer of the model, the model parameters are optimized according to the loss value obtained by back propagation; the weight and bias The update rule is:

[0083]

[0084] S23, Model Validation: Input the validation set into the model and optimize the hyperparameters based on the performance of the model on the validation set;

[0085] Repeat S21-S23 until one of the following training termination conditions is reached to obtain the optimized model parameters and hyperparameters, and then obtain the optimal model. The termination conditions are as follows:

[0086] 1. Converge on the training set, achieve the highest accuracy on the validation set, and achieve the best generalization performance on the test set;

[0087] 2. The maximum number of training times set is reached.

[0088] Among them, the weight matrix (w) and the bias vector (b) are the trainable parameters in the model. The hyperparameters in the network include: learning rate (1r), loss function weight (λ), the number of layers of the neural network (L), and the number of neurons in each hidden layer (j). In this model, the loss function used is the mean square error (MSE).

[0089] In this embodiment, the results obtained after optimization are as follows: Figure 5 Shown

[0090] The whole grain sample still maintains the state of internal tensile stress and surface compressive stress. At the same time, the internal tensile stress is more concentrated on the plane 40mm away from the center of the circle. The maximum stress value is 1.027MPa. The surface stress of the grain is mainly compressive stress. The maximum compressive stress is at the middle end of the side of the grain, with a maximum compressive stress of -2.088MPa. The compressive stress at other surface positions is between -1MPa and -1.5MPa, which is not much different from the test results of the surface residual stress in the experiment.

[0091] On the other hand, the present invention also discloses a three-dimensional residual stress inversion method of explosive cylindrical parts based on deep learning, which is used to predict its three-dimensional stress distribution after peeling, including: a data preprocessing module, which produces a data set based on finite element simulation results; a model training module, which is used to input the data set into the model and obtain the optimal model through a loss function and an optimizer; and a module for inputting the data to be tested into the optimal model to obtain a predicted value.

[0092] Similarly, it should be noted that in order to simplify the description disclosed in this specification and thus help understand one or more embodiments of the invention, in the above description of the embodiments of this specification, multiple features are sometimes combined into one embodiment, figure or description thereof. However, this disclosure method does not mean that the features required by the subject matter of this specification are more than the features mentioned in the claims. In fact, the features of the embodiments are less than all the features of the single embodiment disclosed above.

[0093] In some embodiments, numbers describing the number of components and attributes are used. It should be understood that such numbers used in the description of the embodiments are modified by the modifiers "about", "approximately" or "substantially" in some examples. Unless otherwise specified, "about", "approximately" or "substantially" indicate that the numbers are allowed to vary by ±20%. Accordingly, in some embodiments, the numerical parameters used in the specification and claims are approximate values, which may change according to the required features of individual embodiments. In some embodiments, the numerical parameters should take into account the specified significant digits and adopt the general method of retaining digits. Although the numerical domains and parameters used to confirm the breadth of their range in some embodiments of this specification are approximate values, in specific embodiments, the setting of such numerical values ​​is as accurate as possible within the feasible range.

[0094] Each patent, patent application, patent application publication, and other materials, such as articles, books, specifications, publications, documents, etc., cited in this specification are hereby incorporated by reference in their entirety. Except for application history documents that are inconsistent with or conflicting with the contents of this specification, documents that limit the broadest scope of the claims of this specification (currently or later attached to this specification) are also excluded. It should be noted that if the descriptions, definitions, and / or use of terms in the materials attached to this specification are inconsistent or conflicting with the contents described in this specification, the descriptions, definitions, and / or use of terms in this specification shall prevail.

[0095] Finally, it should be understood that the embodiments described in this specification are only used to illustrate the principles of the embodiments of this specification. Other variations may also fall within the scope of this specification. Therefore, as an example and not a limitation, alternative configurations of the embodiments of this specification may be considered consistent with the teachings of this specification. Accordingly, the embodiments of this specification are not limited to the embodiments explicitly introduced and described in this specification.

Claims

1. A three-dimensional residual stress inversion method for explosive cylindrical parts based on deep learning, characterized in that: include: Establish a data set of delamination strain under various temperature functions; Construct a fully connected neural network inversion model suitable for explosive cylinders; The experimental data are input into the fully connected neural network inversion model, and the three-dimensional residual stress distribution of the experimental sample is output.

2. A deep learning-based three-dimensional residual stress inversion method for explosive cylindrical parts according to claim 1, characterized in that: The method of establishing a data set of delamination strain under each temperature function includes: performing finite element simulation based on the residual stress generation mechanism, and establishing a temperature distribution function according to the distribution of the maximum heat flux during the cooling process; inputting the obtained temperature distribution function into the finite element model, and performing a delamination experiment on a cylindrical sample to obtain data sets containing different temperature function parameters and their delamination strains, and establishing the delamination strain data sets under each temperature function.

3. The method for three-dimensional residual stress inversion of explosive cylindrical parts based on deep learning according to claim 2 is characterized in that: The establishment of the finite element model: constructing a geometric model of the sample by finite element analysis software, and simulating the basic distribution characteristics of the residual stress according to the residual stress generation mechanism of the sample.

4. The method for three-dimensional residual stress inversion of explosive cylindrical parts based on deep learning according to claim 2 is characterized in that: Generation of the temperature distribution function: Based on the residual stress generation mechanism, finite element simulation is carried out, and according to the maximum heat flux distribution form during the cooling process of the explosive cylindrical component, a temperature distribution function (T(x, y, z)) is established as a temperature field, and residual stress is directly preset by presetting the temperature field.

5. The method for three-dimensional residual stress inversion of explosive cylindrical parts based on deep learning according to claim 2 is characterized in that: A peeling experiment was conducted on a cylindrical sample, including: simulating the actual peeling experiment process, applying different temperature distribution functions (T(x, y, z)) to the sample, recording the strain data on the sample surface after each peeling, and using it as a data set for model training.

6. The method for three-dimensional residual stress inversion of explosive cylindrical parts based on deep learning according to claim 1 is characterized in that: The method of establishing a fully connected neural network inversion model includes: designing a fully connected neural network structure, using a ReLU activation function to perform nonlinear processing on the network so that it can process an input temperature distribution function (T(x, y, z)) and predict the residual stress distribution of the sample; using parameters in the temperature distribution function as an input vector, and using the delamination strain after finite element simulation as an output vector, each neuron in the fully connected layer is connected to the neurons in the previous layer to form a fully connected structure FCNN, and establishing a mapping relationship between the parameters in the temperature distribution function and the final delamination strain through the fully connected network.

7. The method for three-dimensional residual stress inversion of explosive cylindrical parts based on deep learning according to claim 6 is characterized in that: The training and optimization of the fully connected neural network inversion model includes: The loss function is constructed using mean square error to evaluate the difference between the predicted results and the actual measured values. The Adam optimizer is used to adjust the network parameters and the learning rate is adaptively adjusted. Use the back-propagation algorithm to update the weights and biases of the neural network; the input data passes through the neural network from the input layer to the hidden layer and then to the output layer, and each layer performs weighted summation and activation function calculation; Calculate the gradient of the loss function at the output layer, propagate back to each layer, and calculate the gradient of each parameter; update the weights and biases of each layer based on the gradient and learning rate; Through hyperparameter search, the optimal hyperparameters are found in each converged model to determine the optimal network structure and learning parameters.

8. The method for three-dimensional residual stress inversion of explosive cylindrical parts based on deep learning according to claim 1 is characterized in that: The method of inputting the experimental data into the fully connected neural network inversion model and outputting the three-dimensional residual stress distribution of the residual stress of the experimental sample includes: experimental data input: inputting the temperature-strain data obtained from the experiment into the trained fully connected neural network inversion model to predict the three-dimensional stress field distribution of the sample.

Images