Intelligent aero-engine bearing fault migration diagnosis method based on multi-source data and attention mechanism

Abstract

The invention discloses an intelligent aero-engine bearing fault migration diagnosis method based on multi-source data and an attention mechanism, and the method comprises the steps: automatically learning the importance difference of different modes in different fault states through an attention weight, and achieving the dynamic optimization combination of mode features; meanwhile, an attention mechanism is utilized to capture deep correlation characteristics among multi-modal data of the aero-engine bearing part, space-time coupling relations among heterogeneous signals such as vibration, temperature and current are excavated, and an attention weight is utilized to automatically suppress modal contribution polluted by noise, so that noise robustness is enhanced; according to the method, the weight difference under different modes and different fault states can be automatically learned, the deep correlation of the modes is captured, the noise pollution is inhibited, and the robustness of the model is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of aircraft engine bearing fault diagnosis, and in particular to an aircraft engine bearing fault intelligent migration diagnosis method based on multi-source data and an attention mechanism. Background Art

[0002] Industrial equipment fault diagnosis technology is a key component in ensuring the reliable operation of modern intelligent manufacturing systems. Data-driven fault diagnosis methods offer significant advantages over traditional mechanism-based approaches in this area. Their core value lies in establishing diagnostic models by mining the inherent patterns in data, rather than relying on prior physical knowledge. Data-driven fault diagnosis utilizes an end-to-end architecture, directly mapping raw sensor data (such as vibration waveforms, current signals, and infrared thermal images) to fault categories. This eliminates the manual feature extraction (such as FFT and wavelet transforms) and multi-stage modeling required in traditional methods, making it more efficient in the field of aircraft engine fault diagnosis.

[0003] With the rapid development of sensing technology, diagnostic methods based on multi-source data fusion have gradually become a research hotspot in the field of aircraft engine bearing fault diagnosis. Compared with single-modal methods, the core advantage of multimodal fault diagnosis lies in achieving more comprehensive and robust equipment status perception by fusing multi-source heterogeneous sensor data. Multimodal fusion can effectively overcome the perception limitations of a single data source and demonstrates significant advantages in information complementarity, anti-interference capabilities, and feature expression capabilities. This collaborative perception mechanism not only improves the distinguishability of fault characteristics but also enhances the diagnostic system's fault tolerance to practical issues such as noise and missing data. It can also more accurately identify complex fault modes, thereby comprehensively improving the accuracy, reliability, and applicability of fault diagnosis. Its essence is to construct a more complete equipment status representation space through multi-dimensional data correlation analysis.

[0004] Among multi-source data fusion methods, direct concatenation or weighted averaging (Gao J, Li P, Chen Z, et al. A survey on deep learning for multimodal data fusion [J]. Neural Computation, 2020, 32 (5): 829-864.) ignores the heterogeneity and dynamic correlation between modalities, resulting in feature space redundancy or information loss. Secondly, static weight allocation (Zhao F, Zhang C, Geng B. Deep multimodal data fusion [J]. ACM computing surveys, 2024, 56 (9): 1-36.) cannot adapt to the dynamic changes in the contribution of each modality at different fault stages, affecting diagnostic sensitivity. Finally, it lacks robustness to noise and missing data, and is easily degraded by the anomaly of a single modality. This coarse-grained fusion method makes it difficult to fully tap the collaborative value of multimodal data. Therefore, how to deeply mine modal correlation patterns and construct an effective multimodal data fusion method remains a challenge.

[0005] In summary, unimodal methods suffer from information limitations and poor robustness compared to multimodal methods. While multimodal methods offer many advantages, they face challenges in deeply mining modal correlation patterns and constructing effective multimodal data fusion methods. Therefore, fusion methods incorporating attention mechanisms have emerged as an effective solution to the challenges of multi-source data fusion. Summary of the Invention

[0006] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide an intelligent migration diagnosis method for aircraft engine bearing faults based on multi-source data and attention mechanism, which can automatically learn the weight differences under different modes and different fault states, capture the deep correlation of modalities and suppress noise pollution to enhance the robustness of the model.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0008] An intelligent migration diagnosis method for aircraft engine bearing faults based on multi-source data and an attention mechanism is proposed. The method automatically learns the importance differences of different modalities under different fault conditions through attention weights, and realizes the dynamic optimization combination of modal features. At the same time, the attention mechanism is used to capture the deep correlation characteristics between the multimodal data of aircraft engine bearing components, explore the spatiotemporal coupling relationship between vibration, temperature, and current heterogeneous signals, and use attention weights to automatically suppress the contribution of noise-contaminated modes to enhance noise robustness.

[0009] An intelligent migration diagnosis method for aircraft engine bearing faults based on multi-source data and attention mechanism includes the following steps:

[0010] Step 1: obtaining original detection signals of multi-modal data of different fault states of bearing mechanical equipment;

[0011] Step 2, data preprocessing: randomly intercept n samples of length L from the first half of the source domain working condition data, and mark the fault type as the training set X m Part of D s (x s ,y s ), randomly intercept n samples of length L in the first half of the target domain working condition without any annotation as the other part of the training set, namely D t (x t ), and finally randomly intercept n samples of length L from the second half of the target domain working condition data as the test set T m ;

[0012] Step 3: Fault migration diagnosis model construction: A fault migration diagnosis model based on a deep residual network is constructed. A corresponding modal feature extraction subnetwork is constructed based on the number of modalities. Each feature extraction subnetwork contains two one-dimensional convolutional layers, two residual blocks, two maximum pooling layers, and two batch normalization layers. Each residual block contains two one-dimensional convolutional layers. Multimodal fusion based on the attention mechanism is performed after the feature extraction subnetwork. Finally, the model is input into the classifier through two fully connected layers for fault classification diagnosis.

[0013] Step 4, include D s (x s ,y s ) and D t (x t ) Two-part training set X m The input is sent to the feature extraction subnetwork for feature extraction and training, and the cross entropy loss function is used as the optimization target. The cross entropy loss function is shown in formula (1). Then the gradient descent algorithm is used to update the model parameters θ. The gradient descent algorithm formula is shown in (2).

[0014]

[0015] where y i is the probability of the i-th category of the true label, is the probability of the i-th category of the model prediction label, N is the number of categories of the classification task, L C is the cross entropy loss function;

[0016]

[0017] In the formula, θ represents the parameters of the model (weights and biases), which is usually a vector; α: learning rate, which controls the step size of each parameter update; L C : Cross entropy loss function, used for classification tasks;

[0018] Step 5: Repeat step 4, iterate and update the parameters of the optimized fault migration diagnosis model until the maximum number of iterations is reached, the training is completed, and the final fault migration diagnosis model is obtained;

[0019] Step 6: The labeled sample test set T containing the target domain working conditions is m The input is fed into the final fault migration diagnosis model to obtain the cross entropy loss and diagnostic accuracy on the target domain working condition.

[0020] The step 3 is specifically as follows:

[0021] 3.1) The specific method of the one-dimensional convolutional layer is: Assume that the input data is a one-dimensional sequence x = [x1, x2, x3, ..., x M ], where M is the length of the input data. The convolution operation is to extract the input features locally through a kernel filter w, that is, the convolution kernel. Assume that the length of the convolution kernel w is F L , its operation is expressed as:

[0022]

[0023] in is a local segment of the input sequence, * T is the transpose of the matrix, w is the weight of the filter kernel, b is the bias term, φ is the nonlinear activation function, z i is the result of the convolution operation, which represents the i-th feature extracted by the convolution kernel; the final result of the convolution is z,

[0024]

[0025] Each z i Corresponding to the local features extracted by the convolution kernel each time, multiple convolution kernels can be used at the same time, and different convolution kernel sizes correspond to the extraction of features of different scales;

[0026] 3.2) The specific method of multimodal fusion based on the attention mechanism is as follows: After extracting different modal features, they are first concatenated in the time step dimension. Then, a multi-head attention mechanism is used to perform weighted output on the concatenated modal features in the time step dimension to obtain the fused multimodal fault features. The multi-head attention mechanism is expressed as follows:

[0027]

[0028] Where Q is the query matrix, K is the key matrix, V is the value matrix, d k is the dimension of the key, is a scaling factor, and softmax is a normalization function. The overall process of the multi-head attention mechanism is as follows: first, linearly transform the query matrix Q, key matrix K, and value matrix V:

[0029] Q i =QW i Q ,K i =KW i K ,V i =VW i V (6)

[0030] Where W i Q 、W i K 、W i V is the learnable linear transformation matrix of the i-th attention head, Q i , K i 、V i are the query, key, and value of the i-th attention head after linear transformation;

[0031] Then use scaled dot product attention to calculate the query matrix Q, key matrix K, and value matrix V corresponding to each attention head, as follows:

[0032] head i =Attention(Q i ,K i ,V i ) (7)

[0033] Then the outputs of each attention head are spliced ​​and linearly transformed again to obtain the comprehensive attention output. The process is as follows:

[0034] MultiHead(Q,K,V)=Concat(head1,head2,...,head h )W O (8)

[0035] Where h is the number of attention heads, W O is a learnable weight matrix, Concat is a concatenation operation, d model is the hidden layer dimension.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] The present invention proposes an intelligent migration diagnosis method for aircraft engine bearing faults based on multi-source data and an attention mechanism. The attention mechanism is used to construct a dynamic feature weight distribution multimodal fusion form for the multimodal data of aircraft engine bearing components, and the multi-faceted fusion of multimodal data is efficiently realized. On the one hand, compared with the static modal weights adopted by traditional fixed weight fusion methods (such as weighted averaging or decision voting), the attention mechanism adaptively learns the contribution of each modality at different fault stages, and weights the modal data according to the contribution, thereby improving the core modal performance and weakening the interference of redundant modalities. On the other hand, compared with traditional splicing and fusion methods, the directly spliced ​​multimodal features are easily interfered by low-quality modalities, while the attention mechanism automatically reduces the influence of high-noise modalities through weight attenuation, thereby enhancing the noise robustness of the fused multimodal features. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Flowchart of a method according to an embodiment of the present invention.

[0039] Figure 2 This is a structural diagram of the fault migration diagnosis model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The present invention is described in detail below with reference to the embodiments and accompanying drawings.

[0041] Reference Figure 1 , an intelligent migration diagnosis method for aircraft engine bearing faults based on multi-source data and attention mechanism, including the following steps:

[0042] Step 1: Acquire multi-modal original signals of different fault states of the bearing equipment, such as horizontal vibration signals, vertical vibration signals, current signals, etc.;

[0043] Step 2, data preprocessing: The original multimodal signal is divided into a training set and a test set. The training set is mainly composed of n samples of length L randomly selected from the first half of the original multimodal signal. However, the training set must contain not only source domain data but also target domain data, that is, the training set X m =(D s (x s ,y s ), D t (x t )), this is because the labeled source domain samples D in the training set s (x s ,y s ) is used to enable the model to learn fault diagnosis, while the unlabeled target domain samples D t (x t) is used for domain adaptation of the model and migration to the target domain; the test set is to randomly select n samples of length L from the second half of the original multimodal signal and label the fault status of each sample. This is because the target domain data of the test set needs to be separated from the target domain data used for training in the training set to ensure the effectiveness of migration diagnosis;

[0044] Step 3: Construction of fault migration diagnosis model: Figure 2 As shown in the figure, the fault migration diagnosis model constructs a corresponding feature extraction subnetwork according to the number of modes of multimodal data. Each feature extraction subnetwork contains two one-dimensional convolutional layers, two residual blocks, two maximum pooling layers, and two batch normalization layers. Each residual block contains two one-dimensional convolutional layers. In traditional deep learning neural networks, as the number of network layers continues to deepen, the model will experience gradient vanishing and degradation problems. This problem can be solved by introducing residual blocks. The function of the residual block is to add the input and output to obtain a new output. The residual block learns the residual of the input and output, rather than directly learning the mapping from input to output. The maximum pooling operation is mainly to extract enhanced features. , reducing the dimension of the feature vector, compressing the data to retain important information, and at the same time reducing the number of model parameters and reducing the risk of overfitting; batch normalization is mainly used to solve the problem of internal covariate shift, thereby greatly reducing the dependency and influence relationship between network layers, making the network training speed and model convergence speed significantly faster, alleviating the gradient explosion or gradient disappearance problem to a certain extent, and the training process is more stable; after passing through the feature extraction sub-network, the features need to be spliced ​​first, and after splicing, it is convenient to perform subsequent multi-source data fusion based on the attention mechanism, and finally input into the classifier through two fully connected layers to perform classification diagnosis of the aircraft engine bearing fault status;

[0045] Step 3 is as follows:

[0046] 3.1) The specific method of the one-dimensional convolutional layer is: Assume that the input data is a one-dimensional sequence x = [x1, x2, x3, ..., x M ], where M is the length of the input data. The convolution operation is to extract the input features locally through a kernel filter w, that is, the convolution kernel. Assume that the length of the convolution kernel w is F L , its operation is expressed as:

[0047]

[0048] in is a local segment of the input sequence, * T is the transpose of the matrix, w is the weight of the filter kernel, b is the bias term, φ is the nonlinear activation function, z iThe result of the convolution operation represents the features extracted by the convolution kernel, which represents the i-th feature extracted by the convolution kernel; the final result of the convolution is z:

[0049]

[0050] Each z i Corresponding to the local features extracted by the convolution kernel each time, multiple convolution kernels can be used at the same time, and different convolution kernel sizes correspond to the extraction of features of different scales;

[0051] In step 3, to enhance model stability, residual connections and layer normalization are used:

[0052] Y=LayerNorm(X+F(X)) (11)

[0053] Where F(X) is the common mapping relationship between the two layers of one-dimensional convolutional networks in the residual block, Y is the final output of the residual block, and X is the input feature of the residual connection;

[0054] 3.2) The specific process of the multi-source data fusion method based on the attention mechanism is as follows: Assume that the input is a two-modal time series feature, and the feature matrix of modality A is The characteristic matrix of mode B is Where T is the time step, d a ,,d b For the feature dimension, first perform the splicing operation along the feature dimension:

[0055]

[0056] After splicing the multimodal features, we can perform weighted operations on the multimodal features in the time step. Each time step contains all the modal features of the time step. Therefore, the weight of each time step represents the comprehensive expression ability of all modal features in the period. In addition, the attention mechanism weights the time step dimension, so it is necessary to process the multi-source data into a unified time step form to facilitate the attention weighted operation.

[0057] In order to fuse each modal feature, it is also necessary to consider all modal features in each time step. Therefore, a multi-head attention mechanism is adopted. The multi-head attention mechanism can dynamically consider the different dimensions of multi-modal splicing features, deeply explore the related features of different modalities, and enhance the multi-modal linkage expression ability. The multi-head attention mechanism divides the feature dimensions according to the number of heads N. h Divide equally, that is, the feature dimension must be the number of heads N h an integer multiple of ;

[0058] The multimodal fusion method used: The core of the multi-head attention mechanism is the scaled dot product attention mechanism. The scaled dot product attention mechanism has the characteristics of high computational efficiency and can greatly improve the efficiency of matrix operations. Therefore, it is suitable for more complex multimodal fusion operations. Its calculation formula is as follows:

[0059]

[0060] Where Q is the query matrix, K is the key matrix, V is the value matrix, d k is the dimension of the key, is a scaling factor, which is used to prevent the gradient from disappearing due to the explosion of the dot product value. Softmax is a normalization function. The above is only the expression of the single-head attention mechanism. For the multi-head attention mechanism, the feature dimension needs to be divided equally according to the number of heads. Each attention head pays attention to a different dimension. This is because when fusing multi-source data, it is necessary to pay attention to different modal data and some key data in each modal data. Therefore, it is necessary to use multiple attention heads to perform weighted fusion of the multi-dimensional multi-source data.

[0061] The overall process of the multi-head attention mechanism is as follows: first, linearly transform the query matrix Q, key matrix K, and value matrix V: For the i-th attention head, linearly project the query matrix Q, key matrix K, and value matrix V respectively:

[0062]

[0063] Where W i Q 、W i K 、W i V is the learnable linear transformation matrix of the i-th attention head, Q i , K i 、V i are the query, key, and value of the i-th attention head after linear transformation. Since the original features of different modal data are usually located in different vector spaces (different dimensions and distributions), the features of each modality are mapped to a unified semantic space through independent linear transformations to make them comparable; different attention heads can capture different dependencies in the sequence (such as short-distance dependency, long-distance dependency, etc.), and then use scaled dot product attention to calculate the query matrix Q, key matrix K, and value matrix V corresponding to each attention head. That is, each attention head calculates the attention weights between all time steps on its own head dimension, which is calculated as follows:

[0064] head i =Attention(Q i ,K i ,V i ) (15)

[0065] Then the outputs of h attention heads are concatenated along the feature dimension and linearly fused:

[0066] MultiHead(Q,K,V)=Concat(Head1,...,Head h )W O (16)

[0067] Where h is the number of attention heads, W O is a learnable weight matrix, Concat is a concatenation operation, d model is the hidden layer dimension; the final comprehensive attention output focuses on different parts of the multi-source data information sequence, captures deeper causal dependencies, and enables the model to learn richer multimodal correlation features;

[0068] In the actual operation of the multi-head attention mechanism, a total of 4 projection matrix operations are performed, including 3 matrix projections corresponding to the query matrix Q, key matrix K, and value matrix V respectively. Finally, the splicing features of all attention heads are subjected to 1 matrix projection; the fourth projection, on the one hand, fuses the complementary information captured by different heads into a unified representation to ensure that the output dimension is consistent with the input (to facilitate residual connection and subsequent layer processing); on the other hand, it enhances modal fusion. In multimodal tasks, this projection can further coordinate the contribution of cross-modal information (such as suppressing redundant heads and strengthening key modal features).

[0069] Step 4: In order to reduce the domain offset between the source domain and the target domain, it is necessary to include D s (x s ,y s ) and D t (x t ) Two-part training set X m Input into the feature extraction subnetwork for feature extraction and training. After the source domain and target domain are input into the network respectively, the maximum mean difference between the source domain and the target domain of each layer of the feature extraction subnetwork and the fully connected layer can be obtained. The cross entropy loss function shown in formula (17) and the maximum mean difference loss shown in formula (18) are used as optimization targets. Through these two optimization targets, the model can not only learn fault diagnosis knowledge, but also achieve migration targets by unsupervised alignment to the target domain. Subsequently, the gradient descent algorithm is used to update the model parameters θ; the gradient descent algorithm formula is shown in formula (19) below;

[0070]

[0071] where y i is the probability of the i-th category of the true label, is the probability of the i-th category of the model prediction label, N is the number of categories of the classification task, L C is the cross entropy loss function;

[0072]

[0073] In the formula, θ represents the parameters of the model (weights and biases), which is usually a vector; α: learning rate, which controls the step size of each parameter update; L C : Cross entropy loss function, used for classification tasks;

[0074] Step 5: Repeat step 4, iteratively updating and optimizing the parameters of the fault migration diagnosis model so that the model can accurately diagnose fault conditions and reuse the learned knowledge in the target domain. This process continues until the maximum number of iterations is reached, completing the training and obtaining the final fault migration diagnosis model.

[0075] Step 6: The labeled sample test set T containing the target domain working conditions is m The input is fed into the final fault migration diagnosis model to obtain the cross entropy loss and diagnostic accuracy on the target domain working condition.

[0076] This embodiment uses the fault diagnosis of a bearing fault dataset of the University of Potsdam in Germany and a machine pump cycle fault experimental dataset of Xi'an Jiaotong University as examples to verify the effectiveness of the present invention.

[0077] The University of Potsdam bearing failure dataset in Germany contains two modal data types: motor current and bearing vibration modes. We selected five artificially damaged bearing conditions, five actual bearing damage conditions caused by accelerated bearing testing, and one healthy bearing condition. The five artificially damaged bearing conditions are KA01 (level 1 EDM groove outer ring damage), KA03 (level 2 electric engraving outer ring damage), KA05 (level 2 electric engraving outer ring damage), KA06 (level 2 electric engraving outer ring damage), and KA07 (level 1 bearing outer ring drilling damage). The five actual bearing damage conditions caused by accelerated bearing testing are KA04 (outer ring fatigue and pitting, single-point distribution), KA15 (outer ring and cage plastic deformation and indentation, single-point distribution), KA16 (outer ring and cage fatigue and pitting, single-point distribution), KA22 (outer ring fatigue and pitting, single-point distribution), and KA30 (outer ring plastic deformation, single-point distribution).

[0078] 100 samples were collected for each fault type, for a total of 11 × 100 = 1100 fault samples for the 11 fault types. This dataset contains four operating conditions: num0, speed 1500 rpm, load torque 0.7 Nm, radial force 1000 N, numbered N15_M07_F10; num1, speed 900 rpm, load torque 0.7 Nm, radial force 1000 N, numbered N09_M07_F10; num2, speed 1500 rpm, load torque 0.1 Nm, radial force 1000 N, numbered N15_M01_F10; and num3, speed 1500 rpm, load torque 0.7 Nm, radial force 400 N, numbered N15_M07_F04. Each operating condition has 1100 fault samples, for a total of 4 × 1100 = 4400 samples for the four operating conditions. 100 samples with a length of 1500 are randomly selected from the first half of a measurement data as a training set, and 100 samples with a length of 1500 are randomly selected from the second half as a test set.

[0079] This experiment uses four modal data in the Xi'an Jiaotong University pump cycle fault experimental data set, including bearing radial horizontal vibration X, axial vibration Y, radial vertical vibration Z, and pump end vertical vibration signal P. In this data set, the bearing fault locations include inner ring fault, outer ring fault, rolling element fault, internal and external composite fault, inner roller composite fault, outer roller composite fault, and inner and outer roller composite fault. The severity of the bearing is divided into mild, moderate, and severe. There are 13 types of bearing faults, including 9 single position faults and 4 composite position faults. This data set has a total of 4 test conditions. The conditions are shown in Table 1 below.

[0080] Table 1. Four working conditions

[0081]

[0082] Experimental parameter settings: Bearing fault dataset from the University of Potsdam, Germany: batch size B s =1024, number of cycles N e =9000, learning rate L r =0.001.

[0083] Xi'an Jiaotong University pump cycle failure experimental dataset: batch size B s =256, number of cycles N e =3000, learning rate L r =0.0001.

[0084] Below, we compare the present invention's multimodal data fusion transfer method, based on an attention mechanism, with a single-modal data transfer method without fusion, on two datasets to demonstrate its effectiveness. The task is set as follows: for example, T12 represents the source domain as 1 and the target domain as 2, with the task performing a transfer diagnosis from source domain 1 to target domain 2. The comparison results are shown in Table 2.

[0085] Table 2 shows that fault transfer diagnosis using a multimodal data fusion method based on an attention mechanism significantly outperforms single-modal methods without multimodal data fusion in all transfer tasks. In the German Potsdam University bearing fault dataset, the diagnostic accuracy of the multimodal data fusion method improved by 49.45%, 54.09%, 25.18%, 52.55%, 22.54%, and 13.73% compared to the single-modal method without fusion in tasks T12, T13, T14, T23, T24, and T34, respectively. In the Xi'an Jiaotong University pump cycle fault experimental dataset, the diagnostic accuracy of the multimodal data fusion method improved by 34.61%, 59.04%, 61.88%, 42.46%, 39.46%, and 12.96% compared to the single-modal method without fusion in tasks T12, T13, T14, T23, T24, and T34, respectively. It can be seen that in both data sets, the intelligent migration diagnosis method for aircraft engine bearing faults based on multi-source data and attention mechanism performed excellently. It can better capture the multimodal correlation characteristics of bearing faults, combine multi-angle information to comprehensively evaluate the fault type and accurately classify it, and realize the dynamic optimization combination of modal features, overcoming the limitations of traditional fixed weight fusion. At the same time, it uses the attention mechanism to capture the deep correlation characteristics between multimodal data, improve the stability of the model in the extremely complex environment of aircraft engines, and enhance noise robustness, which is a huge improvement compared to the single-modal method without fusion.

[0086] Table 2. Comparative experimental results of multimodal fusion and single modality without fusion on two datasets

[0087]

Claims

1. An intelligent migration diagnosis method for aircraft engine bearing faults based on multi-source data and attention mechanism, characterized by: By automatically learning the importance differences of different modes under different fault conditions through attention weights, the dynamic optimization combination of modal features is achieved; at the same time, the attention mechanism is used to capture the deep correlation characteristics between multimodal data of aircraft engine bearing components, explore the spatiotemporal coupling relationship between vibration, temperature, and current heterogeneous signals, and use attention weights to automatically suppress the contribution of noise-contaminated modes to enhance noise robustness.

2. The intelligent migration diagnosis method for aircraft engine bearing faults based on multi-source data and attention mechanism according to claim 1 is characterized in that: The following steps are involved: Step 1: obtaining original detection signals of multi-modal data of different fault states of bearing mechanical equipment; Step 2, data preprocessing: randomly intercept n samples of length L from the first half of the source domain working condition data, and mark the fault type as the training set X m Part of D s (x s ,y s ), randomly intercept n samples of length L in the first half of the target domain working condition without any annotation as the other part of the training set, namely D t (x t ), and finally randomly intercept n samples of length L from the second half of the target domain working condition data as the test set T m ; Step 3: Fault migration diagnosis model construction: A fault migration diagnosis model based on a deep residual network is constructed. A corresponding modal feature extraction subnetwork is constructed based on the number of modalities. Each feature extraction subnetwork contains two one-dimensional convolutional layers, two residual blocks, two maximum pooling layers, and two batch normalization layers. Each residual block contains two one-dimensional convolutional layers. Multimodal fusion based on the attention mechanism is performed after the feature extraction subnetwork. Finally, the model is input into the classifier through two fully connected layers for fault classification diagnosis. Step 4, include D s (x s ,y s ) and D t (x t ) Two-part training set X m The input is sent to the feature extraction subnetwork for feature extraction and training, and the cross entropy loss function is used as the optimization target. The cross entropy loss function is shown in formula (1). Then the gradient descent algorithm is used to update the model parameters θ. The gradient descent algorithm formula is shown in (2). where y i is the probability of the i-th category of the true label, is the probability of the i-th category of the model prediction label; N is the number of categories of the classification task, L C is the cross entropy loss function; In the formula, θ represents the parameters of the model, namely weights and biases, which is a vector; α: learning rate, which controls the step size of each parameter update; L C For classification tasks; Step 5: Repeat step 4, iterate and update the parameters of the optimized fault migration diagnosis model until the maximum number of iterations is reached, the training is completed, and the final fault migration diagnosis model is obtained; Step 6: The labeled sample test set T containing the target domain working conditions is m The input is fed into the final fault migration diagnosis model to obtain the cross entropy loss and diagnostic accuracy on the target domain working condition.

3. The method for intelligent migration diagnosis of aircraft engine bearing faults based on multi-source data and attention mechanism according to claim 2, wherein step 3 specifically comprises: 3.1) The specific method of the one-dimensional convolutional layer is: Assume that the input data is a one-dimensional sequence x = [x1, x2, x3, ..., x N ], where N is the length of the input data. The convolution operation is to extract the input features locally through a kernel filter w, that is, the convolution kernel. Assume that the length of the convolution kernel w is F L , its operation is expressed as: in is a local segment of the input sequence, * T is the transpose of the matrix, w is the weight of the filter kernel, b is the bias term, φ is the nonlinear activation function, z i The result of the convolution operation represents the feature extracted by the convolution kernel, which represents the i-th feature extracted by the convolution kernel; the final result of the convolution is z, Each z i Corresponding to the local features extracted by the convolution kernel each time, multiple convolution kernels can be used at the same time, and different convolution kernel sizes correspond to the extraction of features of different scales; 3.2) The specific method of multimodal fusion based on the attention mechanism is as follows: After extracting different modal features, they are first concatenated in the time step dimension. Then, a multi-head attention mechanism is used to perform weighted output on the concatenated modal features in the time step dimension to obtain the fused multimodal fault features. The multi-head attention mechanism is expressed as follows: Where Q is the query matrix, K is the key matrix, V is the value matrix, d k is the dimension of the key, is a scaling factor, and softmax is a normalization function. The overall process of the multi-head attention mechanism is as follows: first, linearly transform the query matrix Q, key matrix K, and value matrix V: Q i =QW i Q ,K i =KW i K ,V i =VW i V (6) Where W i Q 、W i K 、W i V is the learnable linear transformation matrix of the i-th attention head, Q i , K i 、V i are the query, key, and value of the i-th attention head after linear transformation; Then use scaled dot product attention to calculate the query matrix Q, key matrix K, and value matrix V corresponding to each attention head, as follows: head i =Attention(Q i ,K i ,V i ) (7) Then the outputs of each attention head are spliced ​​and linearly transformed again to obtain the comprehensive attention output. The process is as follows: MultiHead(Q,K,V)=Concat(head1,head2,...,head h )W O (8) Where h is the number of attention heads, W O is a learnable weight matrix, Concat is a concatenation operation, d model is the hidden layer dimension.

Images