A method for detecting key components of aircraft piping systems based on supervised attention mechanism

Through the detection method based on the supervised attention mechanism, the Hiformer network automatically recognizes key components of the aircraft pipeline system in complex environments, solving the problems of low accuracy and insufficient robustness of the traditional methods, and achieving efficient detection and classification , improves the safety and reliability of aircraft engines.

CN118196076BActive Publication Date: 2024-11-12NANJING UNIV OF AERONAUTICS & ASTRONAUTICS
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410510853.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-04-26
Publication Date
2024-11-12
Estimated Expiration
2044-04-26

AI Technical Summary

Technical Problem

Traditional methods have low accuracy and insufficient robustness in identifying key components of aircraft pipeline systems, and are unable to effectively deal with complex and changing aircraft engine environments.

Method used

Using a detection method based on a supervised attention mechanism, a feature map is obtained through the backbone network, combining the perception extraction module and feature extraction layer, an attention supervision matrix is ​​generated, and the Hiformer network is trained to obtain a defect detection model, so as to achieve a critical zero for the aircraft pipeline system. Automatic identification and classification of components.

Benefits of technology

The accuracy and efficiency of detection are significantly improved in multi-scale and complex environments, suitable for various industrial scenarios, improve the safety and reliability of aircraft engines, and provide strong technical support for maintenance and fault diagnosis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118196076B_ABST
    Figure CN118196076B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of mechanical processing and industrial visual inspection technology, and solves the technical problems of low accuracy and insufficient robustness of traditional methods in identifying key components of aircraft piping systems, and in particular, relates to a method for detecting key components of aircraft piping systems based on a supervised attention mechanism. The detection method proposed in the present invention achieves excellent performance in multi-scale and complex environments, performs feature extraction through an attention supervision matrix, and focuses on analyzing the characteristics of key components in aircraft piping systems. At the same time, by adopting a feature extraction layer, the method can simultaneously extract features of different scales, and is suitable for complex piping system target detection in various industrial scenarios.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical processing and industrial visual inspection, and in particular to a method for detecting key components of an aircraft piping system based on a supervised attention mechanism. Background Art

[0002] In current aviation engineering technology, identifying key components in aircraft piping systems has become an important topic. These key components play a vital role in the performance and safe operation of the entire engine. Traditional methods rely on the experience and visual judgment of professionals to identify these key components, but this method is not only time-consuming and labor-intensive, but also has limited accuracy and efficiency.

[0003] With the development of technology, especially the advancement of artificial intelligence technology, the use of convolutional neural networks can realize the detection and identification of key components. However, due to the complexity and variability of the aircraft engine environment, traditional methods still have low accuracy and insufficient robustness when identifying key components of aircraft piping systems. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides a method for detecting key components of an aircraft pipeline system based on a supervised attention mechanism, which solves the technical problems of low accuracy and insufficient robustness of traditional methods in identifying key components of an aircraft pipeline system.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions: a method for detecting key components of an aircraft pipeline system based on a supervised attention mechanism, the method comprising the following steps:

[0006] S1. Obtain the key component data of the aircraft piping system and obtain the feature map containing the key component information through the backbone network Backbone. ;

[0007] S2, feature map In the input perception extraction module PEL, the feature map Extract features and perform perception analysis modeling to finally obtain the attention supervision matrix ;

[0008] S3. Obtain feature maps containing shallow features in key component data through the feature extraction layer FEL ;

[0009] S4. Feature map and feature map After pixel-level superposition, the output is sent to the attention supervision module SA. Obtain coded information under supervision ;

[0010] S5, the coded information Input to the decoding layer DL to obtain the escape information, and minimize the loss function based on the escape information to train the Hiformer network to obtain the final defect detection model;

[0011] S6. Complete the inspection of key components in the aircraft piping system based on the defect detection model.

[0012] Furthermore, in step S2, the specific process includes the following steps:

[0013] S21, feature map Input into the perception extraction module PEL for perception, and obtain the perception ratio output by the perception extraction module PEL ;

[0014] S22. Feature Information The input is sent to the perception thrust layer MLP for perception reasoning, and after activation, the output is the basic attention matrix for subsequent connection with the perception information. ;

[0015] S23. The basic attention matrix and perceived ratio Splicing is performed to obtain the attention supervision matrix for detecting key components in the aircraft piping system .

[0016] Further, in step S21, the perception extraction module PEL includes a method for capturing feature information of key components in a complex piping system. Local guidance module LGM; and for analyzing and capturing feature information And get the perception vector Perception Guidance Layer PGL.

[0017] Further, in step S23, the attention supervision matrix The expression is:

[0018] ;

[0019] In the above formula, Represents the convolution calculation of the perception extraction module; Indicates that the convolution kernel size is Convolution; represents bias; HL represents the number of hidden layers; Adam represents the adaptive learning rate; is the perception vector; It is the output of the LGM module.

[0020] Furthermore, in step S3, the specific process includes the following steps:

[0021] S31. Key component data is input into the feature extraction layer FEL to extract multi-scale feature maps As the input of the attention supervision mechanism Hiformer, it first passes through a 3×3 convolution module, and then uses the feature pyramid FPN to fuse multi-scale features;

[0022] S32. Multi-scale feature map Perform feature extraction and batch normalization, and finally unify the multi-scale feature map With feature map Dimension.

[0023] Further, in step S4, the feature map and feature map Perform pixel-level superposition to obtain the input of the attention supervision module SA ,Right now:

[0024] ;

[0025] Among them, re-shape means matrix size adjustment; Indicates the corresponding addition between matrix elements; Indicates that the convolution kernel size is Convolution; Indicates bias; It is the convolution calculation; Represents an element of a feature graph containing key component information obtained through the backbone network.

[0026] Further, in step S4, in the attention supervision matrix Obtain coded information under supervision , where the encoded information The expression is:

[0027] ;

[0028] In the above formula, It is the input of the attention supervision module SA; To guide matrix operations under the attention supervision mechanism Hiformer; Indicates that the convolution kernel size is Convolution; represents bias; HL represents the number of hidden layers; Adam represents the adaptive learning rate.

[0029] Furthermore, in step S5, the specific process includes the following steps:

[0030] S51, the coded information After being input into the decoding layer DL, it is decoded by the hierarchical attention structure decoding module (Dehiformer) to obtain the decoding information D1;

[0031] S52, input the decoded information D1 into the traditional convolutional layer to extract information to obtain the escaped information D2;

[0032] S53, extract the escape information D2 and input it into the convolution layer with a convolution kernel of 1×1, and then output it to obtain a defect detection model.

[0033] Through the above technical solution, the present invention provides a method for detecting key components of an aircraft pipeline system based on a supervised attention mechanism, which has at least the following beneficial effects:

[0034] 1. The detection method proposed in this paper achieves excellent performance in multi-scale and complex environments. Compared with other segmentation frameworks, the attention supervision matrix Feature extraction is performed to focus on analyzing the features of key components in the aircraft piping system. At the same time, by adopting the feature extraction layer, this method can simultaneously extract features of different scales, which is suitable for complex piping system target detection in various industrial scenarios.

[0035] 2. The present invention uses machine vision technology to automatically identify and classify various key components in the aviation engine piping system. Through training with a large amount of data, the data-driven method can accurately identify different types of components, even those in complex environments or different states. Compared with traditional methods, this deep learning method significantly improves the accuracy and efficiency of segmentation.

[0036] 3. The present invention not only provides an efficient technical means for segmenting key components in aircraft piping systems, but also provides strong technical support for aircraft engine maintenance and fault diagnosis, further improving the safety and reliability of aircraft engines. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0038] Figure 1 It is a framework diagram of the Hiformer network model of the present invention;

[0039] Figure 2 This is a rendering of the key component identification of the aircraft piping system of the present invention. DETAILED DESCRIPTION

[0040] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific implementation methods, so that the implementation process of how the present application uses technical means to solve technical problems and achieve technical effects can be fully understood and implemented accordingly.

[0041] Please refer to Figure 1-Figure 2 , shows a specific implementation of this embodiment. This embodiment achieves excellent performance in multi-scale and complex environments by adopting the attention supervision mechanism Hiformer with a hierarchical attention structure. Compared with other segmentation frameworks, this embodiment adopts a carefully designed attention supervision matrix Feature extraction is performed to focus on analyzing the features of key components in the aircraft piping system. At the same time, by adopting the feature extraction layer, the method can simultaneously extract features of different scales, which is suitable for complex piping system target detection in various industrial scenarios. Based on the above technical objectives, this embodiment proposes a method for detecting key components of an aircraft piping system based on a supervised attention mechanism, which includes the following steps:

[0042] S1. Obtain the key component data of the aircraft piping system and obtain the feature map containing the key component information through the backbone network Backbone. In this embodiment, a two-dimensional industrial camera can be used to capture images of key components of the aircraft piping system, and then image preprocessing is performed on the key components in the captured images to filter out noise in the images.

[0043] S2, feature map In the input perception extraction module PEL, the feature map Extract features and perform perception analysis modeling to finally obtain the attention supervision matrix ; In step S2, the specific process includes the following steps:

[0044] S21, feature map Input into the perception extraction module PEL for perception, and obtain the perception ratio output by the perception extraction module PEL In this embodiment, the perception and extraction module PEL is used to capture the characteristic information of key components in a complex piping system. Local guidance module LGM; and for analyzing and capturing feature information And get the perception vector The perception guidance layer PGL is used to capture the characteristic information of key components in complex pipeline systems through the local guidance module (LGM). , for feature information The Perception Guidance Layer (PGL) is used to analyze the captured information to obtain the perception vector , and then output the perception ratio , used for subsequent evaluation and guidance of Hiformer’s token interactions.

[0045] S22. Feature Information The input is sent to the perception thrust layer MLP for perception reasoning, and after activation, the output is the basic attention matrix for subsequent connection with the perception information. ;

[0046] S23. The basic attention matrix and perceived ratio Splicing is performed to obtain the attention supervision matrix for detecting key components in the aircraft piping system .

[0047] Please refer to Figure 1 , in order to give the attention supervision matrix Better supervision and ensure that the attention supervision mechanism Hiformer can better perform self-attention calculations and add attention supervision matrix Introducing the Perception Ratio , through the perception ratio To better determine and ensure that the attention supervision module SA outputs more effective information and reduces the amount of calculation.

[0048] Among them, the attention supervision matrix The expression is:

[0049] ;

[0050] In the above formula, Represents the convolution calculation of the perception extraction module; Indicates that the convolution kernel size is Convolution; represents bias; HL represents the number of hidden layers; Adam represents the adaptive learning rate; is the perception vector; is the output of the LGM module. By designing the number of HL hidden layers and the Adam adaptive learning rate in the perception reasoning layer MLP, the perception structure is adjusted to obtain a more effective attention supervision matrix. .

[0051] For the perception extraction module PEL, it passes through a local guidance module LGM and is input into the perception reasoning layer MLP and the perception guidance layer PGL respectively. The local guidance module LGM and the perception guidance layer PGL are both two multi-layer convolution modules. For the local guidance module LGM, its purpose is to extract The information is used by the perception guidance layer and the perception layer to extract the perception information. The perception guidance layer PGL is also a convolution module, and its purpose is to generate the perception vector in order to extract the effectiveness of the perception information. , further providing guidance information for the attention supervision mechanism Hiformer.

[0052] for (i.e., attention supervision matrix The elements 1 to n of the i-th column in the attention supervision matrix For the input of the attention supervision module SA The i-th token number ;but, In the attention supervision matrix Under the supervision of Respectively with input Token serial number To token number Perform self-attention calculation. is the perception ratio of the supervision matrix. By token number To token number composition.

[0053] S3. Obtain feature maps containing shallow features in key component data through the feature extraction layer FEL ; The specific process includes the following steps:

[0054] S31. Key component data is input into the feature extraction layer FEL to extract multi-scale feature maps As the input of the attention supervision mechanism Hiformer, it first passes through a 3×3 convolution module, and then uses the feature pyramid FPN to fuse multi-scale features;

[0055] S32. Multi-scale feature map Perform feature extraction and batch normalization, and finally unify the multi-scale feature map With feature map The dimensions of the feature map are as follows: Pixel-level superposition is performed to obtain the input of the attention supervision module SA.

[0056] In this embodiment, the present invention uses the FPN feature pyramid mechanism in the feature extraction layer FEL to fuse various spatial feature information in the channel dimension, fuses multi-scale features at the channel level, and enhances the expression capability of feature fusion and spatial information capture.

[0057] S4. Feature map and feature map After pixel-level superposition, the output is sent to the attention supervision module SA. Obtain coded information under supervision , the input matrix for subsequent guidance operations; in this step, the feature map and feature map Perform pixel-level superposition and add attention supervision matrix The following guidance matrix operation forms a hierarchical attention structure to output the encoded information More specifically, refer to Figure 1 The output feature map is resized and superimposed with the feature information map pixel level to obtain the input:

[0058] ;

[0059] Among them, re-shape means matrix size adjustment; Indicates the corresponding addition between matrix elements; Indicates that the convolution kernel size is Convolution; Indicates bias; It is the convolution calculation; Represents the elements of the feature map containing key component information obtained through the backbone network Backbone. The matrix operation is guided by the attention supervision mechanism Hiformer as follows: , the output is expressed as:

[0060] ;

[0061] Adopting the attention supervision mechanism Hiformer, for In the attention supervision matrix In exist For the i-th token, the calculation method of G (Guidance), K (Key), and V (Value) is as follows:

[0062] ;

[0063] ;

[0064] ;

[0065] ;

[0066] in, , , These are weights to be calculated by gradient descent; is the perceived ratio; through the constant Scaling is performed to avoid dot products being too large.

[0067] S5, the coded information The information is input into the decoding layer DL to obtain the escape information, and the loss function is minimized according to the escape information to train the Hiformer network to obtain the final defect detection model; the specific process includes the following steps:

[0068] S51, the coded information After being input into the decoding layer DL, it is decoded by the hierarchical attention structure decoding module (Dehiformer) to obtain the decoding information D1;

[0069] S52, input the decoded information D1 into the traditional convolutional layer to extract information to obtain the escaped information D2;

[0070] S53, extract the escape information D2 and input it into the convolution layer with a convolution kernel of 1×1, and then output it to obtain a defect detection model.

[0071] In this embodiment, more specifically, for the Deformer decoder, the input of the decoder is first converted into an embedding matrix, and then the position code is added to it, and it is sent as input to the bottom decoder. The decoder receives the input and sends it to the masked multi-head attention layer to generate an attention matrix. The attention matrix and the eigenvalues ​​output by the encoder are then used as the input of the multi-head attention layer, and the new attention matrix is ​​output again. The attention matrix obtained from the attention layer is used as input and sent to the feedforward network layer. The feedforward network layer uses the attention matrix as input and the decoded features as output. Finally, the output obtained from decoder 1 is used as input and sent to decoder 2. Decoder 2 performs the same processing and outputs the decoded information D1. D1 is then input into the traditional convolutional layer for information extraction to obtain the escape information D2. Finally, the Hiformer network is trained to obtain the final defect detection model by minimizing the loss function based on the escape information D2.

[0072] S6. Complete the inspection of key components in the aircraft piping system based on the defect detection model. After minimizing the loss function through the escape information D2, the Hiformer network is trained to finally obtain the defect detection model. When it is necessary to inspect the key components in the aircraft piping system, it is only necessary to use a two-dimensional industrial camera to collect images of the key components of the aircraft piping system, and then input them into the defect detection model to output the corresponding inspection results, such as Figure 2 As shown, it is a diagram showing the effect of applying the detection method proposed in this embodiment to the aircraft piping system for identifying key components.

[0073] Compared with the traditional detection method of key components of aircraft piping system, the method proposed in this invention shows significant superiority. The traditional method relies on the experience and visual judgment of professionals, but this method is not only time-consuming and labor-intensive, but also has limited accuracy and efficiency. In this context, the present invention introduces a new method based on deep learning technology, especially combining the feature extraction ability and learning efficiency of convolutional neural network and Hiformer.

[0074] The present invention uses machine vision technology to automatically identify and classify various key components in the pipeline system of an aircraft engine. Through training with a large amount of data, the data-driven method can accurately distinguish different types of components, even those in complex environments or under different conditions. Compared with traditional methods, this deep learning method significantly improves the accuracy and efficiency of segmentation.

[0075] In addition, considering the complexity and variability of the aircraft engine environment, significant progress has been made in optimizing the network structure and algorithm. This innovative method not only provides an efficient technical means for the segmentation of key components in the aircraft piping system, but also provides strong technical support for the maintenance and fault diagnosis of aircraft engines, further improving the safety and reliability of aircraft engines.

[0076] Those skilled in the art can understand that all or part of the steps in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a program, so the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0077] The above implementation methods have been described in detail. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A method for detecting key components of an aircraft piping system based on a supervised attention mechanism, characterized in that: The method comprises the following steps: S1. Obtain the key component data of the aircraft piping system and obtain the feature map containing the key component information through the backbone network Backbone. S2, feature map In the input perception extraction module PEL, the feature map Extract features and perform perception analysis modeling to finally obtain the attention supervision matrix M p The specific process includes the following steps: S21, feature map The input is sensed in the perception extraction module PEL, and the perception ratio P output by the perception extraction module PEL is obtained. i ; S22. Feature Information The input is sent to the perception reasoning layer MLP for perception reasoning, and after activation, the basic attention matrix M is output for subsequent connection with the perception information. p '; S23, the basic attention matrix M p ' and the perceived ratio P i The attention supervision matrix M is obtained for detecting key components in the aircraft pipeline system. p ; S3, obtain the feature map I' containing shallow features in the key component data through the feature extraction layer FEL p ; S4, feature map I' p and feature map After pixel-level superposition, the output is sent to the attention supervision module SA. p Obtaining coded information under supervision S5, the coded information Input to the decoding layer DL to obtain the escape information, and minimize the loss function based on the escape information to train the Hiformer network to obtain the final defect detection model; S6. Complete the inspection of key components in the aircraft piping system based on the defect detection model.

2. The method for detecting key components of an aircraft piping system according to claim 1, characterized in that: In step S21, the perception extraction module PEL includes a module for capturing the characteristic information of key components in a complex piping system. The local guidance module LGM; And used to analyze and capture feature information And get the perception vector L M Perception Guidance Layer PGL.

3. The method for detecting key components of an aircraft piping system according to claim 1, characterized in that: In step S23, the attention supervision matrix M p The expression is: In the above formula, F PEL Represents the convolution calculation of the perception extraction module; Indicates that the convolution kernel size is N km convolution; b represents bias; HL represents the number of hidden layers; Adam represents the adaptive learning rate; L M is the perception vector; It is the output of the LGM module.

4. The method for detecting key components of an aircraft piping system according to claim 1, characterized in that: In step S3, the specific process includes the following steps: S31, key component data is input into the feature extraction layer FEL to extract multi-scale feature map I' p As the input of the attention supervision module SA, it first passes through a 3×3 convolution module, and then uses the feature pyramid FPN to fuse multi-scale features; S32, multi-scale feature map I' p Perform feature extraction and batch normalization, and finally unify the multi-scale feature map I' p With feature map Dimension.

5. The method for detecting key components of an aircraft piping system according to claim 1, characterized in that: In step S4, the feature map I' p and feature map Perform pixel-level superposition to obtain the input of the attention supervision module SA Right now: Among them, re-shape means matrix size adjustment; Indicates the corresponding addition between matrix elements; Indicates that the convolution kernel size is N km convolution; b represents bias; F cov is the convolution calculation; I e Represents an element of a feature graph containing key component information obtained through the backbone network.

6. The method for detecting key components of an aircraft piping system according to claim 1, characterized in that: In step S4, in the attention supervision matrix M p Obtaining coded information under supervision Among them, the coded information The expression is: In the above formula, It is the input of the attention supervision module SA; To guide matrix operations under attention supervision; Indicates that the convolution kernel size is N km convolution; b represents bias; HL represents the number of hidden layers; Adam represents the adaptive learning rate.

7. The method for detecting key components of an aircraft piping system according to claim 1, characterized in that: In step S5, the specific process includes the following steps: S51, the coded information After being input into the decoding layer DL, it is decoded by the hierarchical attention structure decoding module Dehiformer to obtain the decoding information D1; S52, input the decoded information D1 into the traditional convolutional layer to extract information to obtain the escaped information D2; S53, extract the escape information D2 and input it into the convolution layer with a convolution kernel of 1×1, and then output it to obtain a defect detection model.