A double-wall vane flow prediction system based on hole parameter convolutional neural network

Abstract

The application relates to a double-wall blade flow prediction system based on a hole parameter convolutional neural network, which comprises the following modules: an input module for receiving user input structure parameters and pressure parameters; a parameter preprocessing module for processing the structure parameters and the pressure parameters into tensor form to obtain structure parameter tensors and pressure difference tensors; a flow prediction module for calling a trained hole parameter convolutional neural network and taking the structure parameter tensors and the pressure difference tensors as network inputs to predict gas film hole cold gas flow tensors and impact hole cold gas flow tensors; and an output module for converting the predicted gas film hole and impact hole cold gas flow tensors into a general form of characters and parameters and displaying and outputting the characters and parameters through a UI interface. The application realizes rapid and accurate prediction of the cold gas flow of each hole of a double-wall blade under different internal and external pressure conditions by calling a previously trained convolutional neural network, and simultaneously considers the calculation speed and accuracy.

Description

Technical Field

[0001] This invention relates to the field of aero-engine technology, and in particular to a flow prediction system for double-walled blades based on a hole parameter convolutional neural network. Background Technology

[0002] To improve the thrust of aero-engines, turbine inlet gas temperatures have been continuously rising, currently exceeding the allowable temperature of 500K for turbine blade materials. In next-generation advanced aero-engines, this gap may exceed 700K, necessitating efficient and reliable blade cooling technologies. Turbine blade cooling is mainly divided into two categories: internal cooling and external cooling. Internal cooling utilizes low-temperature airflow for convective heat transfer within the blade, absorbing heat from the inner side to reduce blade temperature; this primarily includes convection cooling and impingement cooling. External cooling involves discharging low-temperature airflow through discrete holes or slots, forming an air film on the outer surface of the blade to isolate direct heat exchange between the high-temperature gas and the blade, thereby reducing the blade's thermal load; this mainly includes film cooling and evaporative cooling. Single cooling methods such as convection, impingement, and film cooling have reached their cooling capacity limits after decades of development and cannot meet the ever-increasing turbine inlet temperatures. Evaporative cooling technology uses porous media materials, ensuring sufficient contact between the cold air and the blade to form a dense and uniform cold air film, thus possessing the strongest heat exchange capacity. However, it suffers from problems such as easy clogging and high thermal stress, and has not yet reached the practical application stage. Laminated plate cooling (also known as double-wall cooling) combines multiple cooling methods such as convection, impingement, and film cooling. It also boasts a large heat exchange area, resulting in high heat exchange efficiency and cooling effect. Furthermore, compared to evaporative cooling, it has lower manufacturing costs and higher reliability in practical applications. Therefore, double-wall cooling is expected to become the basic cooling structure for future advanced turbine blades.

[0003] The basic structure of a double-walled turbine blade includes an inner wall with impingement holes, an outer wall with film cooling holes, and interlayer turbulence pillars connecting and supporting the inner and outer walls. Accurately evaluating the coolant flow distribution of each impingement hole and film cooling hole is crucial for achieving refined cooling design of double-walled turbine blades. Currently, the evaluation of coolant flow in each hole of a double-walled turbine blade mainly relies on computational fluid dynamics (CFD) numerical simulation methods and fluid network calculation methods. CFD methods can simulate the entire flow field inside and outside the blade, thus obtaining complete flow information and details. However, due to the complex structure and small geometry of the double-walled blade, a large number of dense, high-quality meshes are required to achieve higher simulation accuracy, resulting in significant computational resource and time consumption. The fluid network method is a reduced-order calculation model that simplifies the actual flow path to a few nodes and interconnected flow resistance elements, thus retaining the main characteristics of the flow process while also offering extremely fast computation speed. However, the fluid network method has relatively low accuracy when applied to double-walled blades because the interlayer channels of the double-walled structure are interconnected and communicate with numerous impact holes and film vents. Therefore, the gas flow inside the double-walled blade cannot be separated, making it difficult to clearly delineate flow nodes and extract flow resistance elements. Furthermore, under actual operating conditions, significant pressure gradients exist along both the flow direction and radial direction on the outer surface of the turbine blade. This causes cool gas to migrate and flow towards the low-pressure outlet vents within the interlayer between the inner and outer walls of the blade, further complicating the calculation using the fluid network method and reducing its accuracy and applicability.

[0004] Therefore, given the known internal and external pressure conditions of double-walled blades, how to quickly and accurately calculate the cooling air flow rate of all impact holes and film cooling holes for different cooling structures is a technical problem that urgently needs to be solved. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a flow prediction system for double-walled blades based on a convolutional neural network with aperture parameters. This system enables rapid and accurate prediction of the cold air flow rate in each aperture of double-walled blades with different structures under varying internal and external pressure conditions by calling a pre-trained convolutional neural network, while simultaneously balancing computational speed and accuracy.

[0006] The present invention adopts the following technical solution:

[0007] A flow prediction system for double-walled blades based on a pore parameter convolutional neural network includes:

[0008] The input module is used to receive the structural and pressure parameters of the double-walled blade to be calculated, which are input by the user through the UI interface. The structural parameters include N parameters for each impact hole diameter, film pore diameter, hole angle, hole spacing, and row spacing of the double-walled blade, as well as the number of holes in a single row n1 and the number of hole rows n. r The pressure parameters include the inlet pressure of each impact hole and the outlet pressure of each film gas hole;

[0009] The parameter preprocessing module is used to process the structural parameters and pressure parameters transmitted from the input module into tensor form, obtaining tensors of size N·n1·n. r The structural parameter tensor and its size are 1·n1·n r The pressure difference tensor is obtained, and the structural parameter tensor and the pressure difference tensor are stored and transmitted to the flow prediction module respectively;

[0010] The flow prediction module is used to call a pre-trained orifice parameter convolutional neural network, and to perform calculations using the structural parameter tensor and the pressure difference tensor as input parameters to the orifice parameter convolutional neural network, predicting that the size of each is 1·n1·n. r The flow rate tensors of the film cooling orifice and the impingement orifice cooling air are obtained, and the two flow rate tensors are transmitted to the output module;

[0011] The output module is used to convert the predicted cold air flow tensor of the film gas orifice and the cold air flow tensor of the impact orifice into a general text and parameter form, and to display the text and parameters through a UI interface.

[0012] The beneficial effects of this invention are:

[0013] In terms of computational speed, this invention achieves extremely high speed because the training of the convolutional neural network is pre-completed. Therefore, when faced with new double-walled structures and internal / external pressure conditions, the prediction system can directly call the pre-trained convolutional neural network for pore parameter calculations, eliminating the need for iterations. Regarding computational accuracy, firstly, based on the understanding of physical flow, the pore flow rate is only related to the structural flow resistance and the internal / external pressure difference. Structural parameters and pressure difference parameters are selected as input parameters to ensure physical reliability. Secondly, considering that the multi-connectivity of the double-walled structure means that the flow rate of a single pore is affected by the flow of surrounding pores, and that the physical mechanisms of this effect are similar, a convolutional neural network model is chosen. The convolutional kernels are trained to learn the interaction between each pore and its surrounding pores, and the shared parameter characteristic of the convolutional kernels is used to express the similarity of the physical mechanisms of the interaction between each pore and its surrounding pores, ensuring its physical meaning. Through the organic combination of physics and neural network algorithms, the computational accuracy of this invention is significantly improved. Attached Figure Description

[0014] Figure 1This is a schematic diagram of the structure of a double-walled blade flow prediction system based on a hole parameter convolutional neural network according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the convolutional neural network structure and core calculation process for hole parameters, taking a single row of holes as an example. Detailed Implementation

[0015] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings and preferred embodiments.

[0016] In one embodiment, such as Figure 1 As shown, this embodiment provides a flow prediction system for double-walled blades based on a hole parameter convolutional neural network. The system includes an input module, a parameter preprocessing module, a flow prediction module, and an output module.

[0017] Specifically, the input module is mainly used to receive the parameters of the double-walled blade to be calculated, which are input by the user through the UI interface. These parameters include two main categories: structural parameters and pressure parameters. The structural parameters include N parameters such as the impact hole diameter, film pore diameter, hole angle, hole spacing, and row spacing of each double-walled blade, as well as the number of holes in a single row n1 and the number of hole rows n. r The pressure parameters include the inlet pressure of each impact hole on the inner side of the double-walled blade and the outlet pressure of each film gas hole on the outer surface. The input module transmits the received structural and pressure parameters to the parameter preprocessing module.

[0018] The parameter preprocessing module is mainly used to process the user-input structural and pressure parameters, normalize them, and store them in tensor form required by the subsequent flow prediction module, obtaining the corresponding structural parameter tensor and pressure difference tensor, respectively. The structural parameter tensor has a size of N·n1·n. r As shown in equation (1), each element is the i-th parameter among the N structural parameters of the hole at that position.

[0019]

[0020] For the pressure difference tensor, the parameter preprocessing module needs to first calculate the inlet pressure of the inner impact hole at each location of the double-walled blade. and the outlet pressure of the outer surface air film holes The pressure difference is shown in equation (2), and then a pressure difference tensor is formed with a scale of 1·n1·n r As shown in equation (3), each element represents the internal and external pressure difference at that location.

[0021]

[0022]

[0023] The parameter preprocessing module processes the structural parameter tensor and pressure difference tensor according to equations (1) to (3), and then stores and transmits them to the flow prediction module.

[0024] The flow prediction module is primarily used to predict the cooling air flow rates of each impactor and film gas vent based on the pre-processed structural parameter tensors and pressure difference tensors of the double-walled blades. The flow prediction module receives the structural parameter tensors and pressure difference tensors from the parameter preprocessing module, using these as input parameters to the vent parameter convolutional neural network. It then directly calls the pre-trained vent parameter convolutional neural network to predict the flow rates of each of the following sizes: 1·n. r The flow rate tensors of the film cooling orifice and the impingement orifice are calculated, and the two flow rate tensors are transmitted to the output module.

[0025] The core model for flow prediction in the flow prediction module is a pre-trained aperture parameter convolutional neural network. This aperture parameter convolutional neural network consists of an input layer, an intermediate layer, and a flow output layer. The input layer takes the structural parameter tensor and pressure difference tensor of the double-walled blade as inputs. The intermediate layer consists of two convolutional layers. The output layer outputs two cooling air flow tensors: the film cooling orifice and the impingement orifice. and The sizes are both 1·n1·n r As shown in equations (4) and (5), each element represents the cold air flow rate of the hole at that location.

[0026]

[0027]

[0028] This invention, within the convolutional operation and training framework of convolutional neural networks, designs a pore parameter convolutional neural network structure and core computational process that embodies physical meaning, mainly comprising three features (features 1 to 3), such as... Figure 2 As shown, the following is a detailed explanation using a single row of holes as an example. Feature 1 is based on the structural characteristics of the impact holes and film vents in the double-walled structure, which are located on the inner and outer plates respectively. A dual-path prediction method is adopted, that is, the intermediate layer and output layer of the neural network are divided into two identical parts. One intermediate layer and output layer are used to predict the cold air flow rate of the film vents, and the other intermediate layer and output layer are used to predict the cold air flow rate of the impact holes. The two paths are interconnected and trained together through a loss function. Feature 2 is to add k convolutional layers after the input layers of structural parameter tensors and pressure difference tensors. The flow interaction between holes is learned through convolution calculation. The calculation process of the convolutional layers is shown in equations (6) and (7).

[0029] g (l) =h (l) *f (l) (6)

[0030]

[0031] Among them, f (l) and f (l+1) Let f represent the input and output of the l-th convolutional layer, respectively, 1 ≤ l ≤ k, and f (1) =p. h (l) b represents the convolution kernel of the l-th convolutional layer (l) It is the bias added after the convolution calculation. It is an activation function to increase the nonlinear expressive power of the layer. Feature 3 is to follow the convolutional layer with a dot product layer. This layer first classifies all elements of the pressure difference tensor of the initial input layer into 0 and 1 according to whether there is a pressure difference, as shown in Equation (8). Then, it performs a dot product between this zero tensor and the output tensor of the convolutional layer to obtain the final flow tensors of the impact orifice and the film gas orifice, as shown in Equation (9). The role of this dot product layer is to restrict the flow predicted by the model to be multiplied by 0 at orifices without pressure difference, so it must be strictly zero, while multiplying by 1 at orifices with pressure difference, so it is unaffected, thus conforming to physical reality.

[0032] q=sgn(p) (8)

[0033]

[0034] Where p is the initial pressure difference input tensor, and f (k) This represents the output of k convolutional layers. The unified representation represents the tensor of the cold air flow rate at the impact orifice and the cold air flow rate at the film orifice in the final output layer.

[0035] The training dataset consists of simulation data of cold air flow rates through various orifices in different double-walled structures under varying internal and external pressures. The structural and pressure parameters in the dataset cover commonly used values ​​in engineering practice. Since the total flow rate through the film gas orifice and the impingement orifice should be conserved, the loss function of the orifice parameter neural network consists of two additive parts: one part is the orifice flow rate tensor prediction. and simulation values The mean square error between the two parts is the predicted total cooling air flow rate of all film vents. And the predicted total cold air flow rate of all impact holes The square of the difference is shown in equation (10).

[0036]

[0037] The output module is mainly used to receive the predicted cooling gas flow tensor of the film gas orifice and the cooling gas flow tensor of the impingement orifice, add explanatory text such as the orifice location and the corresponding cooling gas flow value, thereby converting the predicted cooling gas flow tensors of the impingement orifice and the film gas orifice into a general text and parameter form, and displaying the converted text and parameters to the user through the UI interface, thus completing the prediction of the flow of each orifice of the double-walled turbine blade to be calculated.

[0038] In terms of computational speed, this invention achieves extremely high speed because the training of the convolutional neural network is pre-completed. Therefore, when faced with new double-walled structures and internal / external pressure conditions, the prediction system can directly call the pre-trained convolutional neural network for pore parameter calculations, eliminating the need for iterations. Regarding computational accuracy, firstly, based on the understanding of physical flow, the pore flow rate is only related to the structural flow resistance and the internal / external pressure difference. Structural parameters and pressure difference parameters are selected as input parameters to ensure physical reliability. Secondly, considering that the multi-connectivity of the double-walled structure means that the flow rate of a single pore is affected by the flow of surrounding pores, and that the physical mechanisms of this effect are similar, a convolutional neural network model is chosen. The convolutional kernels are trained to learn the interaction between each pore and its surrounding pores, and the shared parameter characteristic of the convolutional kernels is used to express the similarity of the physical mechanisms of the interaction between each pore and its surrounding pores, ensuring its physical meaning. Through the organic combination of physics and neural network algorithms, the computational accuracy of this invention is significantly improved.

[0039] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0040] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A flow prediction system for double-walled blades based on a pore parameter convolutional neural network, characterized in that, include: The input module is used to receive the structural and pressure parameters of the double-walled blade to be calculated, which are input by the user through the UI interface. The structural parameters include N parameters for each impact hole diameter, film pore diameter, hole angle, hole spacing, and row spacing of the double-walled blade, as well as the number of holes in a single row n1 and the number of hole rows n. r The pressure parameters include the inlet pressure of each impact hole and the outlet pressure of each film gas hole; The parameter preprocessing module is used to process the structural parameters and pressure parameters transmitted from the input module into tensor form, respectively obtaining tensors of scale. The structural parameter tensor and its size are The pressure difference tensor is obtained, and the structural parameter tensor and the pressure difference tensor are stored and transmitted to the flow prediction module respectively; The flow prediction module is used to call a pre-trained orifice parameter convolutional neural network, and to perform calculations using the structural parameter tensor and the pressure difference tensor as input parameters to the orifice parameter convolutional neural network, predicting flow rates of all sizes. The system calculates the cold air flow rate tensors for the film orifice and the impingement orifice, and transmits these two flow rate tensors to the output module. The aperture parameter convolutional neural network includes an input layer, an intermediate layer, and a flow rate output layer. Both the intermediate layer and the flow rate output layer are divided into two identical parts. One intermediate layer and output layer are used to predict the cold air flow rate of the film orifice, and the other intermediate layer and output layer are used to predict the cold air flow rate of the impingement orifice. The two paths are interconnected and trained together through a loss function. Each intermediate layer includes k convolutional layers following the input layer and a dot product layer following the convolutional layers. The loss function of the aperture parameter convolutional neural network consists of two parts: one part is the mean square error between the predicted cold air flow rate tensors of the film orifice and the impingement orifice and the corresponding simulated values; the other part is the square of the difference between the predicted total cold air flow rate of all film orifices and the predicted total cold air flow rate of all impingement orifices. The output module is used to convert the predicted cold air flow tensor of the film gas orifice and the cold air flow tensor of the impact orifice into a general text and parameter form, and to display the text and parameters through a UI interface.

2. The double-walled blade flow prediction system based on a hole parameter convolutional neural network according to claim 1, characterized in that, The structural parameter tensors are as follows: （1） Each element represents the i-th parameter among the N structural parameters of the hole at the corresponding position.

3. The double-walled blade flow prediction system based on a pore parameter convolutional neural network according to claim 1, characterized in that, When processing the pressure parameters into tensor form, the parameter preprocessing module first calculates the inlet pressure of the inner impact hole at each position of the double-walled blade. and the outlet pressure of the outer surface air film holes pressure difference Then, the pressure difference tensor is constructed based on the calculated pressure difference, and the pressure difference tensor is as follows: （3） Each element represents the internal and external pressure difference at the corresponding location.

Images