In-service damage monitoring method for aviation turbine blades based on digital twin

Abstract

The present invention relates to a method for monitoring damage in service of aviation turbine blades based on digital twins, which belongs to the field of aviation engine technology and solves the problems of high computational complexity, low efficiency and accuracy of existing detection algorithms. The physical entity of the turbine blade to be monitored is determined, and environmental information of the turbine blade is collected; a digital twin is constructed based on the physical entity, wherein a turbine blade three-dimensional model construction module constructs a turbine blade three-dimensional model according to the physical entity, a turbine blade operating environment simulation module performs environmental simulation on the model according to the real-time load, rotation speed and ambient temperature data of the turbine blade, and a turbine blade defect evolution prediction module uses an intelligent defect detection algorithm to identify the surface defects of the blades at each moment of the model, obtains the category, location information and size information of the defects, and also performs evolution simulation and risk factor analysis. The present invention realizes real-time online, high-precision and high-efficiency detection of surface defects of aviation turbine blades, thereby improving the service life of turbine blades.

Description

Technical Field

[0001] The present invention relates to the technical field of aero-engines, and in particular to a method for monitoring in-service damage of aero-turbine blades based on digital twins. Background Art

[0002] As a key part of aircraft engines, turbine blades are directly exposed to high temperature, high mechanical load and complex stress state due to their direct use of high temperature and high speed combustion gas to do work. These factors make turbine blades prone to failure, especially under high speed rotation, blades are susceptible to centrifugal inertia, aerodynamic force and foreign body impact, resulting in cracks, breakage, burns, deformation and other damage. At the same time, design and processing loopholes in the production process will also accelerate the formation of the above defects, making it a key component with a high failure rate in the engine. Under the action of composite material fatigue load, any tiny defect will cause the failure of turbine blades. Microcracks will cause cracking and even fracture of turbine blades under the action of high temperature creep, fatigue and other factors, causing great hidden dangers to the safe service of blades, affecting engine performance and even threatening aircraft safety. Therefore, the real-time monitoring of turbine blade defects is of great significance to ensure aviation safety, improve engine reliability and extend service life. However, due to the harsh environment in which aviation turbine blades are working, it is difficult for traditional defect detection methods to achieve real-time monitoring of aviation turbine blade entities during operation.

[0003] With the continuous development of informatization and intelligence, the concept of digital twins has gradually become popular. Since the working environment of aviation turbine blades is relatively harsh, it is difficult to realize real-time monitoring of the entity during the working process. Therefore, it is necessary to build a three-dimensional model of the turbine blade based on the idea of ​​digital twins and simulate the real-time environment. The real-time monitoring of the turbine blade entity can be realized by real-time defect detection of the virtual twin of the turbine blade. In order to overcome the limitations of traditional defect detection methods and realize real-time monitoring of defects, deep learning target detection methods based on convolutional neural networks (CNN) have gradually emerged. Through autonomous learning, image deep features are obtained to identify multiple and multi-scale defects. The detection range is large, the results are intuitive, and it is not affected by the external environment. Turbine blade defects have the characteristics of multiple types, multiple scales, and irregular shapes. Therefore, designing a high-precision and high-speed detection algorithm suitable for this type of defect detection is of great significance for the real-time monitoring of turbine blade defects.

[0004] However, the existing two-stage visual inspection algorithm has high computational complexity and slow calculation speed; the single-stage inspection algorithm has low accuracy in detecting multi-scale, multi-type and irregular-shaped defects of turbine blades, and is very likely to miss tiny defects and easily confuse the edges of defects of various shapes, making it difficult to accurately classify defects. Summary of the invention

[0005] In order to solve the problems of high computational complexity, low efficiency and accuracy in existing aircraft turbine blade defect detection algorithms, the present invention provides an aircraft turbine blade service damage monitoring method based on digital twins. The method realizes real-time monitoring of defects generated during the operation of aircraft turbine blades by constructing a digital twin of the aircraft turbine blades, and has higher detection accuracy and efficiency.

[0006] To solve the above problems, the present invention adopts the following technical solutions:

[0007] An aviation turbine blade service damage monitoring method based on digital twins includes the following steps:

[0008] Step 1: Determine the physical entity of the turbine blade to be monitored, use a force sensor, a speed sensor, and a temperature sensor to collect environmental information of the turbine blade and transmit the environmental information to the software port;

[0009] Step 2: Based on the physical entity of the turbine blade, a digital twin of the turbine blade is constructed, wherein the digital twin includes a turbine blade three-dimensional model construction module, a turbine blade operating environment simulation module and a turbine blade defect evolution prediction module. The turbine blade three-dimensional model construction module is used to construct a corresponding turbine blade three-dimensional model according to the physical entity of the turbine blade. The turbine blade operating environment simulation module is used to perform environmental simulation on the turbine blade three-dimensional model through Comsol simulation software based on the real-time load data, rotation speed data and temperature data of the environment of the turbine blade obtained through the software port. The turbine blade defect evolution prediction module is used to identify the blade surface defects at each moment in the turbine blade three-dimensional model using an intelligent defect detection algorithm, obtain the category, location information and size information of the blade surface defects, and perform evolution simulation on the blade surface defects on a time scale to obtain an evolution curve of defect size and shape characteristics, and perform a risk factor analysis on the evolution trend of the blade surface defects;

[0010] The intelligent defect detection algorithm includes a backbone network, multiple adaptive multi-scale receptive field attention modules, a neck network and a head network. Each adaptive multi-scale receptive field attention module includes an adaptive spatial feature extraction module and a multi-level cross-channel feature extraction module. The feature map obtained after the original blade surface image is processed by the convolution layer and the pooling layer is input into the backbone network. Each feature map output by the backbone network after the feature extraction operation is processed by the corresponding adaptive multi-scale receptive field attention module to obtain the final output feature map, and the output feature map is input into the neck network. The neck network adopts a feature pyramid network. The feature pyramid network performs multi-scale feature extraction and outputs a fused feature map to the head network. The head network adopts a T-head structure to realize the output of the prediction result, wherein the prediction result includes the category and position of the blade surface defect.

[0011] In each adaptive multi-scale receptive field attention module, the multi-level cross-channel feature extraction module first compresses the input feature map X in the spatial dimension through a global average pooling operation, retaining only the channel features, and compresses the original dimension C×H×W to C×1×1. The compressed feature map is then input into the first fully connected layer for dimensionality reduction feature extraction to obtain the predicted value of the importance of each channel. The second fully connected layer performs a dimensionality increase operation to update the importance value of each channel. Finally, after normalization and nonlinear processing, the channel attention weight is obtained. ;

[0012] The adaptive spatial feature extraction module first selects a convolution with a convolution kernel of corresponding size according to the size of the feature map output by the backbone network to reduce the dimension of the input feature map X, and performs preliminary feature extraction on the feature map, extracts global features through hole convolution, and only increases the expansion rate while expanding the receptive field of the convolution kernel; finally, a 1×1 convolution is used to map the feature map to a feature map of dimension 1×H×W to obtain the spatial attention weight ;

[0013] Channel attention weights are broadcasted and spatial attention weights Expand to the same dimension and add them to get the fused attention weights; assign corresponding weights to the elements in the input feature map X according to the fused attention weights to achieve the output of the fused attention feature map; finally, fuse the fused attention feature map with the input feature map X through the residual structure to get the output feature map X'.

[0014] Beneficial effects of the present invention: The digital twin-based aviation turbine blade service damage monitoring method proposed in the present invention constructs a digital twin of the aviation turbine blade, realizes real-time online detection of surface defects during the operation of the aviation turbine blade, improves the service life of the turbine blade, ensures aviation safety, and designs an intelligent defect detection algorithm for multi-scale, multi-type, and irregularly shaped blade surface defects. According to the characteristics of aviation turbine blade surface defects, an adaptive multi-scale receptive field self-attention module is introduced into the network structure of the intelligent defect detection algorithm, which can realize high-precision and high-efficiency detection of turbine blade surface defects, laying a foundation for the formulation of aviation turbine blade maintenance strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a schematic diagram of the in-service damage monitoring method for aviation turbine blades based on digital twins according to an embodiment of the present invention;

[0016] Figure 2 It is a schematic diagram of the network structure of the intelligent defect detection algorithm;

[0017] Figure 3 Schematic diagram of the network structure of the adaptive multi-scale receptive field attention module;

[0018] Figure 4 Schematic diagram of the principle of adaptively selecting the convolution kernel size. DETAILED DESCRIPTION

[0019] In order to realize real-time monitoring of aviation turbine blades during operation without affecting the actual working state of aviation turbine blades, the present invention provides an intelligent solution based on digital twins, that is, a service damage monitoring method for aviation turbine blades based on digital twins. The method utilizes digital twin technology to construct a 1:1 three-dimensional entity in a virtual space, and to construct the same environment through simulation software. Meanwhile, the environmental changes are updated synchronously, the working state of the target entity is visualized, more operating information is obtained, and the subsequent conditions that may occur in the turbine blades are predicted, so that relevant personnel can take appropriate maintenance measures in time, thereby increasing the service life of the turbine blades and ensuring aviation safety.

[0020] The present invention proposes a method for monitoring damage of aviation turbine blades in service based on digital twins, which mainly includes the following steps:

[0021] Step 1: Determine the physical entity of the turbine blade to be monitored, use a force sensor, a speed sensor, and a temperature sensor to collect environmental information of the turbine blade and transmit the environmental information to the software port;

[0022] Step 2: Based on the physical entity of the turbine blade, a digital twin of the turbine blade is constructed, and the digital twin is used to monitor the surface defects of the turbine blade in real time.

[0023] Specifically, Figure 1 As shown, the aviation turbine blade service damage monitoring method based on digital twin proposed in the present invention mainly includes three parts: physical entity, software control management and virtual entity (i.e. digital twin). The method is based on physical entity, and uses physical entity to guide the construction of virtual entity. The environmental information of turbine blade is collected through multiple sensors and transmitted to the software port (the control end for controlling virtual entity). Professionals review and correct the data and then update the parameters of the turbine blade operation environment simulation module in the virtual environment. After the turbine blade defect evolution prediction module obtains the evolution trend and risk factor analysis of the defect, the relevant information is transmitted to the software port. Professionals build a preventive maintenance plan based on the information and carry out preventive maintenance on the turbine blade based on the plan. Among them, the virtual entity includes three modules: turbine blade three-dimensional model construction module, turbine blade operation environment simulation module and turbine blade defect evolution prediction module. The turbine blade three-dimensional model construction module is used to build a 1:1 turbine blade three-dimensional model of the entity according to the physical entity of the turbine blade, and the geometric features and physical features of the entity are integrated. When building the turbine blade three-dimensional model, Catia software or solidworks software can be used. The turbine blade operation environment simulation module is used to simulate the environment of the three-dimensional virtual entity through the Comsol simulation software based on the real-time load data, rotation speed data and temperature data of the turbine blade obtained by the force sensor, speed sensor and temperature sensor, and simulate the actual working state of the turbine blade during operation. The main functions of the turbine blade defect evolution prediction module are: first, the intelligent defect detection algorithm is used to identify the blade surface defects at each moment in the turbine blade three-dimensional model, and the category information and size information of the defects are read. The size information of the defects mainly refers to large, medium and small in a broad sense. It is obtained based on the image input into the algorithm and the definition of defect scale by the international organization SPIE. When the ratio of the pixel blocks contained in the defect to the pixel blocks of the entire image is less than 0.12%, it is a small-scale target, and the rest are medium or large defects; secondly, the evolution of the blade surface defects on the time scale is simulated to obtain the evolution curve of the defect size and shape characteristics. Finally, the evolution trend of the blade surface defects is analyzed according to the evolution curve, so as to predict the overall danger of the turbine blade parts. The turbine blade defect evolution prediction module also transmits the results of the risk factor analysis to the software port, and professionals can obtain a maintenance plan for this type of blade surface defect after analysis and discussion. According to the above process, the digital twin method can be used to avoid sudden failure of turbine blades in subsequent work.

[0024] In order to realize online nondestructive detection of turbine blade surface defects, the present invention adopts a high-precision intelligent defect detection algorithm that can detect multi-scale, multi-type and irregular-shaped defects of turbine blades. The overall architecture of the algorithm is as follows: Figure 2 As shown in the figure, it is divided into three parts: feature extraction (backbone network), feature fusion (neck network) and defect classification and positioning (head network). Among them, the backbone network uses the classic feature extraction network ResNet-50 neural network. ResNet-50 neural network is a classic feature extraction model. By adding BatchNorm layer and residual structure, it can ensure that the network is deep enough while maintaining high detection accuracy. A standard residual block in ResNet-50 neural network consists of two convolutional layers, and RELU is the activation function. In the feature extraction part, the original leaf surface image is processed by a convolution layer and a pooling layer to obtain a feature map with a size of 64 channels, a height of h, and a width of w. The feature map of size (64, h, w) is input into the backbone network for feature extraction. During the ResNet50 processing, it goes through four stages: stage1, stage2, stage3, and stage4. Each stage outputs feature maps of sizes: (256, h, w), (512, h / 2, w / 2), (1024, h / 4, w / 4), and (2048, h / 16, w / 16), which are denoted as C1, C2, C3, and C4 respectively.

[0025] In order to realize the detection of multi-scale defects in turbine blades, the neck network of the model uses Feature Pyramid Network (FPN) for multi-scale feature extraction. Feature detection methods are mainly divided into three categories: single feature layer detection, multi-scale input detection and multi-scale feature map detection. In the convolutional network, the information contained in feature maps at different levels is different. The deep feature map contains rich semantic information, that is, category information, but due to the low pixel count, it contains less spatial information. The shallow feature map contains rich spatial information, but less semantic information. The backbone network can obtain feature maps of different scales output in four stages. From stage1 to stage4, the semantic information of the output feature map is gradually enriched and the spatial information is gradually reduced. In order to better utilize the information features at different scales, the feature maps of different scales obtained by feature extraction are combined, and the high-level feature map is upsampled and fused with the feature map of the previous layer, so that the feature map of this layer contains both high-level semantic information and low-level spatial information, so as to realize efficient and accurate prediction of feature maps of different scales and complete the detection of minor defects. In the FPN network, feature map C4 is upsampled 2 times and added to feature map C3, and then the obtained fused feature map is upsampled 2 times and added to feature map C2, and so on. At the same time, in order to meet the head network's requirements for the scale characteristics of the input feature map, two downsamplings are required to obtain feature maps with spatial sizes of (256, h / 16, w / 16) and (256, h / 32, w / 32) respectively. The feature maps obtained by each fusion constitute the output feature map set (or fused feature map) of the neck network, denoted as The head network will complete the prediction of defect type and defect location based on the feature maps in the set.

[0026] For the head network of the model, the T-head structure is used to achieve the output of prediction results, where the prediction results include the category and location of the blade surface defects. The T-head structure is different from the traditional single-stage detection output head. This structure realizes the interaction between the classification task and the position positioning task through the alignment task module to reduce the conflict between the results of the two types of tasks and achieve more accurate classification prediction and position positioning. The T-head structure consists of two parts: the feature conversion structure and the task alignment predictor (TAP). First, task interaction feature extraction is performed. In order to enhance the interaction between the two tasks of classification and positioning, the feature maps of different scales output by the FPN network are processed by using n continuous conversion layers with activation functions to obtain the task interaction feature map , in order to promote the interaction between the two tasks and provide them with multi-level features, the calculation formula is as follows:

[0027] (1)

[0028] in, and Indicates convolutional layers and RELU functions, Indicates Task interaction feature map, represents the fused feature map output by FPN. Then the calculated task interaction feature map Input to the TAP module for classification task and location task alignment operation. The model can eliminate the inconsistency of the two tasks to a certain extent by performing classification and location tasks based on task interaction features. However, due to the different types of features that the two tasks focus on, there is a conflict between the two task features in the task interaction features. To resolve this conflict, a layer attention mechanism is introduced to perform feature decomposition specific to the two types of tasks on the task interaction features in each layer. The calculation method is as follows:

[0029] (2)

[0030] in, Represents the task characteristics of each classification or location positioning task, The attention coefficient for each task is calculated by fusion features of cross-level tasks to capture the dependencies between layers. The calculation method is as follows:

[0031] (3)

[0032] in, and are two fully connected layers, Represents the sigmoid function, and then from Results of prediction classification and location positioning tasks :

[0033] (4)

[0034] in, for After completing the preliminary prediction of the task, the two tasks, namely the defect classification task and the location task, are further aligned by adjusting the spatial distribution of the classification task (P) and the spatial distribution of the pre-selected box of the location task (B). For the classification task, the spatial probability map is used. Adjustment:

[0035] (5)

[0036] Among them, M is calculated through interactive features to learn the consistency between two tasks.

[0037] Furthermore, for the alignment of the position positioning task, the offset is learned from the interaction task Adjust the position of each position prediction box:

[0038] (6)

[0039] in, Represents the position of the defect location prediction box, i and j represent the coordinates of the location point in space, and c represents a certain channel. The above formula can be implemented based on bilinear interpolation. M and O are calculated by learning the task interaction features in the feature stack:

[0040] (7)

[0041] (8)

[0042] in, and are two 1×1 convolutions for dimensionality reduction.

[0043] Compared with other single-stage object detection models, the learnable task alignment head structure can unify the results of classification tasks and location positioning tasks, and complete the object detection task with high efficiency and high accuracy.

[0044] In order to adapt to the characteristics of turbine blades with multiple types, multi-scales, and irregular shapes of defects, and to ensure the unity of classification tasks and location positioning tasks while realizing efficient comprehensive detection of complex defects in turbine blades, multiple adaptive multi-scale receptive field (AMRF) attention modules are added between the backbone network and the neck network. The network structures of each AMRF attention module are the same, and the difference lies in the size of the convolution kernel.

[0045] like Figure 3 As shown in , each AMRF attention module contains two modules: an adaptive spatial feature extraction module and a multi-level cross-channel feature extraction module. In the adaptive spatial feature extraction module, the convolution kernel size decreases with the increase of the layer depth, and the number of channels increases with the increase of the layer depth, as shown in Figure 4 As shown in the figure, in the multi-level cross-channel feature extraction module, the feature map is reduced in dimension through the fully connected layer to extract the important features of the feature map, and the tiny features of the feature map are further extracted through the dimension increase operation.

[0046] The multi-level cross-channel feature extraction module first compresses the input feature map X (i.e., the feature map output by the backbone network) in the spatial dimension through a global average pooling operation, retaining only the channel features, and compressing the original dimension C×H×W to C×1×1. The compression process is shown in formula (9). The compressed feature map is further input into the first fully connected layer for dimensionality reduction feature extraction to obtain the predicted value of the importance of each channel. This process extracts the defect features with obvious channel dimensions. In order to further accurately extract the small features that are difficult to detect, the second fully connected layer is used to perform dimensionality increase operations to update the importance values ​​of each channel. Finally, after normalization and nonlinear processing, the channel attention weight is obtained. .

[0047] (9)

[0048] in, Represents the compressed feature map, H and W represent the height and width of the input feature map X respectively. The coordinates in the representative feature map are The pixel value in the pixel grid.

[0049] The adaptive spatial feature extraction module first selects a convolution with a convolution kernel of corresponding size according to the size of the feature map output by the backbone network to reduce the dimension of the input feature map X. In order to reduce the number of parameters required for calculation, the convolution layer Conv2 is used to perform preliminary feature extraction on the feature map. The global features are further extracted through a 3×3 hole convolution, which only increases the expansion rate while expanding the receptive field of the convolution kernel. Finally, a 1×1 convolution is used to map the feature map to a feature map of dimension 1×H×W to obtain the spatial attention weight. .

[0050] Getting the channel attention weight and spatial attention weights After that, the broadcast mechanism will and Expand to the same dimension and add them to get the fused attention weights. Then, according to the fused attention weights, assign corresponding weights to the elements in the input feature map X to achieve the output of the fused attention feature map. Finally, the fused attention feature map is fused with the input feature map X through the residual structure to get the final output feature map X'.

[0051] The digital twin-based aviation turbine blade service damage monitoring method proposed in the present invention constructs a digital twin of the aviation turbine blade to realize real-time online detection of surface defects during the operation of the aviation turbine blade, thereby increasing the service life of the turbine blade and ensuring aviation safety. An intelligent defect detection algorithm is designed for multi-scale, multi-type, and irregularly shaped blade surface defects. According to the characteristics of aviation turbine blade surface defects, an adaptive multi-scale receptive field self-attention module is introduced into the network structure of the intelligent defect detection algorithm, which can realize high-precision and high-efficiency detection of turbine blade surface defects, laying the foundation for the formulation of aviation turbine blade maintenance strategies.

[0052] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described 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.

[0053] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A method for monitoring damage in service of aviation turbine blades based on digital twins, characterized in that: The following steps are involved: Step 1: Determine the physical entity of the turbine blade to be monitored, use a force sensor, a speed sensor, and a temperature sensor to collect environmental information of the turbine blade and transmit the environmental information to the software port; Step 2: Based on the physical entity of the turbine blade, a digital twin of the turbine blade is constructed, wherein the digital twin includes a turbine blade three-dimensional model construction module, a turbine blade operating environment simulation module and a turbine blade defect evolution prediction module. The turbine blade three-dimensional model construction module is used to construct a corresponding turbine blade three-dimensional model according to the physical entity of the turbine blade. The turbine blade operating environment simulation module is used to perform environmental simulation on the turbine blade three-dimensional model through Comsol simulation software based on the real-time load data, rotation speed data and temperature data of the environment of the turbine blade obtained through the software port. The turbine blade defect evolution prediction module is used to identify the blade surface defects at each moment in the turbine blade three-dimensional model using an intelligent defect detection algorithm, obtain the category, location information and size information of the blade surface defects, and perform evolution simulation on the blade surface defects on a time scale to obtain an evolution curve of defect size and shape characteristics, and perform a risk factor analysis on the evolution trend of the blade surface defects; The intelligent defect detection algorithm includes a backbone network, multiple adaptive multi-scale receptive field attention modules, a neck network and a head network. Each adaptive multi-scale receptive field attention module includes an adaptive spatial feature extraction module and a multi-level cross-channel feature extraction module. The feature map obtained after the original blade surface image is processed by the convolution layer and the pooling layer is input into the backbone network. Each feature map output by the backbone network after the feature extraction operation is processed by the corresponding adaptive multi-scale receptive field attention module to obtain the final output feature map, and the output feature map is input into the neck network. The neck network adopts a feature pyramid network. The feature pyramid network performs multi-scale feature extraction and outputs a fused feature map to the head network. The head network adopts a T-head structure to realize the output of the prediction result, wherein the prediction result includes the category and position of the blade surface defect. In each adaptive multi-scale receptive field attention module, the multi-level cross-channel feature extraction module first compresses the input feature map X in the spatial dimension through a global average pooling operation, retaining only the channel features, and compresses the original dimension C×H×W to C×1×1. The compressed feature map is then input into the first fully connected layer for dimensionality reduction feature extraction to obtain the predicted value of the importance of each channel. The second fully connected layer performs a dimensionality increase operation to update the importance value of each channel. Finally, after normalization and nonlinear processing, the channel attention weight is obtained. ; The adaptive spatial feature extraction module first selects a convolution with a convolution kernel of corresponding size according to the size of the feature map output by the backbone network to reduce the dimension of the input feature map X, and performs preliminary feature extraction on the feature map, extracts global features through hole convolution, and only increases the expansion rate while expanding the receptive field of the convolution kernel; finally, a 1×1 convolution is used to map the feature map to a feature map of dimension 1×H×W to obtain the spatial attention weight ; Channel attention weights are broadcasted and spatial attention weights Expand to the same dimension and add them to get the fused attention weights; assign corresponding weights to the elements in the input feature map X according to the fused attention weights to achieve the output of the fused attention feature map; finally, fuse the fused attention feature map with the input feature map X through the residual structure to get the output feature map X'.

2. The method for monitoring damage of aviation turbine blades in service based on digital twin according to claim 1, characterized in that: The backbone network uses the ResNet-50 neural network. The original leaf surface image is processed by the convolution layer and the pooling layer to obtain a feature map of size (64, h, w). The feature map of size (64, h, w) is input into the ResNet-50 neural network for feature extraction. After processing by the adaptive multi-scale receptive field attention module, four final output feature maps of sizes (256, h, w), (512, h / 2, w / 2), (1024, h / 4, w / 4), and (2048, h / 16, w / 16) are obtained.

3. The method for monitoring damage of aviation turbine blades in service based on digital twins according to claim 1 or 2, characterized in that: The process of compressing the input feature map X in the spatial dimension by the multi-level cross-channel feature extraction module is expressed as follows: （9） in, Represents the compressed feature map, H and W represent the height and width of the input feature map X respectively. The coordinates in the representative feature map are The pixel value in the pixel grid.

4. The method for monitoring damage of aviation turbine blades in service based on digital twins according to claim 1 or 2, characterized in that: The T-head structure first uses n continuous conversion layers with activation functions to process the feature maps of different scales output by the feature pyramid network to obtain the task interaction feature map , and then calculate the task interaction feature map Input into the task alignment predictor to perform classification task and location positioning task alignment operations.

5. The method for monitoring damage of aviation turbine blades in service based on digital twin according to claim 4, characterized in that: Task interaction feature map The calculation formula is as follows: （1） in, and Indicates convolutional layers and RELU functions, Indicates Task interaction feature map, Represents the fused feature map output by the feature pyramid network.

6. The method for monitoring damage of aviation turbine blades in service based on digital twin according to claim 5, characterized in that: A layer attention mechanism is introduced into the T-head structure. In each layer, the task interaction features are decomposed into features specific to the two types of tasks. The calculation method is as follows: （2） in, Represents the task characteristics of each classification or location positioning task, The attention coefficient for each task is calculated by fusion features of cross-level tasks to capture the dependencies between layers. The calculation method is as follows: （3） in, and are two fully connected layers, Represents the sigmoid function, and then from Results of prediction classification and location positioning tasks : （4） in, for Obtained by splicing.

7. The method for monitoring damage of aviation turbine blades in service based on digital twin according to claim 6, characterized in that: The two tasks are aligned by adjusting the spatial distribution of the classification task and the spatial distribution of the pre-selected box of the location positioning task; For classification tasks, use spatial probability maps Adjustment: （5） Among them, P is the spatial distribution function of the classification task, and M is calculated through the interaction features to learn the consistency between the two tasks; For the alignment of the position localization task, the offset is learned from the interaction task Adjust the position of each position prediction box: （6） Among them, B is the spatial distribution function of the pre-selected box for the location positioning task, Represents the position of the defect prediction box, i and j represent the coordinates of the position point in space, and c represents a certain channel.

8. The method for monitoring damage of aviation turbine blades in service based on digital twin according to claim 7, characterized in that: M and O are calculated by learning the task interaction features in the feature stack: （7） （8） in, and are two 1×1 convolutions for dimensionality reduction.

9. The method for monitoring damage of aviation turbine blades in service based on digital twins according to claim 1 or 2, characterized in that: The turbine blade defect evolution prediction module is also used to transmit the results of the risk factor analysis to the software port to obtain a maintenance plan for the blade surface defect.

10. The method for monitoring damage of aviation turbine blades in service based on digital twins according to claim 1 or 2, characterized in that: Use Catia software or solidworks software to build a three-dimensional model of the turbine blade.

Images