Agent attention enhancement diffusion aero-engine small sample fault diagnosis method

Abstract

The invention provides an aero-engine small sample fault diagnosis method for agency attention enhancement diffusion, and belongs to the technical field of aero-engine gas circuit fault diagnosis. According to the method, the proxy attention is added on the basis of the DDIM, so that the sample generation can be accelerated, the sample generation quality is obviously improved, and no extra training is needed. Inputting a small amount of pre-processed collected sensor signal data into the agent attention enhanced denoising diffusion implicit model for data enhancement to generate high-quality sample data so as to make up for the deficiency of an original data set and improve the performance of the model under the condition of scarce data; the expanded data set is used as a training set, the original data set is used as a test set for model training, the test set is used for testing the performance and accuracy of the model after training is completed, and a final aero-engine fault diagnosis model is established. According to the method, the problem of insufficient data in fault diagnosis is solved, and the accuracy and robustness of the model are improved.

Description

Technical Field

[0001] The invention relates to an aircraft engine sensor signal fault diagnosis method, and in particular to an aircraft engine small sample fault diagnosis method with agent attention enhanced diffusion. Background Art

[0002] As a complex mechanical system, the fault diagnosis of aircraft engines relies on a large amount of data support. However, in practical applications, due to the high cost, high complexity and low failure rate of aircraft engines, it is often difficult to obtain sufficient fault sample data, which greatly limits the training effect and practical application of fault diagnosis models.

[0003] In recent years, diffusion models have made significant progress in deep generative models, especially in the field of image generation. Diffusion models can generate high-quality sample data through forward denoising and reverse denoising processes. Among them, DDIM (denoising diffusion implicit model) as an improved version of DDPM (denoising diffusion probabilistic model) achieves faster sampling speed and deterministic output through non-Markov reverse process, further improving generation efficiency and quality.

[0004] Attention mechanism has become an important means to improve model performance. In particular, the proxy attention mechanism combines the advantages of Softmax attention and linear attention, reduces computational complexity and retains global modeling capabilities by introducing proxy vectors. This method has achieved remarkable results in visual tasks such as image classification and detection, but its application in the field of aircraft engine fault diagnosis is still blank.

[0005] In actual diagnosis scenarios, due to the installation restrictions of gas path sensors, their measurement parameters are limited, and there is a lot of measurement noise. There are usually problems such as a small number of fault samples, missing fault sample labels, and an unbalanced number of various fault samples. As a result, the engine fault data set is usually insufficient. Existing fault diagnosis technologies usually require a large number of labeled data samples that are balanced between classes. When faced with insufficient data sets, due to the lack of sufficient data, the learning ability of the model decreases, and the generalization performance is also affected, resulting in poor engine fault diagnosis application results under insufficient data conditions. Therefore, studying how to perform engine fault diagnosis in hybrid systems under insufficient data conditions is crucial to promoting the development of intelligent fault diagnosis technology. Summary of the invention

[0006] The present invention proposes a method for adding a proxy attention mechanism to the up- and down-sampling process of a U-Net network in an implicit diffusion model DDIM, which is used to enhance the fault diagnosis capability of aircraft engines under small sample data. By introducing the proxy attention mechanism, the present invention aims to improve the feature retention capability of the diffusion model during up- and down-sampling, reduce computational complexity, and improve the generalization performance of the model under small sample data. Furthermore, the enhanced data is generated using the improved diffusion model, and fault diagnosis is performed in combination with the classification model to improve the accuracy and reliability of the diagnosis.

[0007] The technical solution of the present invention:

[0008] An aircraft engine small sample fault diagnosis method based on agent attention enhanced diffusion, the method comprising:

[0009] (1) Collect sensor data of aircraft engines;

[0010] (2) We fuse Softmax attention and linear attention to construct a proxy attention module, and apply the proxy attention module to the up- and down-sampling process of the core network U-Net of the implicit diffusion model DDIM. We use the proxy attention-enhanced DDIM to accelerate the generation of enhanced data and improve the quality of data generation.

[0011] (3) Use the generated enhanced data to train the fault diagnosis model;

[0012] (4) Apply the trained fault diagnosis model to perform fault diagnosis on small sample data.

[0013] In step (2), the proxy attention module is implemented in the following way: based on introducing a set of additional proxy vectors A in the attention triple (Q, K, V), a four-tuple attention mechanism (Q, A, K, V) is formed. First, A is used as the query vector, and attention calculation is performed between A, K and V, and information is gathered from all features to obtain the proxy feature. This step is equivalent to using the Softmax attention mechanism to calculate the similarity between A and K, and weighted sum V according to the similarity to obtain the proxy feature; then, A is used as the key vector, the proxy feature is used as the value vector, and Q is used for the second attention calculation, the global information in the proxy feature is broadcast back to each feature, and the final output O is obtained. This step also uses the Softmax attention mechanism to calculate the similarity between Q and A, and weighted sum the proxy features according to the similarity to obtain the final output O.

[0014] In the up-sampling and down-sampling process of the noise prediction network U-Net model of DDIM, a proxy attention module is introduced. Its training process includes: a small number of fault samples in the training data set are input into the diffusion model, and the diffusion model adds noise ε to the fault sample with time step t to generate the corresponding noise sample. The noise sample is input into the neural network U-Net model, and the proxy attention module is added after the convolution block of each stage of the U-Net model; the U-Net model predicts the noise component at each time step t, and removes these noises from the noise sample to gradually restore the data. The U-Net model is trained by minimizing the difference between the predicted noise and the actual noise, that is, the denoising ability of the U-Net model is improved by optimizing the loss function. Among them, the backbone network of the diffusion model has an encoder-decoder architecture, including symmetrical downsampling and upsampling paths, and jump connections.

[0015] The fault diagnosis model is a classifier based on a deep convolutional neural network. First, the low-level, medium-level and high-level features of the input data are gradually extracted using multi-layer convolution operations. Each convolution layer slides on the input data through the convolution kernel, calculates the dot product of each position and generates the corresponding feature map to obtain the local features of the input image. The first convolution layer extracts basic features, and the subsequent convolution layers gradually extract higher-level features. After the convolution operation, nonlinear features are introduced through nonlinear activation functions. After that, the pooling layer downsamples the features of the local area to reduce the spatial dimension of the data while retaining the most important information. After multiple layers of convolution and pooling are alternately performed, more abstract feature representations are gradually extracted to form the final feature vector. Next, the fully connected layer maps these features to the classification space and converts them into category probability distributions. Each fully connected layer realizes nonlinear transformation through matrix operations and activation functions, in which each neuron is connected to all neurons in the previous layer, and finally generates output. The output layer is used for classification and calculation of loss, and the output of the last fully connected layer is converted into the probability distribution of the target category. Through the Softmax activation function, the output layer generates a probability distribution, compares it with the true label, and calculates the loss through the loss function. In the whole network training, the loss function plays a guiding role, optimizing the parameters of the network to reduce the classification error. The fault diagnosis model is trained with the enhanced data to improve the diagnostic accuracy of small sample data.

[0016] The fault diagnosis method further comprises the following steps: before inputting into the model, the sensor data is preprocessed, including similarity analysis, characteristic factor screening, tensor quantization, and normalization.

[0017] The fault diagnosis method also includes the following steps: after the classification model is trained, a performance evaluation is performed on it, and the model parameters are adjusted according to the evaluation results to optimize the diagnosis effect.

[0018] Beneficial effects of the present invention: The present invention adds proxy attention on the basis of DDIM, so that it can accelerate sample generation and significantly improve the quality of sample generation without any additional training. A small amount of preprocessed collected sensor signal data is input into the denoising diffusion implicit model enhanced by proxy attention for data enhancement to generate high-quality sample data to make up for the deficiencies of the original data set and improve the performance of the model in the case of scarce data; the expanded data set is used as a training set, the original data set is used as a test set, and the training set is used as the input of the convolutional neural network for model training. After the training is completed, the performance and accuracy of the model are tested with the test set to establish the final aircraft engine fault diagnosis model. The present invention solves the problem of insufficient data in fault diagnosis and improves the accuracy and robustness of the model by performing data enhancement based on the diffusion model enhanced by the proxy attention mechanism. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Flow chart of the small sample fault diagnosis method for aero-engine gas path system based on agent attention enhanced denoising diffusion implicit model;

[0020] Figure 2 Schematic diagram of the architecture of the small sample fault diagnosis method for the air path system of an aircraft engine based on the attention enhanced denoising diffusion implicit model;

[0021] Figure 3 Flow chart of the small sample fault diagnosis method for aero-engine gas path system based on attention enhanced denoising diffusion implicit model;

[0022] Figure 4 Schematic diagram of the agent attention architecture;

[0023] Figure 5 It is a schematic diagram of the data enhancement subsystem structure;

[0024] Figure 6 Schematic diagram of pseudo code for training and generating samples for the data augmentation subsystem;

[0025] Figure 7 This is a schematic diagram of the convolutional neural network classification model structure;

[0026] Figure 8 To generate a sample schematic. DETAILED DESCRIPTION

[0027] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0028] Reference Figure 1 and Figure 2,The present invention enhances the fault diagnosis accuracy of aircraft engines in the case of small samples by enhancing the aircraft engine sensor data. Specifically, by introducing proxy attention, the feature retention ability of the diffusion model in the up-sampling and down-sampling processes is improved, the computational complexity is reduced, the sample generation is accelerated and the sample generation quality is significantly improved. Furthermore, the enhanced data is generated by using the improved diffusion model, and the fault diagnosis is performed in combination with the classification model to improve the accuracy and reliability of the diagnosis.

[0029] Reference Figure 3 , an aircraft engine small sample fault diagnosis method based on agent attention enhanced diffusion, comprising the following steps:

[0030] 101: Collect sensor data of aircraft engine gas path system, capture multiple parameters including temperature, pressure, flow, etc. in real time, and ensure data accuracy and completeness;

[0031] 102: Perform similarity analysis, feature factor screening, tensor quantization, normalization and other preprocessing on the collected sensor data of the aircraft engine gas path system to provide efficient and accurate data support for subsequent model training and diagnosis;

[0032] 103: Integrate Softmax attention and linear attention to build a proxy attention module, which has low computational complexity and high model expression ability, and adopts a larger receptive field without increasing the amount of model calculation, providing strong support for subsequent data generation and fault diagnosis;

[0033] 104: Apply the proxy attention module to the up- and down-sampling process of U-Net, and use the proxy attention-enhanced DDIM to accelerate the generation of enhanced data and significantly improve the data generation quality without any additional training;

[0034] (4.1) The proxy attention module consists of two Softmax attention operations, which combines softmax attention with linear attention. Specifically, we initially treat the proxy token A as the query and perform attention calculation between A, K, and V to aggregate the proxy features V from all values A . We then use A as the key in the second attention calculation, V A As the value and query matrix Q, the global information of the proxy feature is broadcast to each query token and the final output O is obtained. In this way, we avoid calculating the similarity between Q and K, while retaining the information exchange between each query key pair through the proxy token. By default, the proxy vector A is obtained by pooling, that is, A = Pooling (Q). The proxy attention can be expressed as:

[0035] O=Attn s (Q,A,Attns (A,K,V)

[0036] is equivalent to:

[0037]

[0038] Among them, O represents the output feature of the proxy attention module, Attn s (·) represents the attention calculation, σ(·) represents the activation function softmax calculation, φ q (Q) = σ(QA T ) is used as the query vector in the linear attention calculation, φ q (K)=σ(AK T ) is used as the key vector in the linear attention calculation, represents a linear attention computation.

[0039] In order to make better use of position information, a proxy bias is introduced in the proxy attention, namely:

[0040] O=σ(QA T +B2)σ(AK T +B1)V

[0041] Among them, B1 and B2 are two proxy deviations.

[0042] Considering that the model suffers from insufficient feature diversity, a deep convolution (DWC) module is introduced to retain feature diversity, namely:

[0043] O=σ(QA T +B2)σ(AK T +B1)V+DWC(V)

[0044] Proxy attention seamlessly combines powerful Softmax attention and efficient linear attention, while enjoying low computational complexity and high model expressiveness, and adopts a larger receptive field without increasing the model computation. When applied to the diffusion model, proxy attention accelerates generation and significantly improves image generation quality without any additional training.

[0045] (4.2) Data enhancement method based on improved diffusion model

[0046] In data enhancement and feature extraction, the up-sampling and down-sampling processes are key steps. Up-sampling is usually used to increase the resolution of data, while down-sampling is used to reduce the dimension of data. In diffusion models, the up-sampling and down-sampling processes have an important impact on the quality and efficiency of generated data. However, traditional up-sampling and down-sampling methods often fail to effectively retain the key features of the data, resulting in a decrease in the quality of generated data. In the up-sampling and down-sampling stages of the DDIM model, a proxy attention mechanism is introduced to weight the feature map through proxy vectors, enhance the model's ability to capture key features, reduce information loss, and reduce computational complexity. The DDIM model with the proxy attention module is used to generate and enhance data to generate more diverse fault sample data. Through the reverse denoising process, clear fault features are gradually restored from the noisy data to improve the quality of generated data.

[0047] The processing process of DDIM with agent attention enhancement: the preprocessed data is input into the improved diffusion model. In the forward process, noise is gradually added through multiple time steps to obtain a pure noise sample x close to the standard Gaussian distribution. t In the reverse process, time step t is compared with the noise sample x t Input into the U-Net network enhanced by proxy attention, predict the noise, and output a denoised sample. Repeat the above denoising process until the final generated sample is obtained. Through multiple forward denoising and reverse denoising processes, a sufficient amount of enhanced data is generated. Figure 5 shown.

[0048] Noise data is generated through the forward denoising process, and clear fault features are gradually restored through the reverse denoising process. Repeat the above process to generate a sufficient amount of enhanced data. Figure 5 shown.

[0049] 4.2.1. Forward diffusion process

[0050] In the forward diffusion process, the present invention gradually adds a certain variance β at each time step t. t The Gaussian noise of is used to corrupt the original sample x0 to generate noise samples x1, x2, ...x t ,...x T . The process can be expressed as:

[0051]

[0052] Among them, q(x t |x t-1 ) indicates that at the noise level β t Next from x t-1 Generate x t The conditional distribution of is:

[0053]

[0054] Among them, the noise level β t ∈(0,1) is a predefined hyperparameter sequence that satisfies 0<β1<β2<…<β T <1, I is the identity matrix.

[0055] In the process of continuously adding noise, the previously complete image data is transformed into a completely noisy and indistinguishable image x t This process relies on a Markov chain, which is mathematically described as follows:

[0056]

[0057] Among them, α t =1-β t , ε is Gaussian distributed noise,

[0058] 4.2.2. Reverse denoising process

[0059] Different from the generation process based on Markov chain which requires multiple iterations, DDIM proposes a de-Markov sampling process. The training methods of DDIM and DDPM are the same, so the DDIM sampling scheme can be added to DDPM. The training and sampling process is as follows: Figure 6 As shown. Based on the assumption that the conditional denoising distribution to be learned:

[0060] p(x t-1 |x t ,x0)~N(kx0+mx t ,σ 2 )

[0061] x t-1 =kx0+mx t +σε,ε~N(0,1)

[0062] The noise addition formula Substitute in and simplify:

[0063]

[0064] According to the forward noise formula, we can get

[0065] but

[0066] After sorting, the reverse process expression after accelerated sampling is:

[0067]

[0068] Among them, x sis the denoised image obtained in step s, x k is the image obtained in step k, σ is defined as p(x t-1 |x t ,x0),∈ θ (x k ,k) is the data x at time step k k The predicted value of the noise.

[0069] Therefore, the neural network ∈ θ (x t ,t) is defined as minimizing the following loss function:

[0070]

[0071] 4.2.33. Proxy attention mechanism introduces upsampling and downsampling process

[0072] The present invention adopts U-Net as a noise prediction model, and the output is the predicted noise. The number of noise addition steps t and the current noise image are input to the U-Net network, and the model performance is gradually optimized through repeated iterative training. The U-Net network can accurately predict and remove the noise in the image and restore the original image. During the training process, by continuously adjusting the difference between the predicted noise and the actual noise, the model gradually learns how to accurately identify and remove noise, thereby generating high-quality samples, providing a more reliable basis for subsequent fault diagnosis. The specific U-Net network is introduced as follows:

[0073] (1) Encoder (downsampling path):

[0074] Convolutional layer and pooling layer: The downsampling path consists of multiple layers of convolutional layers and pooling layers. Each convolution operation extracts local information of the input features. Convolution operations are usually accompanied by nonlinear activation functions and batch normalization operations.

[0075] Feature compression: Through the maximum pooling operation, each downsampling gradually reduces the spatial dimension of the feature map, but increases its number of channels to extract more high-level abstract features.

[0076] Noise Conditional Injection: Temporal embedding is introduced during the U-Net downsampling process to incorporate the time step information of the noise addition into the feature map so that the model can make predictions based on different noise levels.

[0077] (2) Skip Connection:

[0078] The output of each downsampling path is directly connected to the input of the corresponding upsampling path. The skip connection retains the feature information of the encoder and passes it to the decoder, ensuring that low-level detail features can be referenced during the upsampling process to help restore the details of the original data.

[0079] (3) Decoder (up-sampling path):

[0080] Deconvolution and upsampling: The upsampling path is symmetrical to the downsampling path, and the step-by-step interpolation operation increases the spatial dimension of the feature map and restores the size of the input data. After each upsampling operation, a convolution layer and activation function are added to gradually restore the image features.

[0081] Feature fusion: In the upsampling stage of each layer, the skip connection fuses the encoder features of the corresponding layer with the decoder features, helping the model to comprehensively consider low-level and high-level features.

[0082] (4) Output layer:

[0083] After multiple layers of upsampling, the last 1x1 convolutional layer of the decoder maps the high-dimensional features to the output space and predicts the denoising result or feature of each pixel. The diffusion model generates or restores data by removing noise.

[0084] (5) Time Embedding:

[0085] Temporal embedding usually encodes the time step through a fully connected layer and fuses it with the feature map. The time step information in the diffusion process is embedded into the U-Net network to help the model make denoising decisions based on the current noise level.

[0086] (6) Residual Blocks:

[0087] In the U-Net design, residual blocks are used between convolutional layers to stabilize the training process and enhance the expressiveness of the model.

[0088] (7) Attention Block

[0089] In the U-Net design, attention blocks are used between convolutional layers, that is, adding a proxy attention mechanism during upsampling and downsampling enables the model to enjoy both low computational complexity and high model expression capabilities.

[0090] 105: In order to improve the accuracy of diagnosis, by introducing enhanced data samples and training the fault diagnosis model, the fault diagnosis model can better learn the characteristics and laws of engine faults, thereby showing higher accuracy and robustness in fault diagnosis;

[0091] Based on deep convolutional networks, such as Figure 7As shown in the figure, multi-layer convolution operations are used to gradually extract low-level, medium-level and high-level features of the input data. The convolution operation is to slide the convolution kernel on the image and calculate the dot product at each position to generate a feature map. The first convolution layer extracts basic features, and subsequent convolution layers gradually extract higher-level features. After the convolution operation, nonlinear features are introduced through nonlinear activation functions. After that, the pooling layer downsamples the features of the local area to generate a low-dimensional feature vector while retaining the most important information. Next, the fully connected layer maps these features to the classification space and converts them into category probability distributions. Each fully connected layer implements nonlinear transformation through matrix operations and activation functions, where each neuron is connected to all neurons in the previous layer and finally generates output. The output layer is used for classification and calculation of loss, and the output of the last fully connected layer is converted into a probability distribution of the target category. Through the Softmax activation function, the output layer generates a probability distribution and compares it with the true label, and the loss is calculated through the loss function. The classifier is trained with augmented data to improve the diagnostic accuracy of small sample data.

[0092] 106: Input the fault data to be diagnosed into the trained fault diagnosis model, which will quickly and accurately identify the fault based on its learned knowledge and experience. By outputting the diagnosis results, it provides maintenance personnel with clear fault location and repair suggestions, thereby effectively shortening maintenance time, reducing maintenance costs, and improving the operating efficiency and safety of aircraft engines.

[0093] The present invention introduces proxy attention into the up- and down-sampling process of U-Net, achieving a favorable balance between computational efficiency and model expressiveness. The improved model can adopt a larger receptive field, improving the speed and quality of sample generation without additional training. The method of attention-enhanced diffusion model has significant advantages in generating a wider range of high-quality samples, improving the accuracy of fault diagnosis of small sample data.

Claims

1. A method for small sample fault diagnosis of aircraft engines with agent attention enhanced diffusion, characterized in that: The method comprises: (1) Collect sensor data of aircraft engines; (2) We fuse Softmax attention and linear attention to construct a proxy attention module, and apply the proxy attention module to the up- and down-sampling process of the core network U-Net of the implicit diffusion model DDIM. We use the proxy attention-enhanced DDIM to accelerate the generation of enhanced data and improve the quality of data generation. (3) Use the generated enhanced data to train the fault diagnosis model; (4) Apply the trained fault diagnosis model to perform fault diagnosis on small sample data.

2. The method for diagnosing small sample faults of an aircraft engine according to claim 1, characterized in that: In step (2), the proxy attention module is implemented in the following manner: based on introducing a set of additional proxy vectors A in the attention triple (Q, K, V), a four-tuple attention mechanism (Q, A, K, V) is formed; first, A is used as the query vector, attention calculation is performed between A, K and V, information is aggregated from all features, and proxy features are obtained. This step is equivalent to using the Softmax attention mechanism to calculate the similarity between A and K, and weighted sum V according to the similarity to obtain the proxy feature; Then, A is used as the key vector and the proxy feature as the value vector, and a second attention calculation is performed with Q. The global information in the proxy feature is broadcast back to each feature, and the final output O is obtained. This step also uses the Softmax attention mechanism to calculate the similarity between Q and A, and weighted sum the proxy features according to the similarity to obtain the final output O.

3. The method for diagnosing small sample faults of an aircraft engine according to claim 1, characterized in that: In the up-sampling and down-sampling process of the noise prediction network U-Net model of DDIM, a proxy attention module is introduced, and its training process includes: a small number of fault samples in the training data set are input into the diffusion model, and the diffusion model adds noise ε to the fault samples with time step t to generate corresponding noise samples; the noise samples are input into the neural network U-Net model, and the proxy attention module is added after the convolution block of each stage of the U-Net model; the U-Net model predicts the noise components at each time step t, and removes these noises from the noise samples to gradually restore the data; the U-Net model is trained by minimizing the difference between the predicted noise and the actual noise, that is, the denoising ability of the U-Net model is improved by optimizing the loss function; wherein, the backbone network of the diffusion model has an encoder-decoder architecture, including symmetrical down-sampling and up-sampling paths, and jump connections.

4. The method for diagnosing small sample faults of an aircraft engine according to claim 1, characterized in that: The fault diagnosis model is a classifier based on a deep convolutional neural network. First, a multi-layer convolution operation is used to gradually extract low-level, medium-level and high-level features of the input data. Each convolution layer slides on the input data through a convolution kernel, calculates the dot product of each position and generates a corresponding feature map to obtain the local features of the input image. The first convolutional layer extracts basic features, and subsequent convolutional layers gradually extract higher-level features; after the convolution operation, nonlinear features are introduced through nonlinear activation functions; then, the pooling layer downsamples the features of the local area to reduce the spatial dimension of the data while retaining the most important information; after multiple layers of convolution and pooling are performed alternately, more abstract feature representations are gradually extracted to form the final feature vector; next, the fully connected layer maps these features to the classification space and converts them into category probability distributions; each fully connected layer implements nonlinear transformations through matrix operations and activation functions, in which each neuron is connected to all neurons in the previous layer to finally generate outputs; the output layer is used for classification and calculation of losses, and the output of the last fully connected layer is converted into a probability distribution of the target category; through the Softmax activation function, the output layer generates a probability distribution, compares it with the true label, and calculates the loss through the loss function; in the entire network training, the loss function plays a guiding role, optimizing the parameters of the network to reduce classification errors; the fault diagnosis model is trained using enhanced data to improve the diagnostic accuracy of small sample data.

5. The method for diagnosing small sample faults of an aircraft engine according to claim 1, characterized in that: The fault diagnosis method further comprises the following steps: before inputting into the model, the sensor data is preprocessed, including similarity analysis, characteristic factor screening, tensor quantization, and normalization.

6. The method for diagnosing small sample faults of an aircraft engine according to claim 1, characterized in that: The fault diagnosis method also includes the following steps: after the classification model is trained, a performance evaluation is performed on it, and the model parameters are adjusted according to the evaluation results to optimize the diagnosis effect.

Images