Lobe mixer crack fault identification method based on multi-head mixed attention mechanism

Abstract

The invention discloses a lobe mixer crack fault identification method based on a multi-head mixed attention mechanism, and particularly relates to the field of aero-engines. Comprising the following steps: acquiring a fault signal in the operation process of a mixer, and performing short-time Fourier transform on the fault signal to obtain a time-frequency image; inputting the time-frequency image into a pre-trained crack fault diagnosis model to obtain a crack diagnosis result; wherein the pre-trained crack fault diagnosis model comprises a plurality of feature extraction modules which are connected in sequence, and the last feature extraction module is connected with a first convolution module and a classification module in sequence; and each feature extraction module comprises a second convolution module and a multi-head mixed attention feature extraction module which are connected in sequence. Based on the method, the accuracy of identifying the crack fault of the afterburner mixer of the aero-engine can be improved in a multi-working-condition environment.

Description

Technical Field

[0001] This application relates to the field of aero-engines, and in particular to a method for identifying crack faults in lobe mixers based on a multi-head hybrid attention mechanism. Background Technology

[0002] The afterburner is a key structure in aero-engines that provides "afterburning," delivering maximum thrust for short periods to meet mission profile requirements such as takeoff, climb, acceleration, and pursuit. The lobe mixer, a large, thin-walled, lobe-shaped component within the afterburner, directly affects the mixing of high- and low-temperature airflows in the inner and outer bypass ducts and fuel atomization. Under harsh operating environments such as high temperature, high-speed airflow, and heavy loads, the lobe mixer primarily bears thermal stress caused by the temperature difference between the inner and outer walls, pressure difference stress, aerodynamic stress, and vibration stress, making it a high-risk area for crack initiation. Once a crack forms in the mixer, it may propagate rapidly under loads such as high-speed airflow, leading to poorer mixing of the high-temperature combustion gas and the low-temperature airflow. This, in turn, reduces combustion efficiency, thrust, and increases fuel consumption in the afterburner. In extreme cases, it may even cause aero-engine failure, posing a serious threat to flight safety. Therefore, achieving crack fault diagnosis in the lobe mixer during operation is of great significance for ensuring the operational safety of aero-engines.

[0003] Unlike rotating components such as bearings and gears, which exhibit obvious periodic failure impact characteristics, crack faults in lobe mixers do not display clear, speed-related periodic characteristics. Their fault signals (such as strain and vibration signals) are typically very weak and submerged in the strong background noise of engine operation, making it difficult to effectively diagnose early crack states using traditional signal analysis methods (such as spectral analysis and envelope demodulation). In recent years, deep learning technology has been widely applied in the field of intelligent fault diagnosis for aero-engines, achieving significant results. However, deep learning-based methods still face significant challenges in the specific detection of crack faults under the operation of lobe mixers, mainly in two aspects: First, traditional deep learning models typically use the original time-domain monitoring signal as input, requiring the model to directly learn crack features from the original signal with extremely low signal-to-noise ratio, leading to difficulty in model convergence and poor generalization ability. Second, existing research, when introducing attention mechanisms to improve model performance, often allocates weights in a "global" or "average" manner, lacking the ability to finely enhance the fault-sensitive components in the signal, thus limiting its diagnostic accuracy for weak, non-periodic crack faults in a strong noise background. Summary of the Invention

[0004] The main objective of this application is to provide a method for identifying crack faults in a beam mixer based on a multi-head hybrid attention mechanism, aiming to solve the problem of low diagnostic accuracy in existing diagnostic methods.

[0005] To achieve the above objectives, this application provides a method for identifying crack faults in a lobe mixer based on a multi-head hybrid attention mechanism, comprising: acquiring fault signals during mixer operation; performing short-time Fourier transform on the fault signals to obtain a time-frequency image; inputting the time-frequency image into a pre-trained crack fault diagnosis model to obtain a crack diagnosis result; wherein, the pre-trained crack fault diagnosis model includes multiple sequentially connected feature extraction modules, the last feature extraction module being sequentially connected to a first convolution module and a classification module; each feature extraction module includes a sequentially connected second convolution module and a multi-head hybrid attention feature extraction module; the second convolution module is used to convolve the time-frequency image. The feature map is obtained. The multi-head hybrid attention feature extraction module is used to extract features from the feature map. The extraction process includes: uniformly dividing the channel dimension of the feature map to obtain features of multiple channel attention heads; determining the channel attention weight of each channel attention head, and using the channel attention weight to enhance the feature map to obtain a channel-enhanced feature map; flattening the channel-enhanced feature map to obtain a one-dimensional pixel sequence, uniformly dividing and reshaping the one-dimensional pixel sequence to obtain features of multiple spatial attention heads; determining the spatial attention weight of each spatial attention head, and using the spatial attention weight to enhance the features of the channel-enhanced feature map to obtain a spatially enhanced feature map.

[0006] Optionally, determining the channel attention weights for each channel attention head includes: performing global average pooling and global max pooling on the features of each channel attention head to obtain a first channel feature vector and a second channel feature vector; performing fully connected operations on the first channel feature vector and the second channel feature vector to obtain a first feature vector and a second feature vector; and adding the first feature vector and the second feature vector together and activating them to obtain the channel attention weights.

[0007] Optionally, feature enhancement is performed on the feature map using channel attention weights to obtain a channel-enhanced feature map, including: multiplying the channel attention weights with the feature map spatially to obtain the first enhanced feature corresponding to each channel attention head; and concatenating all the first enhanced features according to the channel dimension to obtain the channel-enhanced feature map.

[0008] Optionally, determining the spatial attention weights for each spatial attention head includes: performing channel-dimensional average pooling and channel-dimensional max pooling on the features of each spatial attention head to obtain a first spatial feature vector and a second spatial feature vector; concatenating the first spatial feature vector and the second spatial feature vector according to the channel dimension, and performing convolution and activation operations to obtain the spatial attention weights.

[0009] Optionally, spatial attention weights are used to enhance the channel enhancement feature map to obtain a spatial enhancement feature map, including: multiplying the spatial attention weights and the channel enhancement feature map position by position to obtain the second enhancement feature corresponding to each spatial attention head; flattening each second enhancement feature to obtain a one-dimensional sequence; and concatenating and reshaping all the one-dimensional sequence features to obtain the spatial enhancement feature map.

[0010] Optionally, the classification module is a fully connected discriminant network.

[0011] Optionally, the cross-entropy loss function is used during the training of the pre-trained crack fault diagnosis model.

[0012] To achieve the above objectives, this application also provides a crack fault identification device for a lobe mixer based on a multi-head hybrid attention mechanism, comprising: an image acquisition module for acquiring fault signals during mixer operation and performing short-time Fourier transform on the fault signals to obtain a time-frequency image; and a fault diagnosis module for inputting the time-frequency image into a pre-trained crack fault diagnosis model to obtain a crack diagnosis result; wherein the pre-trained crack fault diagnosis model includes multiple sequentially connected feature extraction modules, the last feature extraction module being sequentially connected to a first convolution module and a classification module; each feature extraction module includes a sequentially connected second convolution module and a multi-head hybrid attention feature extraction module; the multi-head hybrid... The attention feature extraction module includes: a first segmentation module, used to uniformly segment the channel dimension of the feature map to obtain the features of the channel attention heads; a channel enhancement module, used to determine the channel attention weight of each channel attention head, and use the channel attention weight to enhance the features of the feature map to obtain a channel enhanced feature map; a second segmentation module, used to obtain a one-dimensional pixel sequence from the channel enhanced feature map, and uniformly segment and reshape the one-dimensional pixel sequence to obtain the features of multiple spatial attention heads; and a spatial enhancement module, used to determine the spatial attention weight of each spatial attention head, and use the spatial attention weight to enhance the features of the channel enhanced feature map to obtain a spatial enhanced feature map.

[0013] Compared with the prior art, the beneficial effects of this application are as follows: The present invention provides a method for identifying crack faults in a beam mixer based on a multi-head hybrid attention mechanism. This method divides channel information into multiple heads and uses a channel attention mechanism to adaptively learn channel weights to obtain channel enhancement feature maps. It then divides the spatial information of the channel enhancement feature maps into multiple heads and performs local feature adaptive learning to obtain spatial enhancement feature maps. This improves the model's ability to comprehensively capture and process input data and adapts to the complex operating conditions of afterburner mixers. Furthermore, it extracts fault information from strong background noise, thereby improving the accuracy of crack fault identification in aero-engine afterburner mixers under various operating conditions. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the method for identifying crack faults in a beam mixer based on a multi-head hybrid attention mechanism, as described in this application. Figure 2 This is a schematic diagram of the crack fault diagnosis model in the crack fault identification method for a lobe mixer based on a multi-head hybrid attention mechanism in this application. Figure 3 This is a model diagram of the mixer lobe in an embodiment of this application; Figure 4 This is a diagnostic performance diagram of an embodiment of this application.

[0015] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] The first embodiment of the present invention provides a method for identifying crack faults in a beam mixer based on a multi-head hybrid attention mechanism, such as... Figure 1 As shown, the specific steps include: Step S1: Obtain fault signals during the operation of the mixer, perform short-time Fourier transform on the fault signals to obtain the corresponding time-frequency image of the fault; wherein, the fault signals include strain signals and vibration signals; Specifically, in step S11, strain gauges are attached to multiple convex lobes of the mixer to collect key strain signals. Simultaneously, vibration sensors are placed outside the casing at the afterburner location of the aero-engine to collect vibration signals. Based on the different crack states of the mixer, training and test datasets corresponding to the original strain and vibration signals are obtained, providing an input basis for subsequent feature extraction.

[0018] Step S12: Bandpass filtering is performed on the acquired strain signal and vibration signal to eliminate interference and improve data quality. Chebyshev Type I filter is used to extract data in the 5Hz-400Hz frequency band, retaining specific frequency bands related to the mixer state and removing high and low frequency irrelevant data.

[0019] Step S13 involves performing short-time Fourier transforms on the strain and vibration signals obtained in step S12 to obtain corresponding time-frequency images. The two-dimensional time-frequency image samples are then used as single-channel grayscale images input to a subsequent feature extractor to capture key frequency components and their dynamic characteristics. Through these steps, the introduction of bandpass filtering and short-time Fourier transform for data preprocessing effectively highlights the characteristics of different signals.

[0020] Step S2: Input the time-frequency image into the pre-trained crack fault diagnosis model to obtain the crack diagnosis result; like Figure 2 As shown, the pre-trained crack fault diagnosis model includes a feature extractor and a classification module. The feature extractor comprises multiple sequentially connected feature extraction modules, with the last feature extraction module connected to the first convolutional module. In the fault diagnosis framework, each task first extracts task features through the feature extractor. Assume the task... The input sample is The corresponding tag is The feature extraction module generates feature representations through a multi-scale, multi-channel attention mechanism: ; In the formula, Represents network parameters, For the task The high-quality general feature representation and feature extraction network can extract highly discriminative and generalizable features from different working conditions, providing a stable feature foundation for subsequent diagnostic tasks.

[0021] Specifically, each feature extraction module includes a second convolutional module and a multi-head hybrid attention feature extraction module connected in sequence; the second convolutional module is used to convolve the time-frequency image to obtain a feature map; the multi-head hybrid attention feature extraction module is used to extract features from the feature map. Step S2 specifically includes the following steps: Step S21: The time-frequency image is sequentially extracted using multiple feature extraction modules to obtain the first crack feature; Taking three feature extraction modules as examples, the feature extraction process will be introduced in detail.

[0022] Step S211: Perform a convolution operation on the time-frequency image using the second convolution module to obtain the first feature map; Step S212: Feature extraction is performed using the multi-head hybrid attention feature extraction module to obtain the second feature map; Specifically, in step S2121, the channel dimension of the feature map is uniformly divided to obtain features of multiple channel attention heads; for example, it can be divided into four channel attention heads. Step S2122: Determine the channel attention weights for each channel attention head, as follows: Global average pooling and global max pooling are performed on the features of each channel attention head to compress the spatial dimension, resulting in the feature vectors of the first and second channels. By sharing a fully connected layer, a fully connected operation is performed on the first channel feature vector and the second channel feature vector respectively to obtain the first feature vector and the second feature vector. The first and second feature vectors are added together, and the channel attention weights are obtained by activation, as shown in the following expression:

[0023] In the formula, Indicates the first i Channel attention weights for each channel attention head Indicates the first i Features of a spatial attention head channel attention head Indicates the number of channel attention heads. This represents the Sigmoid activation function. This indicates a multilayer perceptron within a shared fully connected layer. , These represent max pooling and average pooling, respectively.

[0024] Step S2123: Enhance the feature map using channel attention weights to obtain a channel-enhanced feature map, as follows: The channel attention weights are multiplied spatially with the feature map to obtain the first enhanced feature corresponding to each channel attention head; all the first enhanced features are concatenated according to the channel dimension to obtain the channel enhanced feature map, expressed as follows. Note: The splicing operation is omitted in this formula.

[0025] Step S2124: Flatten the channel enhancement feature map to obtain a one-dimensional pixel sequence, and uniformly divide and reshape the one-dimensional pixel sequence into multiple two-dimensional feature maps, i.e., features of the spatial attention head. Step S2125: Determine the spatial attention weights for each spatial attention head, specifically as follows: Perform channel-dimensional average pooling and channel-dimensional max pooling on the features of each spatial attention head to obtain the first spatial feature vector and the second spatial feature vector. The first and second spatial feature vectors are concatenated according to the channel dimension, and then convolution and activation operations are performed to obtain the spatial attention weights, expressed as follows:

[0026] In the formula, Indicates the first j Spatial feature weights for each spatial attention head The function representation of a convolutional layer, For the first j Features of spatial attention heads.

[0027] Step S2126: Perform feature enhancement on the channel enhancement feature map using spatial attention weights to obtain the spatially enhanced feature map, as follows: Multiply the spatial attention weights and channel enhancement feature maps position by position to obtain the second enhancement feature corresponding to each spatial attention head; Flattening each second enhancement feature yields a one-dimensional sequence. Concatenating all one-dimensional sequence features and reshaping them into a two-dimensional feature map, i.e., a spatial enhancement feature map. The expression is:

[0028] Step S213: Repeat steps S211-S212 twice to obtain the first crack feature; Step S22: Perform a convolution operation on the crack features using the first convolution module to obtain the second crack features, i.e., the task... High-quality general features ; Step S23: Classify the second crack features using the classification module to obtain the crack diagnosis result.

[0029] The classification module is a fully connected discriminant network.

[0030] Furthermore, the pre-trained crack fault diagnosis model employs a cross-entropy loss function during training, using vibration and strain signals as inputs to reflect fault characteristics from different perspectives. The goal of model training is to minimize the loss value. The gradient of the loss function provides optimization guidance for parameter update tasks. Backpropagation converges the loss function value, enabling the model to achieve stable performance. The loss function is calculated using the true labels and the discriminant labels from the discriminant network, and its expression is:

[0031] In the formula, It is the number of categories. It's a real label. It is the probability predicted by the model.

[0032] By employing the above method within a deep learning framework, a multi-head hybrid attention mechanism is introduced to enhance key features, thereby improving the accuracy of crack fault identification in aero-engine afterburner mixers under various operating conditions. Specifically, channel information is divided into multiple heads, and the channel weights are adaptively learned through a channel attention mechanism to obtain updated feature maps for each channel. Spatial information is then divided into multiple heads, and local feature adaptive learning is performed to obtain updated spatial feature maps. By introducing multiple parallel attention heads, the model's ability to comprehensively capture and process input data is enhanced, adapting to the complex operating conditions of afterburner mixers, thereby extracting fault information from strong background noise and improving the accuracy of crack fault identification.

[0033] A second embodiment of the present invention provides a beam mixer crack fault identification device based on a multi-head hybrid attention mechanism, comprising: The image acquisition module acquires signals during the mixer's operation and performs a short-time Fourier transform on the signals to obtain a time-frequency image. The fault diagnosis module inputs the time-frequency image into a pre-trained crack fault diagnosis model to obtain crack diagnosis results. The pre-trained crack fault diagnosis model includes multiple sequentially connected feature extraction modules, with the last feature extraction module sequentially connected to a first convolution module and a classification module. Each feature extraction module includes a sequentially connected second convolution module and a multi-head hybrid attention feature extraction module. The multi-head hybrid attention feature extraction module includes: a first segmentation module, used to uniformly segment the channel dimension of the feature map to obtain features from multiple channel attention heads; a channel enhancement module, used to determine the channel attention weights of each channel attention head and use these weights to enhance the feature map, obtaining a channel-enhanced feature map; a second segmentation module, used to flatten the channel-enhanced feature map to obtain a one-dimensional pixel sequence, uniformly segmenting and reshaping the one-dimensional pixel sequence to obtain features from multiple spatial attention heads; and a spatial enhancement module, used to determine the spatial attention weights of each spatial attention head and use these weights to enhance the channel-enhanced feature map, obtaining a spatially enhanced feature map.

[0034] Example The crack fault identification method of this invention was tested using data collected from a certain type of aero-engine test bench. Test vibration data was measured using an accelerometer mounted outside the casing; test strain data was measured using strain gauges mounted outside the mixer lobe, with a sampling frequency set to 20kHz. The monitoring positions of the strain gauges are shown in the diagram. Figure 3 As shown, the crack diagnosis results for the afterburner mixer, including 8 signal acquisition locations, are as follows: normal state, 5mm crack, 10mm crack and 15mm crack fault.

[0035] Four state samples are used, each containing data from n monitoring points. Each monitoring point outputs strain data. Simultaneously, vibration data for each state is monitored outside the casing. Therefore, the data for each state consists of 2*n+1 samples, with each sample containing 2400 data points. Considering the different dimensions of the monitored parameters and the significant differences in their numerical values, the proposed method uses Z-score standardization to standardize each monitoring parameter. The standardized value for each parameter is 0, with a standard deviation of 1. The standard deviation is shown in the following formula:

[0036] In the formula, and These are the mean and standard deviation, respectively. To prevent data leakage, the training and test sets are strictly isolated. Parameters fitted on the training set are used, and the training, validation, and test sets are uniformly transformed. That is: Training set:

[0037] Validation set:

[0038] Test set:

[0039] Fault Diagnosis Task Construction: This embodiment uses data collected from a certain type of aero-engine test bench. The mixer status data includes four types of crack diagnosis results: normal state, 5mm crack, 10mm crack, and 15mm crack fault. Considering that the strain data of different measuring points of the mixer under the same crack state actually represent a state, they are set to the same label; at the same time, the vibration data of the casing under the same state is also set to this label. The ratio of training set to test set data is set to 7:3. The final constructed diagnostic task is shown in Table 1.

[0040] Table 1. Detailed Data of the Example

[0041] After training, the model parameters are saved, and diagnostics are performed on the test data to verify the performance of the saved model. Its diagnostic performance on the test data is as follows: Figure 4 As shown, it can be seen that it can achieve good diagnostic accuracy for all four types of data, indicating that the transfer network model constructed using this method has excellent bearing fault diagnosis capabilities.

[0042] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for crack fault identification of a lobe mixer based on a multi-head hybrid attention mechanism, characterized in that, include: Fault signals during mixer operation are acquired, and short-time Fourier transform is performed on the fault signals to obtain a time-frequency image; The time-frequency image is input into a pre-trained crack fault diagnosis model to obtain crack diagnosis results; The pre-trained crack fault diagnosis model includes multiple sequentially connected feature extraction modules, with the last feature extraction module being sequentially connected to a first convolution module and a classification module. Each of the feature extraction modules includes a second convolutional module and a multi-head hybrid attention feature extraction module connected in sequence; The second convolution module is used to convolve the time-frequency image to obtain a feature map; The multi-head hybrid attention feature extraction module is used to extract features from the feature map. The extraction process includes: The feature map is uniformly divided along its channel dimensions to obtain features from multiple channel attention heads; Determine the channel attention weight for each channel attention head, and use the channel attention weight to enhance the feature map to obtain a channel-enhanced feature map; Flatten the channel enhancement feature map to obtain a one-dimensional pixel sequence, and uniformly divide and reshape the one-dimensional pixel sequence to obtain features of multiple spatial attention heads; Determine the spatial attention weight for each spatial attention head, and use the spatial attention weight to perform feature enhancement on the channel enhancement feature map to obtain the spatial enhancement feature map.

2. The method for identifying crack faults in a beam mixer based on a multi-head hybrid attention mechanism according to claim 1, characterized in that, Determining the channel attention weight for each channel attention head includes: Global average pooling and global max pooling are performed on the features of each channel attention head to obtain the first channel feature vector and the second channel feature vector; Perform a fully connected operation on the first channel feature vector and the second channel feature vector respectively to obtain the first feature vector and the second feature vector; The first feature vector and the second feature vector are added together, and the channel attention weights are activated.

3. The method for identifying crack faults in a lobe mixer based on a multi-head hybrid attention mechanism according to claim 1 or 2, characterized in that, The step of using the channel attention weights to enhance the feature map to obtain a channel-enhanced feature map includes: The channel attention weights are multiplied spatially by the feature map to obtain the first enhanced feature corresponding to each channel attention head; All the first enhancement features are concatenated according to the channel dimension to obtain the channel enhancement feature map.

4. The method for identifying crack faults in a beam mixer based on a multi-head hybrid attention mechanism according to claim 1, characterized in that, Determining the spatial attention weight for each spatial attention head includes: For each spatial attention head, channel-dimensional average pooling and channel-dimensional max pooling are performed respectively to obtain a first spatial feature vector and a second spatial feature vector. The first spatial feature vector and the second spatial feature vector are concatenated according to the channel dimension, and then convolution and activation operations are performed to obtain spatial attention weights.

5. The method for identifying crack faults in a beam mixer based on a multi-head hybrid attention mechanism according to claim 1 or 4, characterized in that, The step of using the spatial attention weights to enhance the channel enhancement feature map to obtain a spatially enhanced feature map includes: Multiply the spatial attention weights and channel enhancement feature maps position by position to obtain the second enhancement feature corresponding to each spatial attention head; Flatten each second enhancement feature to obtain a one-dimensional sequence, then concatenate and reshape all the one-dimensional sequence features to obtain a spatial enhancement feature map.

6. The method for identifying crack faults in a beam mixer based on a multi-head hybrid attention mechanism according to claim 1, characterized in that, The classification module is a fully connected discriminant network.

7. The method for identifying crack faults in a beam mixer based on a multi-head hybrid attention mechanism according to claim 1, characterized in that, The pre-trained crack fault diagnosis model uses a cross-entropy loss function during training.

8. A crack fault identification device for a beam mixer based on a multi-head hybrid attention mechanism, characterized in that, include: The image acquisition module acquires fault signals during the operation of the mixer, performs a short-time Fourier transform on the fault signals, and obtains a time-frequency image. The fault diagnosis module is used to input the time-frequency image into a pre-trained crack fault diagnosis model to obtain crack diagnosis results; The pre-trained crack fault diagnosis model includes multiple sequentially connected feature extraction modules, with the last feature extraction module being sequentially connected to a first convolution module and a classification module. Each of the feature extraction modules includes a second convolutional module and a multi-head hybrid attention feature extraction module connected in sequence; The multi-head hybrid attention feature extraction module includes: The first segmentation module is used to uniformly segment the channel dimension of the feature map to obtain the features of the channel attention head; The channel enhancement module is used to determine the channel attention weight of each channel attention head, and to enhance the feature map using the channel attention weight to obtain a channel-enhanced feature map. The second segmentation module is used to flatten the channel enhancement feature map to obtain a one-dimensional pixel sequence, and to uniformly segment and reshape the one-dimensional pixel sequence to obtain features of multiple spatial attention heads. The spatial enhancement module is used to determine the spatial attention weight of each spatial attention head, and to perform feature enhancement on the channel enhancement feature map using the spatial attention weight to obtain the spatial enhancement feature map.

Images