Precision casting process blade size evolution prediction method based on data driving

Abstract

A complete investment casting experiment is carried out around an autonomously designed turbine blade mold, S1, S2 and S3 section sizes of a casting in different stages are collected, and a time sequence data set is established, including mold, wax pattern, directional solidification, mold shell removal, pouring system removal and ceramic core constraint removal. Analysis shows that the average deviation of the casting size evolves continuously in different stages (from 0.2134 mm to 0.1857 mm), deviation accumulation is achieved, and the influence degrees of all the stages are different. In order to effectively capture the changes of the casting, an LSTM neural network is adopted to mine size deep features and establish a time-varying model. The result shows that the LSTM considers the deformation history by virtue of the unique door structure, the prediction value of the average deviation of the casting in a single-step prediction task is 0.1832 mm, and the test average absolute percentage error (MAPE) is 1.35%; the MAPE of the multi-step prediction is 2.99%, and the MAPE of the final step is 6.63%. The research provides a new thought for intelligent development of the field of investment casting.

Description

Technical Field

[0001] The present invention relates to the research field of casting size evolution prediction in precision casting process, and in particular to a data-driven method for predicting blade size evolution in precision casting process. Background Art

[0002] Nickel-based superalloys are widely used in key hot-end components of aero-engines and industrial gas turbines, especially first- and second-stage turbine blades, due to their excellent high-temperature properties such as creep resistance, fatigue resistance and corrosion resistance. Investment casting, as a near-net-shape manufacturing technology, can achieve mass production of thin-walled, complex parts with controlled grains ranging from a few grams to several kilograms and with little subsequent processing requirements. However, the dimensional accuracy of turbine blades is affected by multiple factors, including molds, wax patterns, directional solidification, and subsequent mold shells, gating systems, and ceramic core constraint removal stages. Therefore, it is crucial to deeply understand the dimensional evolution of nickel-based superalloys at each stage of precision casting and develop accurate prediction models to improve the quality of castings.

[0003] Due to the geometric complexity of parts and their dynamic changes during the solidification process, establishing the relationship between deformation and influencing factors has become the core difficulty of size prediction. At present, traditional casting deformation prediction methods mainly rely on finite element simulation, but these methods have limited adaptability to complex blade structures, high computational costs, and difficulty in capturing the nonlinear deformation laws caused by multi-field coupling effects. In recent years, data-driven methods have performed outstandingly in revealing the potential correlation of large-scale casting data, and can be used for decision support, process diagnosis and quality improvement. Based on this, this study combined a large number of experiments and numerical simulations to construct a time series casting data set, and used the long short-term memory (LSTM) network to mine deep time series features, and proposed a data-driven turbine blade size evolution prediction method. The purpose is to accurately predict the final size of the casting and its evolution process, improve the quality of the casting, and shorten the R&D cycle. Summary of the invention

[0004] A complete investment casting experiment was carried out around a self-designed turbine blade mold. The cross-sectional dimensions of S1, S2, and S3 at different stages of the casting were collected and a time-series dataset was established, including the mold, wax pattern, directional solidification, mold shell removal, gating system removal, and core restraint removal. The analysis shows that the average dimensional deviation of the casting evolves continuously at different stages (from 0.2134 mm to 0.1857 mm), the deviation accumulates, and the influence degree is different at each stage. To effectively capture these changes in the casting, we used an LSTM neural network to mine the deep features of the dimensions and establish a time-varying model. The results show that LSTM, with its unique gate structure considering the deformation history, has a predicted value of 0.1832 mm for the average deviation of the casting in the single-step prediction task, and the test mean absolute percentage error (MAPE) is 1.35%; the MAPE for multi-step prediction is 2.99%, and the MAPE for the final step is 6.63%. This study provides a new idea for the intelligent development of the investment casting field.

[0005] The technical solution for the present invention to achieve the above object includes the following steps:

[0006] Step 1:

[0007] Design and manufacture a mold, and complete six stages t1 - t6 of precision casting, including the mold, wax pattern, directional solidification, mold shell removal, gating system removal, and core restraint removal. Since the casting is wrapped by the mold shell during the directional solidification stage, an attempt was made to measure the internal dimensions of the casting using CT scanning, but a large number of artifacts made it difficult to accurately reconstruct the true contour of the casting. Therefore, numerical simulation results were used to evaluate the dimensional deviation of the casting at the directional solidification stage t3.

[0008] Step 2:

[0009] Use a non-contact blue light scanner to collect the surface contour data of the casting in five stages except for directional solidification where P i (t k ) = {p i1 , p i2 ,..., p in}; i is the casting number, k is the discrete moment of the casting stage, and n represents the total number of measurement points. Before measurement, the surface to be measured needs to be cleaned with anhydrous ethanol, titanium powder is sprayed to reduce reflection, and reference points are pasted on the surface of the casting for alignment during scanning. Register the casting models at different stages with the design model Calculate the deviation D i (t k ) = P i (t k ) - P design .

[0010] Step 3:

[0011] Carry out directional solidification numerical simulation, with the boundary conditions and process parameters being consistent with the experiment. Wait for the casting to cool to room temperature and export the deformed casting *.stl. Similar to step 2, obtain the deviation D3(t k ) at the current stage, and finally obtain the set of time-series deviations D(t k ) = {D1(t k ), D2(t k ), D3(t k ), D4(t k ), D5(t k ), D6(t k )} at different stages of all castings.

[0012] Step 4:

[0013] Divide the time-series data set into a training set and a test set according to a certain ratio. Each data set must contain data of different castings at different cross-sections and be normalized to [-1 1]. After completing the data preprocessing, start building an LSTM network, set the network depth and hyperparameters (such as learning rate, Dropout), and use the mean squared error MSE as the loss function.

[0014] Step 5:

[0015] The flow chart of the data-driven prediction method for the blade size evolution in the precision casting process is shown in the following table.

[0016]

[0017] Description of the Drawings

[0018] Figure 1 is the flow chart of a data-driven prediction method for the blade size evolution in the precision casting process of the present invention;

[0019] Figure 2 is the schematic diagram of the five-group precision casting experiment process of one mold in an embodiment of the present invention;

[0020] Figure 3 is the schematic diagram of the deviation between the wax mold and the casting at the final stage in an embodiment of the present invention;

[0021] Figure 4 is the schematic diagram of the deviation of the casting in the directional solidification stage in an embodiment of the present invention;

[0022] Figure 5 is the schematic diagram of the deviation of the predicted final step size of the casting in an embodiment of the present invention;

[0023] Figure 6 is the schematic diagram of the multi-step size evolution of the predicted casting in an embodiment of the present invention. Detailed implementation manners

[0024] To more clearly illustrate the specific implementation manners of the present invention, the present invention will be described in detail in combination with the accompanying drawings and specific implementation manners. However, the protection scope of the present invention is not limited to the following embodiments. Obviously, the accompanying drawings in the following description are only partial implementation manners of the present invention. For those of ordinary skill in the art, other accompanying drawings can also be obtained according to the present invention without creative efforts.

[0025] Taking the investment casting of turbine blades as an example, the specific implementation process of the present invention is as Figure 1 shown:

[0026] Step 1

[0027] We designed the molds required for this experiment and commissioned Huamaoruifeng Precision Mould Technology Co., Ltd. to process the cavities by CNC. The mold material was selected as high-hardness 6061 aluminum alloy. The wax patterns were made by a ZH-SCW-50 type wax injection machine. The selector was prepared by advanced 3D printing technology and adhered to the wax patterns with low-temperature red wax. Subsequently, a ZY-ZJ-800 type sand spraying machine was used to repeatedly sprinkle sand and apply slurry to make the mold shell, and the pouring and subsequent deconstraint experiments were completed.

[0028] Figure 2 It is a schematic diagram of the process of a five-group precision casting experiment in one mold of an implementation manner of the present invention;

[0029] Step 2

[0030] The three-dimensional scanning measurement of the castings at different stages was carried out using the ATOS 5 blue light scanner produced by Carl Zeiss of Germany to obtain the three-dimensional models of the castings in.stl format at each stage. Except for the directional solidification stage, the surface profile data of the other five stages were recorded where P i (t k ) = {p i1 , p i2 ,..., p i1200}, i represents the 1st - 5th castings, k is the discrete moments of the five stages of the precision casting process, and 1200 is the total number of measurement points. The measurement point cloud of the casting at each stage was registered with the design model to calculate the casting deviation.

[0031]

[0032] where P design is the corresponding point of the design model. R is the rotation matrix, and T is the translation vector.

[0033] D i (t k ) = P i (t k ) - Pdesign (2)

[0034] The deviation sets of all measurement points of five castings (#1 - #5) at five moments are as follows.

[0035] D(t k ) = {D1(t k ), D2(t k ), D4(t k ), D5(t k ), D6(t k )}(3)

[0036] Figure 3 It is a schematic diagram of the deviation between the wax pattern and the casting in the final stage of an embodiment of the present invention;

[0037] Step 3

[0038] Carry out directional solidification numerical simulation, with its boundary conditions and process parameters consistent with the experiment. After the casting cools to room temperature, export the deformed 3D model in.stl format. Subsequently, similar to Step 2, register the model and calculate the dimensional deviation at the current stage. Finally, obtain the time - series deviation sets of the five castings at six stages

[0039]

[0040] Figure 4 It is a schematic diagram of the deviation of the casting in the directional solidification stage of an embodiment of the present invention;

[0041] Step 4

[0042] Divide the time - series data set into a training set and a test set according to 8:2. Each data set must contain data of different castings at different cross - sections After completing data pre - processing, start training using the built LSTM network. The learning rate is 0.001, and the mean squared error MSE is used as the loss function.

[0043] Figure 5It is a schematic diagram for predicting the final step size deviation of a casting in an embodiment of the present invention. As can be seen from the figure, for the first type of task, that is, the prediction of the final time step of the casting size, a total of 1195 test samples were evaluated. Among them, the S1, S2, and S3 cross-sections contained 388, 386, and 421 samples respectively. The deviation between the predicted and actual values is mainly concentrated around 0.18 mm. Among them, the size predictions of S1, S2, and S3 are concentrated in the ranges of 0.0927 mm - 0.2837 mm, 0.1099 - 0.2524 mm, and 0.1372 mm - 0.2250 mm respectively. The RMSE of S2 and S3 is 0.04 mm, and the RMSE of S1 is 0.07 mm. From the fitting equation, there is a strong linear relationship between the predicted value and the actual value. The average deviation evolved from the initial 0.2134 mm to the final 0.1857 mm. The overall average predicted value and actual value are 0.1832 mm and 0.1857 mm respectively, and the relative error is 1.35%.

[0044] Figure 6 It is a schematic diagram for predicting the multi-step size evolution of a casting in an embodiment of the present invention. As can be seen from the figure, for the second type of task, that is, predicting the size of the blade at multiple time steps during the investment casting process, we used 1195 test samples, and each sample contained prediction data for 5 time steps, with a total of 5975 data points. Among them, the S1, S2, and S3 cross-sections had 2065, 1945, and 1965 data points respectively. In this experiment, the actual and predicted values of the test samples at different time steps were between -0.3 mm and 0.4 mm. The average values of the two were 0.1753 mm and 0.1807 mm respectively, and the relative error was 2.99%. The relative error at the final step was 6.63%, indicating that the model is good at capturing the evolution characteristics of the casting size during the entire investment casting process.

Claims

1. A data-driven prediction method for the size evolution of blades in the precision casting process, characterized by the following steps: Step 1: Design and manufacture a mold, and complete the six stages t1 - t6 of precision casting, including the mold, wax pattern, directional solidification, removal of the mold shell, removal of the gating system, and removal of the ceramic core constraint. Since the casting is wrapped by the mold shell during the directional solidification stage, an attempt is made to measure the internal dimensions of the casting using CT scanning, but a large number of artifacts make it difficult to accurately reconstruct the true contour of the casting. Therefore, the numerical simulation results are used to evaluate the dimensional deviation of the casting at the directional solidification stage t3. Step 2: Use a non-contact blue light scanner to collect the surface contour data of the casting in the five stages except directional solidification, where i is the casting number, k is the discrete moment of the casting stage, and n represents the total number of measurement points. Before measurement, the surface to be measured needs to be cleaned with anhydrous ethanol, titanium powder is sprayed to reduce reflection, and reference points are pasted on the surface of the casting for alignment during scanning. Register the casting models at different stages with the design model and calculate the deviation of the measurement points. Step 3: Carry out numerical simulation of directional solidification, with its boundary conditions and process parameters consistent with the experiment. After the casting is cooled to room temperature, export the deformed casting *.stl. Similar to Step 2, obtain the deviation at the current stage, and finally obtain the set of time-series deviations of all castings at different stages. Step 4: Divide the time-series data set into a training set and a test set according to a certain ratio. Each data set must contain data of different castings at different cross-sections and be normalized to [-1 1]. After completing the data preprocessing, start building an LSTM network, set the network depth and hyperparameters (such as learning rate, Dropout), and use the mean square error MSE as the loss function. Step 5: The flow of the data-driven prediction method for the size evolution of blades in the precision casting process is shown in the following table.

Images