CN115014789A

CN115014789A - CNN-GCN-based dual-sensor aeroengine case fault source acoustic emission positioning method

Info

Publication number: CN115014789A
Application number: CN202210608930.XA
Authority: CN
Inventors: 杨国安; 刘曈; 王硕
Original assignee: Beijing University of Chemical Technology
Current assignee: Beijing University of Chemical Technology
Priority date: 2022-05-31
Filing date: 2022-05-31
Publication date: 2022-09-06
Anticipated expiration: 2042-05-31
Also published as: CN115014789B

Abstract

A CNN-GCN-based dual-sensor aeroengine case fault source acoustic emission positioning method belongs to the field of aeroengine case fault positioning, and relates to an aeroengine case fault acoustic emission source positioning method only using dual sensors, in particular to an aeroengine case fault source acoustic emission positioning method with multi-part coupling characteristics under the condition of dual sensors based on the combination of a convolutional neural network and a graph convolutional neural network. The method only uses two acoustic emission sensors to acquire signals and is applied to a casing structure with a coupling interface. The method combines a plurality of neural network models, and improves the positioning precision. The fault source region positioning of the aircraft engine case simulation test bed can be realized.

Description

CNN-GCN-based dual-sensor aeroengine case fault source acoustic emission positioning method

Technical Field

The invention belongs to the field of aeroengine case fault location, and relates to an aeroengine case fault acoustic emission source location method only using double sensors, in particular to an aeroengine case fault acoustic emission location method with multi-part coupling characteristics under the condition of double sensors based on the combination of a convolutional neural network and a graph convolutional neural network.

Background

The aircraft engine provides all the power required by the aircraft in the flight process, is called the heart of the aircraft, is composed of a plurality of precise parts with high complexity, and is inevitably easy to have various faults when working in a severe environment with high temperature, high speed, strong vibration and large stress. Engine failures have a considerable proportion of the failures in flight and often result in catastrophic accidents in flight because of engine failures. Therefore, in order to operate the engine safely and efficiently and save maintenance costs, the condition monitoring and fault diagnosis capability of the aircraft engine must be improved.

The traditional online monitoring technology mainly comprises a vibration monitoring technology, a gas path analysis monitoring technology, a lubricating oil monitoring technology, an electrostatic monitoring technology and the like, but at the initial stage of engine failure (such as cracks, impact, friction, structural deformation and the like), the weak characteristics of early damage cannot cause the engine to have an obvious abnormal state, and at the moment, the technology cannot monitor the early damage in real time. However, engine structural health monitoring methods based on acoustic emission technology have significant advantages in addressing such issues. The acoustic emission refers to a physical phenomenon that when a material or a structure is subjected to external force or internal force, when the stress exceeds the yield strength, internal stress is suddenly released to generate transient elastic waves, and the transient elastic waves carry a large amount of information of the structure or the defect of the material, and the signals are detected and analyzed to determine information related to the damage of the material or the structure, including damage properties, damage positions, damage degrees, residual life and the like, which is called as an acoustic emission detection technology. In recent years, the depth and the breadth of sound emission nondestructive testing technology research are greatly developed and applied to a plurality of fields.

However, many existing structural health monitoring studies based on acoustic emission technology focus primarily on a single structure with little application on components that include coupling interfaces. Most studies were only experimentally validated in some plates with a pore structure. But do not explore the case of multi-part coupling structures, such as those involving complex coupling interfaces, varied dimensions and various bolt fastenings, and even structures involving different materials. Tubular structures containing coupling interfaces are common in the aerospace industry, such as low pressure compressor casing structures containing casing mounting edges. The complex interface coupling relationships make theoretical wave path analysis difficult and time consuming, resulting in very difficult localization of the acoustic emission source on these components. And because the engine has a complex structure, a plurality of components and works in a high-temperature, high-pressure and high-rotating-speed environment, the fault diagnosis of the engine can only arrange sensors at limited positions to provide monitoring capability, so that fewer sensors can be used than the traditional positioning method. However, the existing documents relating to the location of few sensor fault sources are mostly limited to simple structures, and the development is only fresh in parts containing complex coupling interfaces. In addition, the single-sensor acoustic emission monitoring technology can only realize simple identification of different types of fault sources and positioning of asymmetric structures. And effective positioning cannot be achieved for objects having a symmetrical structure, such as a fully symmetrical thin-walled cylindrical structure. Therefore, the invention adopts the double-sensor acoustic emission technology to realize the fault source positioning.

In recent years, the rapid development of deep learning techniques has accelerated the development of the field of fault diagnosis. Meanwhile, in deep learning, a Convolutional Neural Network (CNN) benefits from the hierarchical abstraction capability of deep learning, and has been widely applied to the fields of image classification, object detection, semantic segmentation, and the like. Graph convolution neural networks (GCNs) act as a kind of feed-forward neural network that uses graph convolution to process graph structure data. At present, GCN shows excellent performance in the fields of drug synthesis, power load prediction, link learning and the like. The GCN not only can effectively excavate a complex nonlinear relation between a fault source position and an acoustic emission signal by using a graph convolution layer with strong learning capacity, but also can represent similarity measurement between an unknown sample and a marked sample by using an adjacency matrix, so that the accuracy of fault source positioning is improved. Therefore, it is necessary to establish a model of the CNN in combination with the GCN to achieve acoustic emission based localization of aircraft engine case fault sources including coupling interfaces using only dual sensors.

Disclosure of Invention

The method aims to solve the problem of fault acoustic emission positioning of the aircraft engine case comprising the coupling interface under the condition of double sensors by utilizing the powerful feature extraction capability and the pattern recognition capability of the convolutional neural network and the graph convolutional neural network.

The invention provides a CNN-GCN-based aeroengine case fault acoustic emission positioning method, which only uses two acoustic emission sensors to acquire signals and is applied to a case structure with a coupling interface. The method combines a plurality of neural network models, and improves the positioning precision. The fault source region positioning of the aircraft engine case simulation test bed can be realized.

The method is characterized in that: the specific operation steps are as follows:

the method comprises the following steps: n positioning areas are divided on an aeroengine casing to be monitored, an H-N source is applied to the random position of each area, and two acoustic emission sensors are used for data acquisition. To avoid the chance of diagnostic results, the application was repeated 100 times for each case to form a data set. And an effective data enhancement scheme is adopted: taking a single sensor as an example, the entire data set is shuffled to use a random order of images during training. In addition, the data set is enhanced to 500 by flipping, mirroring and rotating, resulting in a data set with sufficient images. In each case, 300 images were used for training, 100 for validation, and the remaining 100 for testing. The present invention uses two sensors, using a total of 2 x 500 x N images.

Step two: and converting the acquired signals into time-frequency graphs by using continuous wavelet transform, and combining the time-frequency graphs of the fault signals acquired by the two sensors in the same region. Where the wavelet function is selected as a Morlet wavelet, the continuous wavelet transform of the time domain signal s (t) of scale a and time shift b for a given mother wavelet ψ can be expressed as:

wherein is a complex conjugate, W _s (a, b) is the wavelet coefficient of s (t), R represents the whole real number set, s (t) represents the time domain signal to be converted, ψ is the wavelet mother function, a is the scale factor of the wavelet function, the scaling of the wavelet mother function is controlled, b is the time shift factor, the shift position on the time axis is controlled when the scale is fixed. The continuous wavelet transform in the formula canTo be viewed as acoustic emission signal s (t) and a set of wavelet mother functions ψ _a,b (t) inner product.

Step three: and building a CNN-GCN network model based on a Tensorflow deep learning framework. For the CNN part, the modified VGG16 is used to extract features and add a global average pooling layer (GAP). Specifically, the CNN section deletes VGG16 last rolling block5 and subsequent fully connected layers, using block4 the feature that the maximum pooled layer output is 14 × 14 × 512 in size. The size of the feature map obtained in this section is 1/16 of the input image, effectively preserving the detailed information in the time-frequency map. The last full link layer is deleted and GAP is added. GAP reduces the space dimension by calculating the average value of the height and the width of the feature mapping, and avoids overfitting caused by excessive parameters of the full connection layer. The equation for GAP is as follows:

where x is the input feature map and w, h and c are the width, height and number of channels, respectively. Herein, the GAP converts a 14 × 14 × 512 three-dimensional tensor into a one-dimensional eigenvector of size 512-dim.

Step four: and (3) constructing a graph structure, namely constructing a time-frequency graph sample set of the fault acoustic emission signals obtained by the CWT calculation into a graph structure consisting of nodes and edges. Wherein, the nodes represent time-frequency graphs at different positions, and the edges represent the connection among the time-frequency graphs. Then, the dependency relationship between the nodes is captured through the graph convolution operation. Finally, a position feature based on the time-frequency graph is generated by the pooling layer.

Step five: a feature fusion block is designed. The features extracted from the CNN branch and the GCN branch are integrated and transmitted to the final classifier, and the operation of feature fusion is as follows:

wherein, [, ]]The operation is a cascade of elements and,

and

representing the layer i features extracted from CNN and GCN, respectively.

The fused features are sent into a softmax layer, and the parameter values in the model are optimized by using a cross entropy loss function in a training phase, namely:

wherein, y _i Indicates a class i label, 1 indicates an image belonging to the corresponding class, whereas 0 indicates no. M represents the number of time-frequency graph classes. z is a radical of _i Representing the output probability of the model. By combining CNN and GCN, the method effectively improves the discrimination capability of the characteristics.

Step six: and training a CNN-GCN model and realizing positioning of a fault acoustic emission source of the aircraft engine case. Wherein, the CNN-GCN model uses Adam to optimize the network, the learning rate can be multiplied by (1- (iter/maxIter)) ^0.5 And performing dynamic updating, wherein iter is iteration time, and maxIter is maximum iteration time. dropout layers are set in the middle of each layer, and the ReLU activation function and batch processing normalization are used for all layers except the last layer, wherein the batch normalization momentum (momentum for weight optimization) is set to 0.9. During the network training process, the maximum epoch number is set to 150. Further, the regular term coefficient is l of 0.001 ₂ Norm regularization is applied to the weight decay to stabilize network training and reduce overfitting. During training, the input image is adjusted to 224 × 224, randomly flipped, mirrored, and rotated for data enhancement.

Drawings

FIG. 1 is a flow chart of a CNN-GCN-based double-sensor aeroengine case fault acoustic emission positioning method

FIG. 2 is a CNN-GCN network model overall framework

FIG. 3 is a partial framework of model CNN

FIG. 4 is a model GCN partial framework

FIG. 5 is a CNN-GCN model learning curve

Detailed Description

In the following description, the present invention will be described in detail according to exemplary embodiments.

The invention is verified on an aeroengine casing simulation test bed. The test piece is formed by combining two simulation casings, wherein the simulation casings have the outer diameter of 1000mm, the wall thickness of 3mm and the height of 300 mm. The two simulation casings are connected and fixed by a mounting edge structure with the height of 10 mm. The simulation bench was divided into 192 regions by a 100 x 100mm grid, so a total of 192 experimental scenarios were used to train, test and validate the proposed CNN-GNN model. The failure source was applied at random locations in each area, and in order to avoid the chance of diagnostic results, the application was repeated 100 times for each case to form a data set. Some effective enhancement schemes are also taken here, such as shuffling the entire data set to use a random order of images during training. In addition, the data set is enhanced to 500 by flipping, mirroring and rotating, resulting in a data set with sufficient images. In each case, 300 images were used for training, 100 for validation, and the remaining 100 for testing. A total of 96000 images were used in this experiment.

The one-time training verification result of the CNN-GCN model provided by the invention is shown in FIG. 5, which shows the verification accuracy and loss along with the change of iteration times. It can be seen that the model trains 150 epochs, and in the early phase, the loss function of the training set decreases rapidly as the number of iterations increases. After 107 epochs, the verification precision of the model reaches 99.89%, and the model is not increased in the next 43 periods and reaches the highest level, at the moment, the training accuracy reaches 99.85%, and the testing accuracy is 99.67%. The result shows that the method can accurately identify the position of most fault sources.

Claims

1. A CNN-GCN-based dual-sensor aeroengine case fault source acoustic emission positioning method is characterized by comprising the following steps:

the method comprises the following steps: dividing N positioning areas on an aeroengine casing to be monitored, applying an H-N source at a random position of each area, and performing data acquisition by using two acoustic emission sensors; and data enhancement is adopted:

step two: converting the acquired signals into time-frequency graphs by using continuous wavelet transform, and combining the time-frequency graphs of the fault signals acquired by the two sensors in the same region; where the wavelet function is selected as a Morlet wavelet, the continuous wavelet transform of the time domain signal s (t) of scale a and time shift b for a given mother wavelet ψ can be expressed as:

wherein is a complex conjugate, W _s (a, b) is wavelet coefficient of s (t), R represents whole real number set, s (t) represents time domain signal to be converted, psi is wavelet mother function, a is scale factor of wavelet function to control scaling of wavelet mother function, b is time shift factor to control shift position on time axis when scale is fixed; the continuous wavelet transform in the formula can be regarded as an acoustic emission signal s (t) and a group of wavelet mother functions psi _a,b (t) inner product;

step three: building a CNN-GCN network model based on a Tensorflow deep learning framework; for the CNN part, the modified VGG16 is used to extract features and add a global average pooling layer GAP; specifically, the CNN partially deletes the last rolling block5 of VGG16 and the subsequent fully connected layers, using the feature that block4 has the maximum pooled layer output size of 14 × 14 × 512; the size of the feature map obtained by this section is 1/16 of the input image; deleting the last full link layer and adding GAP;

the equation for GAP is as follows:

wherein x is the input feature map, w, h and c are width, height and channel number, respectively; GAP converts a 14 × 14 × 512 three-dimensional tensor into a one-dimensional eigenvector with the size of 512-dim;

step four: constructing a graph structure, namely constructing a time-frequency graph sample set of fault acoustic emission signals obtained by CWT calculation into a graph structure consisting of nodes and edges; wherein, the nodes represent time-frequency graphs at different positions, and the edges represent the connection among the time-frequency graphs; capturing the dependency relationship between the nodes through graph convolution operation; generating a position feature by a pooling layer based on the time-frequency graph;

step five: designing a characteristic fusion block; the features extracted from the CNN branch and the GCN branch are integrated and transmitted to the final classifier, and the operation of feature fusion is as follows:

wherein, [, ]]The operation is a cascade of elements and,

and

respectively representing the layer i features extracted from the CNN and GCN;

wherein, y _i Indicating a label of the ith class, 1 indicating an image belonging to the corresponding class, and conversely, 0 indicating no; m represents the number of time-frequency graph categories; z is a radical of _i Representing the output probability of the model;

step six: training a CNN-GCN model,

CNN-GCN modelAdam is used to optimize the network, the learning rate is multiplied by the base learning rate by (1- (iter/maxIter)) ^0.5 Performing dynamic updating, wherein iter is iteration time, and maxIter is maximum iteration time; the dropout layer is arranged in the middle of each layer, and the ReLU activation function and batch processing normalization are used for all layers except the last layer, wherein the batch normalization momentum, namely momentum, is set to be 0.9; in the network training process, the maximum epoch number is set to 150; further, the regular term coefficient is l of 0.001 ₂ Norm regularization is applied to the weight attenuation; during training, the input image is adjusted to 224 × 224, randomly flipped, mirrored, and rotated for data enhancement.