Turbine disc microstructure nondestructive determination method and system based on convolutional neural network

Abstract

The invention discloses a turbine disk microstructure nondestructive determination method and system based on a convolutional neural network. The method comprises the following steps: (1) selecting key points needing to be subjected to organization control in a turbine disk; (2) establishing finite element simulation in the forging process of the turbine disc; (3) microstructure requirements are put forward for the average grain size and recrystallization fraction of simulation results, classification is carried out, and forging process data, mainly including time, upper cross beam displacement, upper die load, lower die load, upper cross beam speed and temperatures of two observation points, of each group of simulation are extracted; (4) establishing and training a convolutional neural network quality judgment model; and (5) importing the forging process data of the turbine disc into the convolutional neural network quality judgment model, and judging the average grain size and recrystallization fraction requirement of the turbine disc. According to the method, under the condition that the integrity of the turbine disc is not damaged, whether the average grain size and the recrystallization fraction of the turbine disc meet the target requirements or not can be effectively judged, and a theoretical basis and technical support are provided for high-quality forging.

Description

Technical field:

[0001] The invention belongs to the technical field of forging, and relates to a method and system for non-destructive determination of turbine disk microstructure based on a convolutional neural network. Background technology:

[0002] There is a very serious waste of raw materials or resources in the production process of high-temperature alloy turbine disks. Among them, the sunk cost of scrap caused by the low yield rate of forgings accounts for a large proportion. This is because the quality inspection of high-temperature alloy turbine disks is often carried out during or after machining, and in practice many forgings themselves have certain defects, such as residual stress, uneven structure, pores, looseness, etc., which lead to the microstructure of the key areas of the forgings after forming cannot meet the requirements, but cannot be directly detected. The final forgings with defects flow to the next production process, and a large part of the parts will crack, deform and other problems, resulting in a lot of waste of manpower and material resources. The quality of forgings can be judged without destructive testing, which can reduce subsequent waste, improve the yield rate, and has engineering value.

[0003] In view of the above difficulties, the present invention proposes to use advanced artificial intelligence and machine learning methods for the die forging process of turbine disks, and build a corresponding execution system to perform data mining on the forging process data, distinguish the data differences between finished products and waste products in the forging process, and make quality judgments, which is a new means to solve this problem. Finally, an artificial intelligence model is established that can make microstructure quality judgments based on forging process data. The invention and application of this method has important technical guiding significance for reducing waste in the manufacturing process of turbine disks.

[0004] The patent specification with publication number CN112464533B discloses a method for controlling the gradual microstructure of a turbine disk based on numerical simulation of the entire preform and forging process. The method mainly controls the internal distribution of grain size by adjusting the preform and forging process parameters. It lacks an exploration of the relationship between the two and is unable to determine the microstructure quality of the forging based on the forging process data.

[0005] The patent specification with publication number CN115921734A discloses a method for local forming and manufacturing a turbine disk of a high-temperature alloy GH738. The method only involves the forging process parameters of the turbine disk forging of the GH738 alloy, and does not deeply explore the changes in the internal structure and final form of the turbine disk during the forging process, and cannot theoretically connect the forging process data and the microstructure. Summary of the invention:

[0006] The purpose of the present invention is to provide a method and system for non-destructive determination of turbine disk microstructure based on convolutional neural network. According to this method, a complete execution system can be built to extract reasonable features of forging process data, which can achieve the effect of determining the microstructure status of key points without destroying the forgings, and effectively improve the quality of forgings.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is:

[0008] 1. A method for non-destructive microstructure determination of a turbine disk, characterized in that the method extracts reasonable features from forging process data, thereby achieving the effect of determining the microstructure status of key points without destroying the forging, and effectively improving the quality of the forging, and the method comprises the following steps:

[0009] Step 1: Select the key points in the turbine disc that need to be organized and controlled;

[0010] Step 2: Establish finite element simulation of the whole process of turbine disk forging for multiple working conditions;

[0011] Step 3: Propose microstructural requirements for the average grain size and recrystallization fraction of the simulation results, perform artificial classification, and extract the forging process data of each group of simulations;

[0012] Step 4: Establish and train the convolutional neural network quality discrimination model;

[0013] Step 5: Import the turbine disk forging process data into the convolutional neural network quality judgment model to judge whether the average grain size and recrystallization fraction of the turbine disk are qualified or not, and verify the accuracy of the model.

[0014] Step 6: Based on the above-mentioned non-destructive determination method of turbine disk microstructure based on convolutional neural network, in order to ensure the smooth implementation of the method, a complete application system is designed for it to execute the above-mentioned non-destructive determination method of turbine disk microstructure. The system includes a forging monitoring module, a data receiving module, a data processing module, and a determination module.

[0015] 2. The method as claimed in claim 1, characterized in that the process data to be collected in step 3 include the displacement of the upper beam of the press and the upper die load, lower die load, upper beam speed and the temperature of two observation points on the upper die at that moment. The die forging process data must first be processed uniformly to generate a data structure such as (x1, x2…, x n ,y) matrix, where x irepresents a vector of process data, and y represents whether the final forging meets the requirements, which is 1 if it meets the requirements and 0 if it does not. Due to the algorithm requirements, some columns of all 0 are added to complete the matrix, and the final matrix size is 32×32. These matrices will be used as training data to establish the convolutional neural network model.

[0016] 3. The method as described in claim 1 is characterized in that: the output value of the quality discrimination convolutional neural network model described in step 4 is not the standard 0 and 1, so 0.5 is selected as the discrimination point, that is, if the result is greater than or equal to 0.5, it is judged as 1, that is, it meets the microstructure requirements; if it is less than 0.5, it is judged as 0, that is, it does not meet the requirements.

[0017] 4. Based on the above-mentioned non-destructive determination method of turbine disk microstructure based on convolutional neural network, in order to ensure the smooth implementation of the method, a complete application system is designed for it to execute the above-mentioned non-destructive determination method of turbine disk microstructure. The system includes a forging monitoring module, a data receiving module, a data processing module, and a determination module.

[0018] The forging monitoring module performs multiple sets of finite element simulations on key points of turbine disc structure control and extracts and organizes forging process data.

[0019] The data receiving module is used to receive the input finite element processing data of the key points of the turbine disk structure control and the extracted forging process data, and pass them to the subsequent processing links to provide basic reference data for the entire system.

[0020] The data processing module is used to further process the forging process data, train and build a corresponding convolutional neural network quality discrimination model, and store the corresponding data set to improve the model.

[0021] The determination module inputs the obtained forging process data into the trained convolutional neural network model for microstructure determination and outputs the final result.

[0022] 5. According to the convolutional neural network-based turbine disk microstructure non-destructive determination system of claim 4, it is characterized in that, in the receiving module, the following target data attributes are selected as data features used by the model:

[0023] (1) Simulation information, including finite element simulation of the entire turbine disc forging process to extract key point deformation information

[0024] (2) Forging process information, including the specific implementation steps of the forging process and the requirements for various parameters.

[0025] 6. According to claim 4, a convolutional neural network-based turbine disk microstructure non-destructive determination system is characterized in that, in the data processing module, a convolutional neural network quality discrimination model can be trained according to the existing forging process data, and the corresponding data set can be stored to continuously improve the model. Description of the drawings:

[0026] Figure 1 This is the quality determination process of the key points of the turbine disk in the specific embodiment of the present invention;

[0027] Figure 2 The distribution position of the key points of the turbine disk in the specific embodiment of the present invention;

[0028] Figure 3 Turbine disk temperature measurement point distribution

[0029] Figure 4 Convolutional neural network structure diagram

[0030] Figure 5 Sparse connection diagram

[0031] Figure 6 Flowchart of establishing CNN neural network model

[0032] Figure 7 Schematic diagram of the receptive field of a convolutional neural network

[0033] Figure 8 ROC curve of CNN discriminant model

[0034] Fig. 9 CNN model quality judgment results Specific implementation method:

[0035] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0036] The present invention is a method and system for non-destructive determination of the microstructure of a turbine disk based on a convolutional neural network. Taking the non-destructive microstructure determination of a certain type of aircraft engine high-pressure turbine disk as an example, the implementation details of the quality determination method and system involved in the present invention are introduced in detail.

[0037] This method mainly includes the following steps:

[0038] Step 1: In a real service environment, most areas of the turbine disk will not fail. Therefore, in order to simplify the manufacturing problem and reduce the manufacturing cost, the microstructure distribution of a specific area of ​​the turbine disk can be guaranteed to meet the use requirements.

[0039] The wheel rim and wheel core are the high-risk areas for low-cycle fatigue failure of turbine disks. The fatigue performance of this area largely determines the service life of the turbine disk. Each start and stop of the engine will cause a small amount of plastic deformation in this area, which will then cause fatigue failure due to the accumulation of plastic deformation and the expansion of cracks. Therefore, key points P1, P5 and P6 should be set on the wheel rim and wheel core. The distribution positions of the key points are as follows: Figure 2 In order to obtain good fatigue performance, the microstructure corresponding to these three key points should be uniform and fine-grained.

[0040] Step 2: Some samples are needed to establish the CNN neural network model. Therefore, the finite element method is used to simulate the isothermal die forging process of the final forging of the GH4169 alloy engine turbine disk, and the die forging process data is extracted for CNN neural network modeling. The simulation model is the same as that described in Chapter 3. The collected process data include time, upper beam displacement, upper die load, lower die load, upper beam speed and two observation point temperatures. The temperature measurement points of the die are as follows: Figure 3 The initial parameter settings of the finite element model are shown in Table 1.

[0041] Table 1 Initial parameter settings of the finite element model

[0042]

[0043]

[0044] In the actual high-temperature alloy mold, the arrangement of the temperature measuring holes can be different, which does not affect the feature learning of the CNN neural network model. The process parameter settings of each case in the isothermal die forging simulation of the high-temperature alloy turbine disk are shown in Table 2. Table 2 Process parameter table of turbine disk numerical simulation

[0045]

[0046] The temperature interval of each simulation group is 10℃.

[0047] Step 3: Propose microstructure requirements for the average grain size and recrystallization fraction of the simulation results, and perform artificial classification to extract the forging process data of each group of simulations; the grain size of the wheel rim and the wheel core is less than 12μm, and the recrystallization fraction is greater than 90% to be qualified. Finally, 9 groups of simulated forgings meet the requirements. All 22 groups of data will be formatted and processed to generate 22 data structures in the form of (x1, x2..., x n ,y) matrix. irepresents a vector of process data, and y represents whether the final forging meets the requirements, which is 1 if it meets the requirements and 0 if it does not. Due to the algorithm requirements, some columns of all 0 are added to complete the matrix, and the final matrix size is 32×32. These matrices will be used as training data to establish the CNN neural network model.

[0048] Step 4: Establish and train the convolutional neural network quality discrimination model;

[0049] As the most important development field in the field of target applied mathematics, machine learning methods have been widely infiltrated into all walks of life. Convolutional neural network (CNN) is an important algorithm in machine learning methods, and its main function is to identify features of large amounts of data.

[0050] CNN is a deep neural network with a structure like Figure 4 As a special deep neural network model, it has the following characteristics:

[0051] Shared weights: The weights of the convolution kernels in the convolutional layer of CNN are shared, which greatly reduces the order of magnitude of parameters in the network.

[0052] Sparse connection: The convolution kernels or feature matrices of each layer in CNN are not connected in a fully connected form, but in a partially connected form, such as Figure 5 As shown in the figure, this reduces the complexity of the model and suppresses the occurrence of overfitting.

[0053] Change the activation function: Use ReLU as the activation function. Since the derivative of ReLU is always 1 for positive input, it can well transfer the gradient to the previous network.

[0054] GPU programming: Using GPU for computing has improved computing performance by orders of magnitude compared to the CPU era.

[0055] In addition, convolutional neural networks have another biggest feature - convolution operation. The essence of convolution is to define a series of neighborhoods of a target point according to certain conditions and then superimpose them. For example, the target features in an image are mainly reflected in the relationship between pixels and surrounding pixels. These neighborhood pixel relationships form lines, corners, contours, etc. The convolution operation is exactly this operation that uses neighborhood points to redefine the value of the point according to certain weights. After the above features are effectively implanted, the powerful expression ability of convolutional neural networks and the advantages of supervised automatic feature extraction are fully demonstrated.

[0056] Based on these characteristics, CNN models are widely used in the field of image feature recognition. On this basis, image data is a two-dimensional spatial topological structure, and the time topological structure of forging process data has certain similarities with it. In the time topological structure, with time as the axis, at the same time, the displacement, load, temperature and other data are interconnected. And the related data of the previous and next moments are interconnected with the current moment. Since the features and relationships between these data will not be translated or rotated, the forging process data can be regarded as a simplified two-dimensional spatial topological structure. Therefore, the CNN model can be used to perform mechanical learning on the high-temperature alloy turbine disc die forging process data, and the quality of forgings can be judged by learning and identifying the data features of the forging process data.

[0057] The CNN model structure includes convolutional layers and fully connected neuron layers. In the convolutional layer of CNN, learnable convolution kernels and activation functions are included. The feature matrix of the previous layer is obtained by the operation of convolution and activation function to obtain the feature mapping matrix of the next layer. The feature mapping calculation is shown in formula (1).

[0058]

[0059] represents the j-th feature mapping matrix of the l-th layer, M j represents the feature matrix of the previous layer selected as input, is the convolution kernel used to convolve the i-th selected feature matrix of the l-1th layer, is the bias vector.

[0060] The error E of each unit in the reverse transmission process can be regarded as a function of the disturbance of the bias b, and the influence of the change of the bias b on the error E is called sensitivity, that is,

[0061]

[0062] Since the associated sensitivity map δ of a unit in the next layer corresponds to a block in the matrix of the previous layer, the sensitivity δ of a unit in the convolutional layer l can be obtained by summing all the sensitivities of the next layer to which the unit is connected and multiplying it by the derivative of the associated weight connecting the l+1 layer to the l layer and the activation function of the input u, as shown in formula (3)

[0063]

[0064] Where up() refers to the function that spreads the parameters in the brackets into a matrix, and β is the weight matrix of the downsampling layer. The up() function can be expanded using a simple Kronecker algorithm, as shown in formula (4).

[0065]

[0066] At this point, after obtaining the sensitivity matrix, the gradient of the bias b can be obtained by summing it, as shown in Equation 5.

[0067]

[0068] Finally, the gradient of the convolution kernel can be calculated using the back-propagation formula. As mentioned earlier, the weights in the CNN architecture are shared. When calculating the gradient of the convolution kernel, the gradients of all weights connected to the unit should be summed.

[0069]

[0070] in, In the previous layer The small block selected when performing element-by-element convolution calculations with the convolution kernel. The value at the (u,v)th position in the output convolution feature matrix is ​​the result of the calculation of the small block at that position in the previous layer with the convolution kernel.

[0071] In the CNN structure, there is also a special mapping structure called downsampling. The result of downsampling is to compress the feature mapping matrix of the previous layer and output the feature mapping matrix of the next layer. Different from the above, the compression process will not reduce the number of mapping matrices. That is, if the upper layer has N mapping matrices as input, the lower layer will also output N mapping matrices, but the size of the matrix will be compressed. The formula is as follows,

[0072]

[0073] Among them, β and b are the multiplication and addition bias of the downsampling function respectively, and the down() function is the downsampling function. Usually, the downsampling function divides the input mapping matrix into m blocks, each of which is n×n small blocks. And the elements of each small block are directly summed, so the output feature mapping matrix will be n times smaller than the input matrix.

[0074] The sensitivity calculation for the downsampling layer is similar to the above method. The sensitivity for the additive bias is shown in Equation 8.

[0075]

[0076] To calculate the multiplication bias sensitivity, it is necessary to define an intermediate matrix d. The expression of the d matrix is ​​shown in Formula 9.

[0077]

[0078] The sensitivity of the multiplication bias is shown in Equation 10.

[0079]

[0080] Finally, the delta rule can be used to update the weights of each neuron or convolution kernel, as shown in Equation 11 and

[0081] As shown in formula 12.

[0082]

[0083] Among them, η is the learning rate, and the learning and updating of the convolution kernel and weights of the CNN neural network can be performed.

[0084] In order to ensure faster operation speed and better embedding in the die forging knowledge management system, the CNN neural network model will be established based on the MATLAB platform. The establishment process of the CNN neural network model for identifying the quality of turbine disk forgings is as follows: Figure 6 As shown,

[0085] The key to establishing a CNN network lies in the construction of a neural network framework. The CNN network framework is the core of information and data transmission, and its suitability directly determines the performance of the entire model. For the process data of high-temperature alloy turbine disc forging, the current CNN neural network framework is mainly divided into input layer, convolution layer, pooling layer, fully connected layer and output layer. Its node parameter configuration is shown in Table 3.

[0086] Table 3 CNN framework node parameters

[0087]

[0088] The input layer is a single-layer structure, and its main function is to import the input matrix into the CNN model to prepare for further calculation and processing. Since the size of the input matrix is ​​32×32, the size of the input layer is also 32×32.

[0089] In convolutional neural networks, "receptive field" is an important model parameter. If it is not set properly, the model will not be able to effectively identify the input feature matrix during the convolution and pooling process, resulting in poor final model accuracy. The definition of receptive field is the size of the area mapped by the pixel points on the feature map output by each layer of the convolutional neural network on the original image. It can be described by its center position and its size, such as Figure 7 As shown,

[0090] The size of the receptive field of the convolutional layer is calculated by deducing from the back to the front, while the size of the receptive field of the pooling layer can be set in a custom way. After repeated verification, the optimal receptive field size distribution is shown in the following table

[0091] Table 4 CNN model receptive field size settings

[0092]

[0093] In addition to the framework of the CNN model, there are some general types of training parameters that need to be defined. These parameters are similar to the training parameters of the traditional ANN neural network model, as shown in the following table.

[0094] Table 5 CNN model training parameters

[0095]

[0096]

[0097] Write the above parameters into the MATLAB program and import the input matrix to start the training of the CNN neural network model. The training goal is to identify whether the quality of high-temperature alloy turbine disk forgings is qualified through forging process data.

[0098] Step 5: Import the turbine disk forging process data into the convolutional neural network quality judgment model to judge whether the average grain size and recrystallization fraction of the turbine disk are qualified or not, and verify the accuracy of the model.

[0099] ROC (Receiver Operating Characteristic) curve is used to analyze and evaluate the effect of binary classification. Generally, the independent variable is a continuous variable, and the dependent variable is a binary variable. Its basic principle is: by moving the judgment point, multiple pairs of sensitivity and error rate are obtained, with sensitivity as the vertical axis and error rate as the horizontal axis, connecting each point to draw a curve, and then calculating the area under the curve. The larger the area, the higher the judgment value. Among them:

[0100] Sensitivity: the probability of judging a value that is actually true as true;

[0101] Specificity: The probability of judging a value to be false when it is actually false.

[0102] False positive rate: the probability of judging a false value as true, which is equal to "1-specificity";

[0103] This method is simple and intuitive. The clinical accuracy of the analysis method can be observed through diagrams, and judgments can be made with the naked eye. The ROC curve combines sensitivity and specificity in a graphical way, which can accurately reflect the relationship between the specificity and sensitivity of a certain analysis method and is a comprehensive representation of the test accuracy. The ROC curve provides an intuitive comparison between different tests under a common scale. The more convex and closer the curve is to the upper left corner, the greater its diagnostic value is, which is conducive to the comparison between different indicators. The drawn curve can also be compared with a straight line with an angle of 45 degrees. If it is close to overlap, it means that the judgment value of the independent variable on the dependent variable is very poor; conversely, if it is farther away from the straight line with an angle of 45 degrees, that is, the larger the area under the curve, it means that the judgment accuracy of the independent variable on the dependent variable is higher.

[0104] from Figure 8 It can be seen that the curve is far away from the 45-degree reference line, and the area surrounded by it occupies most of the area of ​​the image. This shows that the quality discrimination CNN neural network model trained by the process data obtained by the high-temperature alloy turbine disk forging simulation has good discrimination ability. Through the input process data matrix, it can accurately identify whether the quality of most turbine disk forgings is qualified or not. In the ROC diagram, it can also be found that there is a large curve bend at "1-specificity" at 0.27, which means that it can be used as a clearer discrimination point, that is, it is judged as 1 if it is greater than or equal to 0.27, and it is judged as 0 if it is less than 0.27. The specific results of the CNN model discrimination are shown in Table 6.

[0105] Table 6 Discrimination accuracy of CNN model

[0106]

[0107] As can be seen from the table, the CNN discrimination model can accurately judge whether the quality of the final forgings is qualified through process data. Among the 22 sets of input data, only one set of data has a judgment error, and both sets of judgment errors are qualified forgings are judged as unqualified forgings. Therefore, according to the statistical definition, the misjudgment rate and specificity are 0 and 1 respectively.

[0108] From the results, the CNN model only misjudged one set of process data, with a sensitivity of 88.9%, a good specificity of 100%, and an overall accuracy of 95.5%. Fig. 9 As shown, if the critical value for judging whether a product is qualified or not is adjusted from 0.5 to 0.35, the overall judgment accuracy can reach 100%.

[0109] In the actual turbine disk manufacturing industry, the obtained process data can be continuously added to the case library of the die forging knowledge management system, especially the failed cases. These cases continuously enrich the case library, and the trained CNN neural network model can be trained through the continuously updated case library, and continuously refined during the production process, making the judgment more and more accurate. In subsequent use, as the case library is gradually refined, the failed cases can also be subdivided according to the type of defects generated. At this time, the structure of the CNN neural network model can be modified, the output type can be increased, and the model can be evolved. The evolved CNN model can not only judge the quality of the sword, but also determine what causes the quality problems of the forgings.

[0110] Step 6: Based on all the items in the above steps 1 to 5, this step proposes a turbine disk microstructure non-destructive determination system based on a convolutional neural network, which is used to execute the above turbine disk microstructure non-destructive determination method; the system includes a forging monitoring module, a data receiving module, a data processing module, and a determination module;

[0111] The forging monitoring module performs multiple sets of finite element simulations on the key points of turbine disc structure control and extracts and organizes the forging process data;

[0112] The data receiving module is used to receive the input finite element processing data of the key points of the turbine disk structure control and the extracted forging process data, and pass them to the subsequent processing link to provide basic reference data for the entire system;

[0113] The data processing module is used to further process the forging process data, train and build a corresponding convolutional neural network quality discrimination model, and store the corresponding data set to improve the model;

[0114] The determination module inputs the obtained forging process data into the trained convolutional neural network model for microstructure determination and outputs the final result.

Claims

1. A method and system for non-destructive determination of turbine disk microstructure based on convolutional neural network, characterized in that By performing reasonable feature extraction on forging process data through the method, the microstructure condition of key points can be determined without destroying the forgings, and the quality of the forgings can be effectively improved. The construction of the method and system includes the following steps: Step 1: Select the key points in the turbine disc that need to be organized and controlled; Step 2: Establish finite element simulation of the whole process of turbine disk forging for multiple working conditions; Step 3: Propose microstructural requirements for the average grain size and recrystallization fraction of the simulation results, perform artificial classification, and extract the forging process data of each group of simulations; Step 4: Establish and train the convolutional neural network quality discrimination model; Step 5: Import the turbine disk forging process data into the convolutional neural network quality judgment model to judge whether the average grain size and recrystallization fraction of the turbine disk are qualified or not, and verify the accuracy of the model. Step 6: Based on the above-mentioned method for non-destructive determination of turbine disk microstructure based on convolutional neural network, in order to better implement this method, a non-destructive determination system for turbine disk microstructure based on convolutional neural network is constructed. The system mainly includes a forging monitoring module, a data receiving module, a data processing module, and a determination module.

2. The method according to claim 1, characterized in that: The process data that needs to be collected in step 3 include the displacement of the upper beam of the press and the upper die load, lower die load, upper beam speed and the temperature of two observation points on the upper die at that moment. The die forging process data must first be processed uniformly to generate a data structure such as (x1, x2…, x n ,y) matrix, where x i represents a vector of process data, and y represents whether the final forging meets the requirements, which is 1 if it meets the requirements and 0 if it does not. Due to the algorithm requirements, some columns of all 0 are added to complete the matrix, and the final matrix size is 32×32. These matrices will be used as training data to establish the convolutional neural network model.

3. The method according to claim 1, characterized in that: The output value of the quality discrimination convolutional neural network model described in step 4 is not the standard 0 and 1, so 0.5 is selected as the discrimination point, that is, if the result is greater than or equal to 0.5, it is judged as 1, that is, it meets the microstructure requirements; If it is less than 0.5, it is judged as 0, which means it does not meet the requirements.

4. The method according to claim 1, a non-destructive determination system of turbine disk microstructure based on convolutional neural network, characterized in that: The system includes forging monitoring module, data receiving module, data processing module and judgment module, and its applicable scopes are: The forging monitoring module performs multiple sets of finite element simulations on the key points of turbine disc structure control and extracts and organizes the forging process data; The data receiving module is used to receive the input finite element processing data of the key points of the turbine disc structure control and the extracted forging process data, and pass them to the subsequent processing links to provide basic reference data for the entire system; The data processing module is used to further process the forging process data, train and build the corresponding convolutional neural network quality discrimination model, and store the corresponding data set to improve the model; The judgment module inputs the obtained forging process data into the trained convolutional neural network model for microstructure judgment and outputs the final result.

5. The non-destructive determination system of turbine disk microstructure based on convolutional neural network according to claim 4 is characterized in that: In the receiving module, the following target data attributes are selected as data features used by the model: (1) Simulation information, including extracting key point deformation information through finite element simulation of the entire turbine disk forging process; (2) Forging process information, including the specific implementation steps of the forging process and the requirements for various parameters.

6. The non-destructive determination system of turbine disk microstructure based on convolutional neural network according to claim 4 is characterized in that ,In the data processing module, the convolutional neural network quality ,discrimination model can be trained based on the existing forging process ,data, and the corresponding data set can be stored to ,periodically improve the model.

Images