A double-wall structure feature extraction method based on an attention mechanism

Abstract

The application discloses a double-wall structure feature extraction method based on an attention mechanism, relates to the technical field of aero-engine cooling structures, and solves the limitation of a double-wall structure heat exchange characteristic / strength performance rapid evaluation model based on a deep learning method in feature extraction, which leads to problems such as inaccurate prediction of the model, and the like. The method cuts a double-wall structure into a plurality of sub-regions through a sub-region cutting module, utilizes a self-attention mechanism in a deep learning method, performs self-attention calculation operation on all the sub-regions through a self-attention feature extraction module, and finally obtains a full-field information feature tensor. The method can effectively extract features from full-field information of a double-wall structure unit, and the features are used in downstream tasks, such as heat exchange characteristic prediction and strength performance prediction, so that the prediction accuracy and generalization of a prediction model can be effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of aero-engine cooling structure, and particularly relates to a double-wall structure feature extraction method based on an attention mechanism. BACKGROUND

[0002] An aero-engine is an important weapon equipment determining the performance of an airplane. At present, the aero-engine is developing towards a high thrust-to-weight ratio. The turbine inlet temperature is an important index determining the thrust of the aero-engine. Generally speaking, increasing the turbine inlet temperature can effectively improve the performance of the engine, but also causes the hot end components to face a more severe working environment. The double-wall structure is a high-efficiency cooling structure, which can be used in hot end components such as the combustion chamber and the turbine blade, and can effectively improve the turbine inlet temperature and help to increase the thrust-to-weight ratio of the engine. The double-wall cooling structure guides the cold air to impact the outer wall and form an air film by densely opening holes in the inner and outer walls, and uses multiple cylinders to connect between the inner and outer walls to increase the heat exchange area. The double-wall cooling structure efficiently combines impact heat exchange, circular rib convection heat exchange and air film heat insulation, and therefore has strong heat exchange capacity. However, the complex structure of the double-wall cooling structure also causes the evaluation of its performance to consume more resources in the design process.

[0003] With the rapid development of information technology such as artificial intelligence, some deep learning method-based double-wall structure heat exchange characteristic / strong performance rapid evaluation models have been gradually proposed. For the model based on the deep learning method, feature extraction is an important step affecting the prediction performance of the model. The existing model mainly adopts the form of convolution for feature extraction: the convolution network is used to automatically extract features, the double-wall structure is abstracted as a three-dimensional tensor, and a two-dimensional convolution or a three-dimensional convolution is used to extract features with a pooling operation. The main problem of the current feature extraction method is that the receptive field of the convolution operation depends on the size of the convolution kernel, and a small convolution kernel of 3x3 or 5x5 is usually used. Therefore, the convolution-based feature extraction operation has locality and is difficult to extract features representing the full-field information. This leads to the fact that when the full-field information of the double-wall structure changes in actual application, the model cannot accurately predict. Although the convolution operation used in the existing technology has strong local feature extraction capability and can be applied to the field of computer vision with image data, for physical field information, the information is usually continuous, and the local information does not have the ability to represent the full field. Therefore, the features extracted by the convolution have weak robustness, and the deep learning model based on the features cannot effectively generalize.

[0004] Therefore, a feature extraction method capable of effectively extracting full-field features of the double-wall structure is urgently needed. SUMMARY

[0005] The application provides a double-wall structure feature extraction method based on an attention mechanism to solve the limitations of a double-wall structure heat exchange characteristic / strength performance rapid evaluation model based on a deep learning method in feature extraction, and the method can comprehensively extract features from full-field information, so that the features are more obvious and the prediction performance is improved.

[0006] A double-wall structure feature extraction method based on an attention mechanism, and the implementation process of the method is as follows:

[0007] Step 1: The double-wall structure unit is divided into sub-regions to obtain an input tensor.

[0008] Step 2: Full-field information feature extraction based on a self-attention mechanism.

[0009] Three fully connected neural networks are used to input the input tensor obtained in step 1 into the fully connected neural networks respectively to obtain three two-dimensional tensors, which are defined as Q matrix, K matrix and V matrix respectively.

[0010] Step 3: The Q matrix, K matrix and V matrix in step 2 are calculated to obtain an attention matrix, and after the attention matrix is normalized and nonlinearly mapped by a fully connected neural network, step 2 and step 3 are repeated, and finally a full-field information feature tensor is obtained.

[0011] Further, the method is realized through a sub-region division module and a self-attention feature extraction module, the sub-region division module is used to divide the double-wall structure into a plurality of sub-regions to obtain an input tensor.

[0012] The self-attention feature extraction module uses a fully connected neural network to extract Q matrix, K matrix and V matrix to obtain an attention matrix.

[0013] Further, in step 1, the solid domain of the double-wall structure is uniformly cut into P h ,P w ,P d segments in three dimensions of length, width and height to obtain P h ×P w ×P d sub-regional slices, and the total number of sub-regions is N; each sub-regional slice is uniformly sampled under a two-dimensional grid to obtain a three-dimensional vector with a shape of C, H, W, wherein H is a length dimension sampling frequency, W is a width dimension sampling frequency, and C is the total number of coordinate information and physical information stored by a node; the three-dimensional vector representing the sub-region is converted into a two-dimensional vector with a shape of C, H×W, all the sub-regions are represented as a three-dimensional tensor with a shape of N, C, H×W, and the three-dimensional tensor is used as an input tensor.

[0014] Further, in step two, three fully connected neural networks, mlp1, mlp2, mlp3, are defined, and the input tensor is transmitted into the fully connected neural network to obtain three two-dimensional tensors of N, HxW.

[0015] Further, in step three, the Q matrix, the K matrix and the V matrix are calculated according to the following formula to obtain an attention matrix.

[0016]

[0017] In the formula, d k is the size HxW of the third dimension of the input tensor, and softmax() is an activation function.

[0018] The attention matrix is N, HxW, and after normalization and fully connected neural network nonlinear mapping of the attention matrix, steps two and three are repeated to obtain a final full-field information feature tensor.

[0019] The beneficial effects of the present application are:

[0020] The double-wall structure unit feature extraction method based on the attention mechanism disclosed in the present application utilizes the self-attention mechanism in the deep learning method, can effectively extract features from the full-field information of the double-wall structure unit, and utilizes the features to input into downstream tasks such as heat exchange characteristic prediction and strength performance prediction, so as to effectively improve the prediction accuracy and generalization of the prediction model. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 FIG. 1 is a schematic diagram of a sub-region segmentation module in the double-wall structure feature extraction method based on the attention mechanism disclosed in the present application;

[0022] Figure 2 FIG. 2 is a schematic diagram of a self-attention feature extraction module. DETAILED DESCRIPTION

[0023] In combination Figure 1 and Figure 2 The present embodiment is a double-wall structure feature extraction method based on the attention mechanism, which cuts the double-wall structure into a plurality of sub-regions by a sub-region segmentation module, utilizes the self-attention mechanism in the deep learning method, performs self-attention calculation operation on all sub-regions through a self-attention feature extraction module, and finally obtains a full-field information feature tensor. The specific process is as follows:

[0024] I. Double-wall structure unit sub-region segmentation

[0025] The sub-region segmentation module uniformly cuts the solid domain of the double-wall structure in three dimensions of length, width and height, and obtains P hP w P d segment, obtaining P h ×P w ×P d sub-region slices, and the total number of the sub-regions is set as N. Each sub-region slice is uniformly sampled under a two-dimensional grid to obtain a three-dimensional vector in the shape of (C, H, W), where H is a length dimension sampling frequency, W is a width dimension sampling frequency, and C is a total number of coordinate information and physical information stored by a node. The three-dimensional vector representing the sub-region in the input tensor is flattened into a two-dimensional vector in the shape of (C, HxW), and therefore, all the sub-regions can be represented as a four-dimensional tensor in the shape of (N, C, H, W), which is named as an input tensor.

[0026] II. Full-field information feature extraction based on self-attention mechanism (feature extractor based on self-attention mechanism);

[0027] First, a full connection neural network is used to extract quary, key and value matrices. That is, three full connection neural networks mlp1, mlp2 and mlp3 are defined, the input tensor (N, C, HxW) flattened from the four-dimensional tensor is input into the corresponding full connection neural network, and three two-dimensional tensors in the shape of (N, HxW) are obtained, which are respectively named as Q (quary matrix), K (key matrix) and V (value matrix). Then, attention matrix calculation is performed; the quary, key and value matrices are calculated according to the following formula to obtain an attention matrix.

[0028]

[0029] wherein, d k is the size of the third dimension of the input tensor HxW, and softmax() is an activation function.

[0030] The shape of the attention matrix is (N, HxW). After normalization and full connection neural network nonlinear mapping of the attention matrix, the above process is repeated L times to obtain the final full-field feature information. The obtained feature is used as the input of a downstream task, such as heat exchange characteristic prediction, strength performance prediction, etc., which can effectively improve the prediction accuracy and generalization of the prediction model.

[0031] The technical features of the above-described embodiments can be combined in any manner. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described, however, as long as the combinations of the technical features do not exist contradictions, they should be considered as the scope of the present disclosure.

[0032] The above embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be noted that for ordinary skilled persons in the art, without departing from the concept of the present application, several modifications and improvements can be made, which are within the protection scope of the present application, and therefore, the protection scope of the present application patent should be subject to the appended claims.

Claims

1. A feature extraction method for double-walled structures based on an attention mechanism, characterized by: This method is implemented by the following steps: Step 1: Divide the double-walled structural unit into sub-regions to obtain the input tensor; specifically: The solid domain of the double-walled structure is uniformly cut into P along the length, width, and height dimensions. h ,P w ,P d Segment, obtain P h ×P w ×P d Each sub-region is sliced ​​into N slices. Each sub-region slice is sampled on a two-dimensional uniform grid to obtain a three-dimensional vector of shape C, H, W, where H is the sampling frequency of the length dimension, W is the sampling frequency of the width dimension, and C is the total number of coordinate and physical information stored in the node. The three-dimensional vector representing the sub-region is transformed into a two-dimensional vector of shape C, H, W. All sub-regions are represented as four-dimensional tensors of shape N, C, H, W, which are used as input tensors. Step 2: Full-field information feature extraction based on self-attention mechanism; Three fully connected neural networks are used. The input tensor obtained in step one is fed into the fully connected neural networks respectively to obtain three two-dimensional tensors, which are defined as Q matrix, K matrix and V matrix respectively. Step 3: Calculate the Q matrix, K matrix, and V matrix obtained in Step 2 to obtain the attention matrix. After normalizing the attention matrix and performing nonlinear mapping with a fully connected neural network, repeat Step 2 and Step 3 to obtain the final full-field information feature tensor.

2. The method for extracting features of a double-walled structure based on an attention mechanism according to claim 1, characterized in that: This method is implemented through a sub-region segmentation module and a self-attention feature extraction module. The sub-region segmentation module is used to segment the double-wall structure into several sub-regions to obtain the input tensor. The self-attention feature extraction module uses a fully connected neural network to extract the Q matrix, K matrix, and V matrix to obtain the attention matrix.

3. The method for extracting features of a double-walled structure based on an attention mechanism according to claim 1, characterized in that: In step two, three fully connected neural networks, mlp1, mlp2, and mlp3, are defined. The input tensor is flattened into N, C, H×W form and fed into the fully connected neural networks to obtain three N, H×W two-dimensional tensors.

4. The method for extracting features of a double-walled structure based on an attention mechanism according to claim 1, characterized in that: In step three, the Q matrix, K matrix, and V matrix are calculated using the following formula to obtain the attention matrix; In the formula, d k The third dimension of the input tensor is H×W, and softmax() is the activation function; The attention matrix N,H×W is obtained. After normalizing the attention matrix and performing a nonlinear mapping with a fully connected neural network, steps two and three are repeated to obtain the final full-field information feature tensor.

Images