High-temperature alloy bolt creep characteristic prediction method based on load fluctuation

Abstract

The invention provides a high-temperature alloy bolt creep characteristic prediction method based on load fluctuation, and the method comprises the following steps: collecting extension data and tension data of a bolt test piece under target temperature and load, and obtaining a tensile curve and a tension curve; the extension data, the tension data, the tensile curve and the tension curve of the bolt test piece are processed, and load fluctuation and a true value used for calculating loss are determined; the load fluctuation, the cross section area of the test piece and the set load are input into a preset deep learning model, a prediction result is output, the prediction result comprises a prediction curve, and the prediction curve is used for representing the creep characteristic trend of the bolt. According to the method, the load fluctuation can be quickly measured by cheaper equipment, so that the use cost of the model in a prediction stage is remarkably reduced, the required training data is less, the creep curve of the thread test piece can be accurately predicted by training the creep test data of less test pieces, and the method is suitable for popularization and application. And the predicted curve is consistent with a true value in length and trend.

Description

Technical Field

[0001] The present invention relates to the technical field of display devices, and in particular to a prediction method for the creep characteristics of superalloy bolts based on load fluctuation. Background Art

[0002] For the superalloy bolts used in aeroengines, creep is a plastic deformation that occurs at high temperatures and under loads lower than the yield strength. The creep of bolts may cause a decrease in the pre-tightening force and the fracture of bolts, which will seriously endanger the safe operation of equipment.

[0003] Traditional creep life prediction methods require full-scale creep experiments on specimens with the same specifications as the specimens to be predicted before prediction to obtain the data required for modeling. Most of them need to use specimens with the same specifications as the threaded specimens to be predicted for the full-scale creep experiment to obtain data to determine the parameters in the creep prediction model. The full-scale creep experiment often takes up to dozens of hours;

[0004] Once the thread specifications of the specimens change, the above steps need to be carried out again, and it is difficult to use the existing data for accurate prediction;

[0005] In both the modeling stage and the prediction stage, an extensometer is required to measure the elongation of the specimen. Whether a non-contact or contact extensometer is used, it needs to be carried out in a laboratory environment, and it is difficult to conduct rapid tests in actual use scenarios. Summary of the Invention

[0006] Technical Problem to be Solved

[0007] In view of the deficiencies of the prior art, the present invention provides a prediction method for the creep characteristics of superalloy bolts based on load fluctuation to solve the problems raised in the above background art.

[0008] Technical Solution

[0009] To achieve the above objectives, the present invention is achieved through the following technical solutions: A prediction method for the creep characteristics of superalloy bolts based on load fluctuation, comprising the following steps:

[0010] Collect the extensometer data and tensile force data of the bolt specimen at the target temperature and load to obtain a tensile curve and a tensile force curve;

[0011] Process the extensometer data, tensile force data, tensile curve and tensile force curve of the bolt specimen, and determine the load fluctuation and the true value for calculating the loss;

[0012] Input the load fluctuation, the cross-sectional area of the specimen and the set load into a preset deep learning model, and output a prediction result, where the prediction result includes a prediction curve, and the prediction curve is used to characterize the trend of the bolt creep characteristics.

[0013] Preferably, in this embodiment, the process of collecting the extensometer data and tensile data of the bolt specimen under the target temperature and load includes:

[0014] Testing with a creep testing machine that complies with the HB5151-96 standard;

[0015] All of the threaded area of the bolt specimen is located within the area measured by the extensometer;

[0016] During collection, first heat up to the target temperature and then apply the load. After applying the load to the target load, maintain the pressure;

[0017] Among them, the experimental data of a single specimen includes two parts: the force sensor reading and the extensometer reading.

[0018] Preferably, in this embodiment, during the specimen sampling process, it also includes: taking one sampling point per second during the loading stage, taking one sampling point every 10 minutes from the end of loading to 1 hour, and then taking one sampling point every half hour.

[0019] Preferably, in this embodiment, the process of determining the load fluctuation includes:

[0020] Intercept the rising stage of the tensile curve and calculate the load fluctuation data according to the formula;

[0021] The calculation formula is:

[0022] Load fluctuation = force sensor reading - loading rate * time value of this sampling point;

[0023] Zero-fill the right side of all the load fluctuation data sequences to a length of 256. At the same time, collect the cross-sectional area of the specimen and the target load value (unit: N), and then perform maximum-minimum normalization on them and the load fluctuation data in the batch size dimension.

[0024] Preferably, in this embodiment, the process of calculating the true value of the loss includes:

[0025] Extract the elongation-time series corresponding to the ET curve and use the least squares method to find the parameters therein;

[0026] The calculation formulas include:

[0027] ;

[0028] Among them, εPM is the specimen elongation, t is the time, and θ1~θ4 are the parameters to be determined;

[0029] Based on the parameters and the true fracture time of the bolt, obtain the true value.

[0030] Preferably, in this embodiment, the training process of the deep learning model includes:

[0031] Obtain the load fluctuation, the cross-sectional area of the specimen, and the set load, analyze the relationship among the three, and extract the input features;

[0032] Input the input features of the load fluctuation into the linear layer of the neural network model;

[0033] Among them, the last two dimensions of the linear layer are swapped with each other;

[0034] Extract and reduce the dimension of the input features of the load fluctuation through the encoder;

[0035] Input the input features of the cross-sectional area of the specimen and the set load into the convolutional neural network model for linear transformation to obtain the predicted θ1~θ4 and the fracture time;

[0036] Judge whether the root mean square error between the predicted θ1~θ4 and the fracture time and the true value decreases;

[0037] If so, continue training after the model converges;

[0038] If not, determine the predicted θ1~θ4 and the fracture time as the target true value.

[0039] Beneficial effects

[0040] The present invention provides a method for predicting the creep characteristics of superalloy bolts based on load fluctuations, having the following beneficial effects:

[0041] Speculate the creep curve of specimens with similar size specifications based on the existing creep test data, effectively reducing the cost;

[0042] Moreover, the present invention requires less training data, and only uses the creep test data of fewer specimens to train and more accurately predict the creep curve of the threaded specimens, and the predicted curve fits the true value in terms of length and trend;

[0043] The load fluctuation can be quickly measured by a cheaper device, thus significantly reducing the usage cost of the model in the prediction stage;

[0044] While ensuring the effect of feature extraction, the present invention can achieve smaller input and output, which reduces the workload of the downstream network of the model, and the entire model can be more lightweight to reduce the required training data. The traditional Transformer network is designed to extract information from text. In this way, the network pays more attention to the relationship between a token corresponding to a word and all tokens. If it is forcibly used to analyze time series data such as the load fluctuation data in this patent, the model pays more attention to the relationship between the token corresponding to the time value and the fluctuation value at each sampling point and the tokens at other sampling points. However, iTransformer swaps the last two dimensions of the input based on Transformer. In this way, the model pays more attention to the relationship between the tokens corresponding to the time values of all sampling points and the tokens corresponding to the fluctuation values of all sampling points. In this way, it is more in line with the analysis logic of time series, and the model can also analyze the time series more comprehensively. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a diagram of the collected data style of an embodiment of the present invention;

[0046] Figure 2 It is a diagram of the collected load fluctuation data style of an embodiment of the present invention;

[0047] Figure 3 It is a flowchart of the training process and prediction process of the deep learning model of the present invention;

[0048] Figure 4 It is a specific usage process of an embodiment of the present invention and the prediction results of specific calculation examples;

[0049] Figure 5 It is a structural block diagram of the deep learning model of the present invention;

[0050] Figure 6 It is a flowchart of the method for predicting the creep characteristics of superalloy bolts based on load fluctuations of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0051] The following details the embodiments of the present invention. The examples of the embodiments are shown in the drawings, where the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions from beginning to end. The embodiments described below with reference to the drawings are exemplary and are only used to explain the present invention and should not be construed as a limitation of the present invention.

[0052] The following disclosure provides many different embodiments or examples for implementing different structures of the present invention. To simplify the disclosure of the present invention, the components and settings of specific examples are described below. Of course, they are only examples and are not intended to limit the present invention. In addition, the present invention may repeat reference numerals and / or reference letters in different examples. Such repetition is for the purpose of simplification and clarity, and does not itself indicate the relationship between the various embodiments and / or settings discussed. In addition, the present invention provides examples of various specific processes and materials, but those of ordinary skill in the art can be aware of the application of other processes and / or the use of other materials.

[0053] As Figure 1 shown, an embodiment of the present invention provides a method for predicting the creep characteristics of a superalloy bolt based on load fluctuation, including the following steps:

[0054] Step 1: Collect the extensometer data and tensile force data of the bolt specimen under the target temperature and load to obtain the tensile curve and the tensile force curve;

[0055] Among them, the original data can be measured by a creep testing machine that complies with the HB5151-96 standard;

[0056] Exemplarily, for a rolled thread specimen, it should be ensured that the measurement area of the extensometer is entirely within the thread area. During the experiment, first heat up to the target temperature and then apply the load. After loading to the target load, hold the pressure. The experimental data of a single specimen includes two parts: the force sensor reading and the extensometer reading.

[0057] As Figure 1 shown, FT is the force sensor reading, ET is the extensometer reading, the data comes from the Φ10 standard specimen in HB5151-96, the experimental temperature is 650 °C, the load is 480 Mpa, the loading rate is 100 N / s, one sampling point is taken per second during the loading stage, one sampling point is taken every 10 minutes from the end of loading to 1 hour, and then one sampling point is taken every half hour.

[0058] Step 2: Process the extensometer data, tensile force data, tensile curve and tensile force curve of the bolt specimen, and determine the load fluctuation and the true value used for calculating the loss;

[0059] Specifically, the load fluctuation is extracted from the Figure 1 FT curve in. This index first climbs with a stable trend and then remains unchanged. Intercept its climbing stage and calculate the load fluctuation value according to the formula;

[0060] The calculation formula is:

[0061] Load fluctuation = force sensor reading - loading rate * time value of this sampling point;

[0062] As shown Figure 2 in the left figure, the left figure shows the general causes of load fluctuations, and the right figure shows the load fluctuation data pattern. The data is a bivariate sequence with a length of approximately 200 (variable 1 is the fluctuation value, and variable 2 is the time value corresponding to the sampling point). Note: The error of the force sensor is generally about 5N.

[0063] Zero-fill the right side of all load fluctuation data sequences to a length of 256. At the same time, collect the cross-sectional area of the specimen and the target load value (unit: N), and then perform max-min normalization on them and the load fluctuation data in the batch size dimension.

[0064] In addition to the input of the deep learning model, it is also necessary to determine the true value used for calculating the loss during model training. This true value is extracted from the ET curve. The parameters of the elongation-time series corresponding to the ET curve are obtained by the least squares method, and these parameters and the true fracture time of the specimen are used as the true values during model training.

[0065] The formula for calculating the parameters is:

[0066] ;

[0067] Among them, the θ function method equation, where εPM is the specimen elongation, t is the time, and θ1 to θ4 are the parameters to be determined;

[0068] Among them, θ1 and θ2 describe the first stage of creep, that is, adjusting these two parameters can adjust Figure 1 the first inflection point of the ET curve in Figure 1 , and θ3 and θ4 describe the third stage of creep, that is, adjusting these two parameters can adjust

[0069] the second inflection point of the ET curve in

[0070] It should be noted that the second stage of creep is the equilibrium state of the first and third stages.

[0071] As shown Figure 3 and 4 in, the verification mode of the deep learning model does not update the gradient. When outputting the prediction result, the output needs to be inverse-transformed by max-min normalization (the maximum and minimum values are both from the training set) and then input into the θ function method equation.

[0072] Among them, in Figure 4 , the solid line with the suffix 'true' is the measured time-elongation curve (creep curve), and the dashed line with the suffix 'pred' is the prediction curve.

[0073] Moreover, by using the creep data of known Φ10 standard specimens, M10 threaded specimens, and Φ5 smooth rod specimens, the creep curve of the M5 threaded specimen can be inferred based on the existing data and the load fluctuations rapidly measured using an inexpensive force sensor gasket for the M5 threaded specimen.

[0074] Furthermore, the training process of the deep learning model includes:

[0075] Obtain the load fluctuations, the cross-sectional area of the specimen, and the set load, analyze the relationship among the three, and extract the input features;

[0076] Input the input features of the load fluctuations into the linear layer of the neural network model;

[0077] Among them, the last two dimensions of the linear layer are swapped with each other;

[0078] Extract and reduce the dimension of the input features of the load fluctuations through an encoder;

[0079] Input the input features of the cross-sectional area of the specimen and the set load into the convolutional neural network model for linear transformation to obtain the predicted θ1~θ4 and the fracture time;

[0080] Judge whether the root mean square error between the predicted θ1~θ4 and the fracture time and the true value has decreased;

[0081] If so, continue training after the model converges;

[0082] If not, determine the predicted θ1~θ4 and the fracture time as the target true values.

[0083] As Figure 5 shown, the layers starting with Emb and L are single-layer linear (Liner) layers, and the layers starting with C are one-dimensional convolutional layers (1D-CNN). Among them, the settings of C3, 4, 5 (kernel size = 8, stride = 2, padding = 6) are different from the settings of other one-dimensional convolutional layers (kernel size = 1, and the rest are default). LN is layer normalization (Layer Normalization), Gelu is the gelu activation function, ABS is to take the absolute value, and the global dropout is set to 0.1.

[0084] It should be noted that Figure 5 the parameters of each layer are as follows in the table:

[0085]

[0086] Specifically, a deep learning model composed of a convolutional neural network and an iTransformer network is used to analyze the force fluctuations during the loading process to predict the parameters in the θ-function method equation and the fracture time of the specimen. Finally, the predicted creep curve is plotted using this equation and the fracture time. This method can build a model based on existing data and only requires the force fluctuations during the loading stage of the target specimen for prediction. This force fluctuation can be measured by the currently mass-produced and inexpensive force sensor gasket in just a few minutes, which means that this method can build a model with the data accumulated in the laboratory and make predictions based on the data quickly measured in a non-laboratory environment. The predicted creep curve can provide more reference information when determining the bolt life.

[0087] The method for predicting the creep characteristics of superalloy bolts based on the loading load fluctuation provided by the present invention infers the creep curve of specimens with similar size specifications based on the existing creep test data. However, the traditional method must use an extensometer threaded specimen to conduct a time-consuming full-process creep experiment to determine its creep curve. The extensometer is much more expensive than the force sensor and has more usage conditions restrictions, so the existing data cannot be fully utilized.

[0088] Moreover, the data required for training in the present invention is less. Only using the creep test data of a few specimens for training can accurately predict the creep curve of the threaded specimen, and the predicted curve coincides with the true value in terms of length and trend.

[0089] The load fluctuation can be quickly measured by a more inexpensive device, thus significantly reducing the usage cost of the model during the prediction stage.

[0090] While ensuring the feature extraction effect, the present invention can achieve smaller input and output, which reduces the workload of the downstream network of the model. The entire model can be more lightweight to reduce the required training data. The traditional Transformer network is designed to extract information from text. In this way, the network pays more attention to the relationship between a token corresponding to a word and all tokens. If it is forced to be used to analyze time series data such as the load fluctuation data in this patent, the model pays more attention to the relationship between the token corresponding to the time value and the fluctuation value at each sampling point and the tokens at other sampling points. However, the iTransformer swaps the last two dimensions of the input based on the Transformer. In this way, the model pays more attention to the relationship between the tokens corresponding to the time values of all sampling points and the tokens corresponding to the fluctuation values of all sampling points. In this way, it is more in line with the analysis logic of time series, and the model can also analyze this time series more comprehensively.

[0091] Although embodiments of the present invention have been shown and described, those of ordinary skill in the art will appreciate that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting creep characteristics of high-temperature alloy bolts based on loading load fluctuations, characterized in that: The following steps are involved: Collect the extension data and tension data of the bolt specimen under the target temperature and load to obtain the tensile curve and tension curve; Process the extensibility data, tension data, tensile curves and tension curves of bolt specimens and determine the load fluctuation and true value for calculating losses; The load fluctuation, the cross-sectional area of ​​the specimen and the set load are input into a preset deep learning model, and a prediction result is output. The prediction result includes a prediction curve, and the prediction curve is used to characterize the creep characteristic trend of the bolt.

2. The method for predicting creep characteristics of high-temperature alloy bolts based on loading load fluctuation according to claim 1 is characterized in that: The process of collecting the extension data and tension data of the bolt specimen under the target temperature and load includes: Passed the creep testing machine test in accordance with HB5151-96 standard; The threaded area of ​​the bolt specimen is entirely within the extensometer measurement area; When collecting, first heat up to the target temperature and then load, and then maintain the pressure after loading to the target load; Among them, the experimental data of a single specimen includes two parts: force sensor reading and extensometer reading.

3. The method for predicting creep characteristics of high-temperature alloy bolts based on loading load fluctuation according to claim 2 is characterized in that: The specimen sampling process also includes: taking a sampling point per second during the loading phase, taking a sampling point every 10 minutes until 1 hour after loading is completed, and taking a sampling point every half hour thereafter.

4. The method for predicting creep characteristics of high-temperature alloy bolts based on loading load fluctuation according to claim 1 is characterized in that: The process of determining load fluctuation includes: Intercept the rising stage of the tension curve and calculate the load fluctuation data according to the formula; The calculation formula is: Load fluctuation = force sensor reading - loading rate * time value of the sampling point; All load fluctuation data sequences are padded to a length of 256 with zeros on the right. At the same time, the cross-sectional area and target load value (unit: N) of the specimen are collected and then normalized with the load fluctuation data in the batch size dimension to the maximum-minimum value.

5. The method for predicting creep characteristics of high-temperature alloy bolts based on loading load fluctuation according to claim 4 is characterized in that: The process of calculating the true value of the loss includes: Extract the elongation-time series corresponding to the ET curve and use the least square method to find the parameters; The calculation formula includes: ； Among them, εPM is the elongation of the specimen, t is the time, and θ1~θ4 are parameters to be determined; Based on the parameters and the real fracture time of the bolt, a true value is obtained.

6. The method for predicting creep characteristics of high-temperature alloy bolts based on loading load fluctuation according to claim 5, characterized in that: The training process of the deep learning model includes: Obtain load fluctuation, specimen cross-sectional area and set load, analyze the relationship between the three, and extract input features; Input features of load fluctuations are fed into the linear layer of the neural network model; Wherein, the last two dimensions of the linear layer are swapped with each other; The input features of load fluctuations are extracted and reduced in dimension through the encoder; The input features of the specimen cross-sectional area and the set load are input into the convolutional neural network model and linearly transformed to obtain the predicted θ1~θ4 and fracture time; Determine whether the root mean square error between the predicted θ1~θ4 and the fracture time and the true value decreases; If so, continue training after the model converges; If not, the predicted θ1~θ4 and fracture time are determined as the target true values.

Images