Cross-domain aircraft rotor wing water fetching fault diagnosis method

Abstract

The invention provides a cross-domain aircraft rotor water fetching fault diagnosis method, and relates to the technical field of fault diagnosis, and the method comprises the steps: constructing a multi-source heterogeneous data set; preprocessing and standardizing the data; converting the one-dimensional sensor signal into a two-dimensional image representation with three-channel characteristics; inputting the two-dimensional image representation into a residual network, and then sequentially carrying out expansion convolution and multi-scale feature extraction, attention mechanism and weighted fusion and bottom-up feature fusion to obtain feature representation of fault data; constructing a few-sample double-graph cooperative propagation network, and training the few-sample double-graph cooperative propagation network in combination with a dynamic early stop method based on graph collaboration improvement; and identifying and diagnosing the fault type or state. According to the technical scheme, the problem that in the prior art, fault detection and diagnosis cannot be accurately carried out on the cross-domain aircraft in real time under the condition of lack of labeled data is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and particularly to a method for diagnosing the fault of rotor water hitting of a cross-domain vehicle. Background Art

[0002] A cross-domain vehicle can operate in two media, underwater and air, and is widely used in mission scenarios such as ocean exploration, underwater rescue, and environmental monitoring. During the process of medium conversion, especially when emerging from water, the rotor inevitably interacts with the water surface, resulting in the phenomenon of "rotor water hitting". This phenomenon seriously affects the performance and stability of the vehicle. The problem of rotor water hitting has become a key technical bottleneck for cross-domain vehicles to perform tasks across media.

[0003] Mechanically, due to the effects of water hammer force and cavitation, the rotor blades are extremely prone to wear, deformation, and even fracture. In severe cases, it may lead to the interruption of thrust output, affecting the normal operation of the vehicle. At the level of the power system, the high resistance and thrust discontinuity at the underwater-air interface may cause the motor load to be too high, resulting in phenomena such as overheating or system protective shutdown, and even water ingress into the system due to insufficient sealing performance, endangering the safety of the entire power module. In addition, the instability of attitude control is also an issue that cannot be ignored. During the process of the rotor emerging from water, due to the uneven distribution of thrust and the interference of fluid reaction forces, the vehicle is prone to pitch, yaw, or roll instability, leading to out-of-control during the emerging process. The problem of fluid resonance further amplifies this hazard. When the working frequency of the rotor approaches the natural frequency of the vehicle structure, the resonance phenomenon will cause severe vibrations, affecting the structural strength and service life of the equipment, and even leading to system failures.

[0004] Due to the above-mentioned various fault risks hidden in the process of rotor water hitting, it is particularly necessary to establish an efficient fault diagnosis method and system. On the one hand, the fault diagnosis system can realize the real-time monitoring of the states of the rotor and the power system. By collecting multi-faceted signals such as rotor vibration, thrust, and attitude, it can timely identify fault problems such as mechanical damage of the rotor and motor overload, and avoid the further expansion of faults. On the other hand, the fault diagnosis system can provide feedback and optimization for the attitude control during the vehicle's emerging process. By early detecting abnormal phenomena such as thrust imbalance or turbulent interference, it can provide data support for the improvement of control algorithms, thereby enhancing the stability and safety of the vehicle during the cross-domain conversion process. In addition, for the fatigue damage of the rotor structure and the fluid resonance phenomenon, the fault diagnosis system can predict the trend of fault occurrence in advance, extend the service life of the vehicle equipment, and reduce the maintenance cost.

[0005] In summary, the problem of the rotor splashing water in the cross-domain vehicle not only affects the stability of mission execution, but may also lead to systematic failures and safety hazards. Therefore, developing a method for diagnosing the rotor splashing water fault of the cross-domain vehicle is the key way to improve the performance and reliability of the cross-domain vehicle. The establishment of this system will provide real-time and accurate fault detection and diagnosis functions for the cross-domain vehicle, further ensuring its stable operation in complex mission environments and laying a solid foundation for the development and application of future cross-domain vehicle technologies.

[0006] Therefore, there is a need for a method for diagnosing the rotor splashing water fault of the cross-domain vehicle that can perform real-time and accurate fault detection and diagnosis on the cross-domain vehicle in the absence of labeled data. Summary of the Invention

[0007] The main object of the present invention is to provide a method for diagnosing the rotor splashing water fault of the cross-domain vehicle to solve the problem that in the prior art, it is impossible to perform real-time and accurate fault detection and diagnosis on the cross-domain vehicle in the absence of labeled data.

[0008] To achieve the above object, the present invention provides a method for diagnosing the rotor splashing water fault of the cross-domain vehicle, which specifically includes the following steps:

[0009] S1. Obtain sensor data and construct a multi-source heterogeneous data set.

[0010] S2. Preprocess and standardize the data in the multi-source heterogeneous data set.

[0011] S3. Perform multi-order difference analysis on the data processed in step S2, and construct a statistical pattern based on the improved Markov method to convert the one-dimensional sensor signal into a two-dimensional image representation with three-channel characteristics.

[0012] S4. First input the two-dimensional image representation into a residual network, and then perform dilated convolution and multi-scale feature extraction, attention mechanism and weighted fusion, and bottom-up feature fusion in sequence to obtain the feature representation of the fault data.

[0013] S5. Use the representation after feature extraction to construct a few-shot dual-graph collaborative propagation network and train the few-shot dual-graph collaborative propagation network.

[0014] S6. After passing the data to be detected through steps S1 to S4, input it into the trained few-shot dual-graph collaborative propagation network to identify and diagnose the fault type or state.

[0015] Further, step S3 specifically includes the following steps:

[0016] S3.1. Assume that the original signal after step S2 is x = {x 1 ,x2 , …, x N}, the first-order difference is Δ 1 x t′ , and the second-order difference is Δ 2 x t′ , then:

[0017] Δ 1 x t′ = x t′ - x t′-1 , t = 2, …, N(1);

[0018] Δ 2 x t′ = Δ 1 x t′ - Δ 1 x t′-1 , t = 3, …, N(2);

[0019] where, x t′ represents the t'-th data in the original signal, and N represents a positive integer.

[0020] S3.2. Divide the original signal x into M quantile bins according to the numerical range, and each quantile bin corresponds to a state, where m = 1, 2, …, M. The quantile bin division is completed by formula (3):

[0021] S m = [Q m-1 , Q m ),

[0022] where, S m represents the range of the m-th quantile bin, Q m is the m-th quantile point, represents the quantile point at which of the data in x is less than or equal to this value.

[0023] Map each value x t′ in the signal to the corresponding number s t :

[0024] s t = m, x t′ ∈ S m (4). S3.3. The state transition matrix P represents the probability P i of transitioning from state S j to state S ij :

[0025]

[0026] where, n ij represents state Si The number of transitions to state S j , where J is the total number of states.

[0027] Construct the state transition matrix P into I, where the pixel value I ij represents the state transition probability:

[0028] I ih = P ih , i, j = 1, 2, …, J (6);

[0029] where I is a single-channel Markov transition field.

[0030] S3.4. For the original signal x, the first-order difference Δ 1 x t′ , and the second-order difference Δ 2 x t′ perform the Markov method respectively to obtain three single-channel transition field images, and merge the three single-channel transition field images into a three-channel image I 3C :

[0031] I 3C = Concat(I (0) , I (1) , I (2) ) (7);

[0032] where I (0) is the transition field of the original signal, I (1) is the transition field of the first-order difference, I (2) is the transition field of the second-order difference, and Concat represents the channel merging operation.

[0033] S3.5. Scale the pixel value range of the image, scale the pixel values of the image to [0, 1], and perform normalization processing according to the mean and standard deviation.

[0034] Furthermore, step S4 specifically includes the following steps:

[0035] S4.1. Input the three-channel two-dimensional image Extract multi-scale features through ResNet50 to obtain the feature map F f , where f = 2, 3, 4, 5.

[0036] S4.2. For each feature map F f , where f = 2, 3, 4, 5, through the dilated convolution operation with different dilation rates k, obtain three feature maps with different scales, and then splice the feature maps along the channel dimension to form a multi-scale fusion feature map.

[0037] In S4.3, after performing global average pooling on the multi-scale fusion feature map and generating an attention weight vector through a fully connected layer, weighted fusion is performed by channels to generate an attention-enhanced feature map.

[0038] In S4.4, the deepest attention-enhanced feature map is used as the top-level feature map, and the shallow features are successively upsampled and fused to form the final multi-scale fusion feature set, and finally converted into a feature vector through a global average pooling operation.

[0039] Further, step S5 specifically includes the following steps:

[0040] In S5.1, the fault feature data is divided according to the ratio of 80% training set and 20% test set. The support set S and the query set Q are extracted from the training set, where the support set contains labeled samples and unlabeled samples; for each labeled sample, the label of the sample is represented by one-hot encoding and concatenated with to generate the node feature vector V node :

[0041]

[0042] where C and d represent the vector dimensions.

[0043] For unlabeled samples, the label vector is initialized as a zero vector The node feature V' of the unlabeled sample node is represented as:

[0044]

[0045] where Concat represents the concatenation operation.

[0046] In S5.2, the node feature matrix in the instance graph is composed of the node features of the support set and the query set. Among them, N ins is the total number of samples in the support set and the query set; the node features of the labeled samples in the support set are composed of ; the node features of the unlabeled samples in the support set are composed of and the label vector is set as a zero vector; the node features of the query set samples are composed of ; all node features are combined into the instance graph node feature matrix

[0047] In S5.3, the K-means clustering algorithm is used to divide into N dis distribution clusters, and the center point C k of the distribution cluster is determined. The center point is concatenated with the initialized label vector as the node feature matrix in the distribution graph It represents the distribution information of the support set and query set samples in the feature space.

[0048] S5.4, Instance graph edge weight matrix It is initialized through the feature similarity between node features, specifically expressed as:

[0049]

[0050] Among them, and respectively represent the features of the i-th and j-th nodes in the instance graph, and σ′ is the scale parameter controlling the feature similarity.

[0051] Distribution graph edge weight matrix It is initialized through the feature similarity between the distribution center points:

[0052]

[0053] Among them, and respectively represent the features of the k-th and l-th nodes in the distribution graph.

[0054] Furthermore, step S5 also includes:

[0055] S5.5, Input instance graph node features and instance graph edge weights The enhancement of instance graph node features is achieved through the dual-graph collaborative Transformer method; the process of enhancing instance graph node features is the same as that of enhancing distribution graph node features. The specific steps for enhancing instance graph node features are as follows:

[0056] S5.5.1, Map the instance graph node features to the instance graph query vector Q ins , key vector K ins and value vector V ins , and the calculation formula is:

[0057]

[0058] Among them, is the mapping weight matrix, h is the number of heads of multi-head attention, and d a is the dimension of a single head.

[0059] S5.5.2, Calculate the attention weights between nodes using the instance graph mask matrix and the instance graph edge weight matrix.

[0060] S5.5.3, Multiply the instance graph attention weight α ins with the instance graph value vector Vins Multiply them to obtain the instance graph attention output X att,ins , and map it back to the original feature dimension through a linear transformation:

[0061]

[0062] where, is the weight matrix of the output mapping.

[0063] S5.5.4, perform a residual connection between the instance graph attention output X att,ins and the instance graph node feature , and through a layer normalization operation, obtain the instance graph node feature X res1,ins enhanced in the first step:

[0064]

[0065] where, LayerNorm is the layer normalization operation.

[0066] S5.5.5, input X res1,ins into a two-layer feed-forward network FFN, and perform a residual connection and layer normalization again to finally obtain the instance graph enhanced feature

[0067]

[0068] Furthermore, step S5.5.2 specifically includes the following steps:

[0069] First, calculate the original attention score through the dot product between the query vector Q ins,i of instance graph node i and the key vector K ins,j of instance graph node j:

[0070]

[0071] where, α raw,ij is the attention score between instance graph node i and node j, is the scaling factor, is K ins,j transpose;

[0072] Subsequently, combine the edge weight matrix of instance graph node i and node j and the instance graph mask matrix M ins,ij , mask and adjust the attention score to obtain the final attention weight α ins,ij :

[0073]

[0074] Among them, Softmax is the activation function, and the instance graph mask matrix M ins,ij is defined as:

[0075]

[0076] Furthermore, step S5 further includes: S5.6, performing cross-graph interaction using the collaborative attention mechanism.

[0077] Specifically, it includes the following steps:

[0078] S5.6.1, mapping the instance graph enhanced feature to the cross-graph interaction attention query vector Q ins2dis , mapping the distribution graph enhanced feature to the cross-graph interaction attention key vector K dis2ins and the cross-graph interaction attention value vector V dis2ins respectively. The specific calculation is as follows:

[0079]

[0080] Among them, is the cross-graph interaction attention linear mapping matrix, h is the number of heads of the multi-head attention, and d a is the feature dimension of a single head.

[0081] S5.6.2, calculating the cross-graph attention weight α ins2dis through Q dis2ins and K ins←dis :

[0082]

[0083] Among them, α ins←dis,ij represents the cross-graph attention weight between the i-th node in the instance graph and the j-th node in the distribution graph, is the scaling factor, Q ins2dis,i is the query vector of the i-th node of the instance graph enhanced feature , K dis2ins,j is the key vector of the j-th node of the distribution graph node feature , is the transpose of K dis2ins,j .

[0084] S5.6.3, multiplying the cross-graph attention weight by the value vector V dis2ins of the distribution graph to obtain the cross-graph fusion feature V ins←dis of the instance graph node:

[0085]

[0086] Among them, It represents the representation of the $i$-th node in the instance graph after the fusion of the distribution map features.

[0087] S5.6.4, perform a residual connection and layer normalization on the cross-graph fusion feature $V$ of the instance graph nodes ins←dis and the instance graph enhancement feature to obtain the cross-graph interaction feature of the instance graph

[0088]

[0089] Furthermore, step S5 also includes: S5.7, fuse the graph node features.

[0090] Specifically, it includes the following steps:

[0091] S5.7.1, concatenate the instance graph enhancement feature and the cross-graph interaction feature of the instance graph to obtain the intermediate feature representation $Z$ ins :

[0092]

[0093] where Concat(·) means concatenating and along the feature dimension to obtain $Z$ ins .

[0094] S5.7.2, perform weighted fusion on the enhancement feature and the cross-graph fusion feature through the gating coefficient $g$ ins to obtain the final node feature of the instance graph

[0095]

[0096] where, is the weight matrix of the fusion gate; is the bias vector; σ(·) represents the Sigmoid activation function to ensure that the fusion coefficient is between [0,1].

[0097]

[0098] where ⊙ represents element-wise multiplication.

[0099] Furthermore, step S5 also includes: S5.8, update the graph edge weights;

[0100] Specifically, it includes the following steps:

[0101] S5.8.1, calculate the square of the feature difference between node $i$ and node $j$:

[0102]

[0103] Among them, represents the squared feature difference between the i-th node and the j-th node.

[0104] S5.8.2, taking as the input, through two-layer MLP mapping to obtain the edge weight score h ins,ij , and the calculation formula is:

[0105]

[0106] Among them, W e1 and W e2 are the weight matrices of the MLP; b e1 and b e2 are the bias vectors; ReLU(·) represents the activation function.

[0107] S5.8.3, through exponential mapping of the edge weight score h ins,ij to ensure that the edge weight is positive, specifically expressed as:

[0108]

[0109] Among them, represents the edge weight after the l-th update.

[0110] Furthermore, step S5 also includes: S5.9, training the few-shot dual-graph collaborative propagation network with dual-graph collaborative iteration based on improved early stopping;

[0111] Specifically, it includes the following steps:

[0112] S5.9.1, for the labeled node i in the support set, the prediction distribution is calculated through the Gaussian kernel function between nodes and normalized by Softmax, and the specific expression is as follows:

[0113]

[0114] Among them, t represents the number of the current iteration, is the predicted label distribution of node i; is the edge weight between nodes i and j in the instance graph; φ(y j ) is the one-hot encoding representation of the true label of node j; ζ s (·) is the Softmax function.

[0115] S5.9.2, the loss of the instance graph is obtained through the Euclidean distance between the prediction distribution and the true label:

[0116]

[0117] Among them, y i is the true label of node i, and S label represents the set of all labeled nodes in the support set.

[0118] S5.9.3, Combining the losses of the instance graph and the distribution graph, adding a regularization term to constrain the model parameters W and the edge weights E, the total loss function L (t+1) is defined as:

[0119]

[0120] Among them, λ ins and ξ dis are the weight parameters of the losses of the instance graph and the distribution graph respectively, and λ reg is the weight coefficient of the regularization term, is the loss of the distribution graph; is the instance graph edge weight matrix at the (t + 1)-th iteration; is the distribution graph edge weight matrix at the (t + 1)-th iteration.

[0121] S5.9.4, The change rate ΔL of the loss function (t+1) is defined as:

[0122]

[0123] The synergy distance measures the difference in node features between the instance graph and the distribution graph is defined as:

[0124]

[0125] Among them, is the instance graph key vector at the (t + 1)-th iteration, represents the feature representation obtained after iteration of the distribution center in the same cluster as the instance node i.

[0126] S5.9.5, Within the set observation window Ω, if any of the following conditions is satisfied, early stopping is triggered:

[0127] ΔL (t+1) <∈ loss or

[0128] Among them, ∈ loss is the loss change rate threshold, and ∈ coh is the synergy distance threshold.

[0129] After the dynamic early stopping is triggered, using the final features V of node i and node j final,i and V final,iPerform label propagation and fault classification; the label of the unlabeled node i is obtained by weighting the Gaussian similarity of the labeled nodes in the support set:

[0130]

[0131] where I(y j = c) indicates whether the true label y j of node j belongs to class c; The Gaussian kernel function measures the feature similarity between nodes; is the predicted label of node i, and arg max is a function that finds the parameter in the function arguments that makes the function value maximum.

[0132] The present invention has the following beneficial effects:

[0133] (1) By introducing a few-shot learning framework of dual-graph co-evolution, the present invention realizes efficient feature propagation between the support set and the query set, effectively addresses cross-domain and data scarcity problems, and improves the generalization ability of the diagnostic model under new domain and few-shot conditions.

[0134] (2) Compared with traditional diagnostic methods based on a single sensor signal or simple features, the present invention performs unified preprocessing and differential-Markov mapping on multi-source sensor data, and converts the time series signal into a multi-channel two-dimensional feature representation image. This process maximally mines the time series characteristics and statistical distribution information in the data, provides more discriminative input features for the subsequent deep network, and thus effectively improves the accuracy and robustness of fault recognition.

[0135] (3) In the feature extraction stage, the present invention uses a multi-level feature pyramid network integrating a residual structure and an attention mechanism for feature extraction, greatly improving the flexibility and depth of the model in extracting key features, being able to better capture the subtle changes of time series signals, and suppressing irrelevant or redundant features, thereby improving the accuracy and efficiency of fault discrimination.

[0136] (4) Traditional graph update methods often rely on a large amount of experimental data to determine fixed iteration times in practical applications, which easily leads to model overfitting, a decline in generalization ability, or insufficient graph structure update, resulting in the diagnostic performance being difficult to reach the optimal. The present invention designs a dynamic early stopping strategy based on improved graph synergy. This strategy adaptively determines the timing of iteration termination by taking into account the change trend of the loss function and the synergy of the dual-graph features. When the performance improvement of the graph co-evolution process is no longer significant, the system automatically terminates the iteration, ensuring that the training process converges to the optimal state, reducing the waste of computing resources, further improving the training efficiency and the stability of the diagnostic model, and showing significant advantages under complex cross-domain and few-shot conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0137] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings. In the drawings:

[0138] Figure 1 The flowchart of a method for diagnosing the rotor water - hitting fault of a cross - domain vehicle is shown.

[0139] Figure 2 The flowchart of applying a method for diagnosing the rotor water - hitting fault of a cross - domain vehicle to a cross - domain vehicle rotor water - hitting fault diagnosis system is shown. Specific embodiments

[0140] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the drawings. Obviously, the described embodiments are some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0141] As Figure 1 shown, a method for diagnosing the rotor water - hitting fault of a cross - domain vehicle specifically includes the following steps:

[0142] S1. Obtain sensor data and construct a multi - source heterogeneous data set.

[0143] S2. Pre - process and standardize the data in the multi - source heterogeneous data set.

[0144] S3. Use the improved multi - order difference Markov mapping method for feature representation conversion. Perform multi - order difference analysis on the data processed in step S2, and construct a statistical pattern based on the improved Markov method to convert the one - dimensional sensor signal into a two - dimensional image representation with three - channel characteristics.

[0145] S4. First input the two - dimensional image representation into a residual network, and then perform dilated convolution and multi - scale feature extraction, attention mechanism and weighted fusion, and bottom - up feature fusion in sequence to obtain the feature representation of the fault data.

[0146] S5. Using the representation after feature extraction, construct a few-shot dual-graph collaborative propagation network and train the few-shot dual-graph collaborative propagation network. Using the representation after feature extraction, realize fault identification and classification by constructing a few-shot dual-graph collaborative propagation network. Divide the feature representation of the data into a training set and a test set, with the training set set to 80% and the test set set to 20%. Input the data of the training set into this network, construct a support set and a query set for the labeled data and unlabeled data, build a dual-graph structure of an instance graph and a distribution graph, and perform graph propagation based on the dual-graph co-evolution mechanism to achieve decision generalization for cross-domain fault diagnosis under few-shot conditions. Train the model parameters and iteratively optimize the graph structure through a dynamic early stopping strategy based on graph collaboration improvement to make the training process converge and ensure better classification performance.

[0147] S6. After the data to be detected goes through steps S1 - S4, input it into the trained few-shot dual-graph collaborative propagation network to identify and diagnose the fault type or status. After freezing the model parameters, input the data to be detected into the trained few-shot dual-graph collaborative propagation network after going through the same preprocessing and feature extraction process to identify and diagnose the fault type or status.

[0148] Specifically, the sensor data in step S1 includes: key signal data obtained from an inertial measurement unit, an acceleration sensor, a strain gauge sensor, a current-voltage sensor, a rotational speed sensor, and a thrust sensor. By deploying multiple types of sensors such as an inertial measurement unit, an acceleration sensor, a strain gauge sensor, a current-voltage sensor, a rotational speed sensor, and a thrust sensor, key signal data related to attitude, acceleration, structural strain, power status, rotor speed, and thrust during the operation of the vehicle are obtained. Uniformly collect the data collected by different sensors to construct a multi-source heterogeneous data set.

[0149] Specifically, step S2 specifically includes the following steps:

[0150] S2.1. Clean the data.

[0151] S2.2. Use median filtering to smooth the cleaned data. Based on the sliding window size k, sort the data within the window and take the median to replace the center value of the window.

[0152] S2.3. Normalize and standardize the data, linearly map the data to the interval [0, 1]

[0153] Convert the normalized data to a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0154] S2.4. To address the issues of inconsistent sampling frequencies and asynchronous time bases among multi-source sensors, a timing alignment method is adopted to unify the time scale of the data. Based on the unified time base, the sensor data is resampled.

[0155] Specifically, step S2.1 includes the following steps:

[0156] S2.1.1. Calculate the median of each sensor data as the central tendency of the data.

[0157] S2.1.2. Calculate the absolute deviation between each data point and the median, and find the median of the absolute deviations, i.e., the MAD value.

[0158] S2.1.3. Set the threshold to 2.5 times the MAD. If the absolute deviation of a certain data point from the median exceeds the threshold, then this data point is determined to be an outlier, and the outlier is removed or replaced.

[0159] Specifically, step S3 includes the following steps:

[0160] S3.1. Assume that the original signal after step S2 is x = {x 1 , x 2 , …, x N}. Calculate the first-order difference Δ 1 x t′ and the second-order difference Δ 2 x t′ ;

[0161] S3.2. Divide the original signal x into M quantile bins according to the numerical range. Each quantile bin corresponds to a state, where m = 1, 2, …, M. The quantile bin division is completed through formula (1):

[0162] S m = [Q m-1 , Q m ),

[0163] where S m represents the range of the m-th quantile bin, Q m is the m-th quantile point, represents the quantile point where of the data in x is less than or equal to this value.

[0164] Map each value x t′ in the signal to the corresponding number s t according to the quantile bin:

[0165] s t = m, x t′ ∈ S m (2). S3.3, the state transition matrix P represents the probability P i of transitioning from state S j to state S ij :

[0166]

[0167] where n ij represents the number of transitions from state S i to state S j , and J is the total number of states.

[0168] Construct the state transition matrix P into I, where the pixel value I ij represents the state transition probability:

[0169] I ij = P ij , i, j = 1, 2, …, J (4);

[0170] where I is a single-channel Markov transition field.

[0171] S3.4, perform the Markov method on the original signal x, the first-order difference Δ 1 x t′ , and the second-order difference Δ 2 x t′ respectively, to obtain three single-channel transition field images, and merge the three single-channel transition field images into a three-channel image I 3C .

[0172] S3.5, scale the pixel value range of the image to [0, 1] and perform normalization processing according to the mean and standard deviation.

[0173] Specifically, step S4 specifically includes the following steps:

[0174] S4.1, input a three-channel two-dimensional image to extract multi-scale features through ResNet50. The network first passes through the initial convolutional layer and the max-pooling operation, and then successively passes through 4 stages of bottleneck residual blocks to extract features of different scales:

[0175] Perform a 7×7 convolution with a stride s = 2 and a padding value p = 3 and a 3×3 max-pooling with a stride s = 2 and a padding value p = 1 to obtain the initial feature map F 1 ; F 1 passes through 3 bottleneck residual blocks to output the first-stage feature map F 2 ; F 2 passes through 4 bottleneck residual blocks to output the second-stage feature map F 3 ; F 3 passes through 6 bottleneck residual blocks to output the third-stage feature map F4 ; Stage 4: F 4 Pass through 3 bottleneck residual blocks and output the feature map F of the fourth stage 5 .

[0176] Among them, F 2 , F 3 , F 4 , F 5 correspond to the features of different scales respectively, with the resolution decreasing gradually and the number of channels being 256, 512, 1024, and 2048 in sequence.

[0177] S4.2, For each feature map F f , where f = 2, 3, 4, 5, through dilated convolution operations with different dilation rates k, obtain three feature maps T of different scales f,1 , T f,2 , T f,3 , and then splice the feature maps along the channel dimension to form a multi-scale fusion feature map.

[0178] Splice the feature maps T f,1 , T f,2 , T f,3 obtained by the three dilated convolutions respectively along the channel dimension to form a multi-scale fusion feature map U i .

[0179] Extract the global statistical information of the channels through the global average pooling GAP operation, and the formula is expressed as:

[0180]

[0181] Among them, represents the global average pooling result of the c-th channel in the feature map U i ; H and W are the height and width of the feature map U i respectively; h and w represent the row and column indices in the feature map, represents the feature value of the c-th channel at the position (h, w);

[0182] S4.3, Perform global average pooling on the multi-scale fusion feature map, and after generating the attention weight vector through the fully connected layer, fuse with weights by channels to generate the attention enhanced feature map.

[0183] Specifically:

[0184] Input the vector p i after global average pooling into a fully connected layer to learn the dependencies between channels and generate the attention weight vector a i , and the calculation formula is:

[0185] a i = σ(Wfc ·p i +b fc ) (6);

[0186] Wherein, W fc and b fc respectively represent the weight matrix and bias vector of the fully connected layer; p i is the input vector of global average pooling; σ represents the Sigmoid activation function;

[0187] Multiply the attention weight vector a i channel by channel with the multi-scale fusion feature map U i to combine the weighted features of all channels together to obtain the attention-enhanced multi-scale feature map A i .

[0188] S4.4. Use the deepest attention-enhanced feature map as the top-level feature map, upsample and fuse the shallow features in sequence to form the final multi-scale fusion feature set, and finally convert it into a feature vector through global average pooling operation.

[0189] Specifically: The feature fusion starts from the deepest attention-enhanced feature map A 5 , and use A 5 as the top-level output feature map F top , and the formula is expressed as:

[0190] F top =A 5 (7);

[0191] Wherein, F top is the top-level output feature map.

[0192] In order to transfer the high-level semantic information to the low level, upsample and feature fusion are carried out step by step. Starting from A 5 , at each level of A a , where a = 4, 3, 2, it is fused with the feature map of the previous layer through the upsample operation; the nearest neighbor interpolation method is used for the upsample operation, and the representation of the target feature map after upsample is U up ; for the feature map A a of the a-th layer, the calculation formula for feature fusion is:

[0193] F a =Conv2D(A a +Upsample(F a+1 , s = 2)) (8);

[0194] Wherein, F a represents the fused feature map of the a-th layer; A a is the attention-enhanced feature map of the a-th layer; Upsample(Fa+1 , s = 2) indicates that the feature fusion map F of the previous layer is a+1 Through the upsampling operation, with parameter s = 2, that is, each dimension of the input feature map is enlarged by 2 times, so that the resolution of the output feature map is 2 times that of the input, and finally enlarged to the same size as A a ; + represents the element-wise addition operation; Conv2D represents the 3×3 convolution operation.

[0195] Through the bottom-up path, successively fuse A 2 , A 3 , A 4 , A 5 , and finally obtain the multi-scale fusion feature map F 2 , F 3 , F 4 , F top , where F 2 represents the fusion feature of the lowest layer, while F top is the top layer feature. The final set of multi-scale fusion feature maps is represented as:

[0196]

[0197] Among them, is the representation of the final multi-scale fusion feature set.

[0198] Convert the final multi-scale fusion feature into a vector through the global average pooling operation

[0199] Specifically, step S5 specifically includes the following steps:

[0200] S5.1, divide the fault feature data according to the ratio of 80% training set and 20% test set, extract the support set S and query set Q from the training set, where the support set contains labeled samples and unlabeled samples; for each labeled sample, represent the label of the sample through one-hot encoding and splice it with to generate the node feature vector V of the labeled sample node :

[0201]

[0202] Among them, C and d represent the vector dimensions;

[0203] For unlabeled samples, the label vector is initialized as a zero vector The node feature V' of the unlabeled sample node is represented as:

[0204]

[0205] Among them, Concat represents the concatenation operation;

[0206] S5.2, the node feature matrix in the instance graph is formed by concatenating the node features of the support set and the query set. Among them, N ins is the total number of samples in the support set and the query set; the node features of the labeled samples in the support set are composed of ; the node features of the unlabeled samples in the support set are composed of ; the label vector is set to the zero vector; the node features of the query set samples are composed of ; all node features are combined into the instance graph node feature matrix

[0207] S5.3, the node feature matrix in the distribution graph represents the distribution information of the support set and query set samples in the feature space. The nodes in the distribution graph do not directly correspond to the samples, but are the distribution centers obtained by clustering the sample features. Use the K-means clustering algorithm to be divided into N dis distribution clusters, and determine the distribution cluster center C k . Concatenate the center point with the initialized label vector as the node feature matrix in the distribution graph represents the distribution information of the support set and query set samples in the feature space;

[0208] S5.4, the instance graph edge weight matrix is initialized through the feature similarity between node features, specifically expressed as:

[0209]

[0210] Among them, and respectively represent the features of the i-th and j-th nodes in the instance graph, σ′ is the scale parameter controlling the feature similarity; exp is the natural exponential function;

[0211] The distribution graph edge weight matrix is initialized through the feature similarity between the distribution centers:

[0212]

[0213] Among them, and respectively represent the features of the k-th and l-th nodes in the distribution graph.

[0214] Specifically, the node calculation process of the distribution graph in step S5.3 is as follows:

[0215] S5.3.1, Use the sample features in the support set and the query set as input, and use the K-means clustering algorithm to divide the samples into N dis distribution clusters, and the center point C of each cluster k is defined as:

[0216]

[0217] where S k is the set of samples belonging to the k-th cluster, and |S k | represents the number of samples in the cluster;

[0218] S5.3.2, Use the center point C of each cluster k as the distribution map node feature, and initialize the label vector as a zero vector Concatenate to obtain the distribution map node feature

[0219]

[0220] S5.3.3, After concatenating all the distribution center points, form a distribution map node feature matrix where N dis is the number of nodes in the distribution map, that is, the total number of distribution clusters.

[0221] Specifically, step S5 also includes:

[0222] Step S5 also includes:

[0223] S5.5, Input the instance graph node feature and the instance graph edge weight Enhance the instance graph node feature through the dual-graph collaborative Transformer method; the process of enhancing the instance graph node feature is the same as that of enhancing the distribution map node feature. The specific steps for enhancing the instance graph node feature are as follows:

[0224] S5.5.1, Map the instance graph node feature to the instance graph query vector Q ins , key vector K ins and value vector V ins respectively. The calculation formula is:

[0225]

[0226] where is the mapping weight matrix, h is the number of heads of multi-head attention, and d a is the dimension of a single head;

[0227] S5.5.2, calculate the attention weights between nodes using the instance graph mask matrix and the instance graph edge weight matrix;

[0228] S5.5.3, multiply the instance graph attention weight α ins by the instance graph value vector V ins to obtain the instance graph attention output X att,ins , and map it back to the original feature dimension through a linear transformation:

[0229]

[0230] where, is the weight matrix of the output mapping;

[0231] S5.5.4, perform a residual connection between the instance graph attention output X att,ins and the instance graph node feature , and through layer normalization operation, obtain the instance graph node feature X res1,ins enhanced in the first step:

[0232]

[0233] where, LayerNorm is the layer normalization operation;

[0234] S5.5.5, input X res1,ins into a two-layer feed-forward network FFN, and perform residual connection and layer normalization again to finally obtain the instance graph enhanced feature

[0235]

[0236] The enhancement process of the distribution map is similar to that of the instance graph. Input the distribution map node feature and the distribution map edge weight to enhance the distribution map node feature through the dual-graph collaborative Transformer method.

[0237] Specifically, step S5.5.2 specifically includes the following steps:

[0238] First, calculate the original attention score through the dot product between the query vector Q ins,i of instance graph node i and the key vector K ins,j of instance graph node j:

[0239]

[0240] where, α raw,ij is the attention score between instance graph node i and node j, is the scaling factor used to stabilize the attention distribution, is K ins,j is the transpose of;

[0241] Subsequently, combine the edge weight matrix of instance graph nodes i and j and the instance graph mask matrix M ins,ij , mask and adjust the attention scores to obtain the final attention weight α ins,ij :

[0242]

[0243] where Softmax is the activation function, and the instance graph mask matrix M ins,ij is defined as:

[0244]

[0245] This mask ensures that when the edge weight between nodes i and j is zero, the attention score is masked to negative infinity, so that the Softmax output is zero.

[0246] Specifically, step S5 further includes: S5.6, perform cross-graph interaction using the collaborative attention mechanism; perform cross-graph interaction between the instance graph and the distribution graph using the collaborative attention mechanism; the cross-graph interaction process between the instance graph and the distribution graph is the same.

[0247] The cross-graph interaction of the instance graph specifically includes the following steps:

[0248] S5.6.1, map the instance graph enhanced feature to the cross-graph interaction attention query vector Q ins2dis , and map the distribution graph enhanced feature to the cross-graph interaction attention key vector K dis2ins and the cross-graph interaction attention value vector V dis2ins respectively, and the specific calculation is:

[0249]

[0250] where is the cross-graph interaction attention linear mapping matrix, h is the number of heads of multi-head attention, and d a is the feature dimension of a single head;

[0251] S5.6.2, calculate the cross-graph attention weight α ins2dis and K dis2ins through Q ins←dis :

[0252]

[0253] where α ins←dis,ijDenotes the cross-graph attention weight between the $i$-th node in the instance graph and the $j$-th node in the distribution graph. is the scaling factor, $Q$ ins2dis,i is the enhanced feature of the instance graph for the query vector of the $i$-th node, $K$ dis2ins,j is the enhanced feature of the distribution graph for the key vector of the $j$-th node, is for $K$ dis2ins,j transpose;

[0254] S5.6.3, Multiply the cross-graph attention weight with the value vector $V$ of the distribution graph dis2ins to obtain the cross-graph fusion feature $V$ of the instance graph node ins←dis :

[0255]

[0256] where, Denotes the representation of the $i$-th node in the instance graph after fusing with the distribution graph features.

[0257] S5.6.4, Perform residual ins←dis connection and layer normalization on the cross-graph fusion feature $V$ of the instance graph node and the enhanced feature of the instance graph to obtain the cross-graph interaction feature of the instance graph

[0258]

[0259]

[0260] Similarly, for the distribution graph, the node features are fused through the key vector $K$ of the instance graph ins2dis and the value vector of the instance graph

[0261] $V$ ins2dis The calculation process is the same as above. Finally, the cross-graph interaction feature of the distribution graph is:

[0262]

[0263] where, is the enhanced feature of the distribution graph.

[0264] Specifically, step S5 also includes: S5.7, Fuse the features of the instance graph and distribution graph nodes; the process of fusing the features of the instance graph and distribution graph nodes is the same;

[0265] The feature fusion of the instance graph nodes specifically includes the following steps:

[0266] S5.7.1, Combine the enhanced feature of the instance graph and the cross-graph interaction feature of the instance graph Perform splicing to obtain the intermediate feature representation Z of the instance graph ins :

[0267]

[0268] Among them, Concat(·) means to and Splice along the feature dimension to obtain Z ins ;

[0269] S5.7.2, through the instance graph gating coefficient g ins Perform weighted fusion on the enhanced feature of the instance graph and the cross-graph fusion feature of the instance graph to obtain the final node feature of the instance graph

[0270]

[0271] Among them, is the weight matrix of the instance graph for fusion gating; is the bias vector; σ(·) represents the Sigmoid activation function to ensure that the fusion coefficient is between [0,1].

[0272]

[0273] Among them, ⊙ represents element-wise multiplication.

[0274] For the node features of the distribution graph, weighted fusion is also performed through the fusion gating mechanism, and the specific calculation steps are the same as those of the instance Figure 1 Finally, the final node feature of the distribution graph is obtained

[0275] Specifically, step S5 also includes: S5.8, update the edge weights of the instance graph and the edge weights of the distribution graph; the processes of updating the edge weights of the instance graph and the weights of the distribution graph are the same;

[0276] The specific steps for updating the edge weights of the instance graph are as follows:

[0277] S5.8.1, calculate the square of the feature difference of the nodes:

[0278]

[0279] Among them, represents the square of the feature difference between the i-th node and the j-th node of the instance graph;

[0280] S5.8.2, take as the input, and obtain the edge weight score h of the instance graph through two-layer MLP mapping ins, ij, and the calculation formula is:

[0281]

[0282] Among them, W e1 and W e2 are the weight matrices of the MLP; b e1 and b e2 are the bias vectors; ReLU(·) represents the activation function to ensure non-linear mapping;

[0283] S5.8.3, the edge weight score h of the instance graph ins,ij is passed through an exponential mapping to ensure that the edge weights of the instance graph are positive, specifically expressed as:

[0284]

[0285] Among them, represents the edge weight of the instance graph after the l-th update.

[0286] For the edge weight update of the distribution graph, the operation is exactly the same as that of the instance graph.

[0287] Specifically, step S5 further includes: S5.9, training the few-shot dual-graph collaborative propagation network with dual-graph collaborative iteration based on improved dynamic early stopping;

[0288] Specifically, it includes the following steps:

[0289] S5.9.1, the nodes of the instance graph perform label propagation through the similarity of edge weights with other nodes to obtain the predicted distribution. For the labeled node i in the support set, the predicted distribution is calculated through the Gaussian kernel function between nodes and normalized by Softmax, specifically expressed as follows:

[0290]

[0291] Among them, t represents the number of the current iteration, is the predicted label distribution of node i; is the edge weight between nodes i and j in the instance graph at the (t + 1)-th iteration, obtained through node features; φ(y j ) is the one-hot encoding representation of the true label of node j; ζ s (·) is the Softmax function to normalize the label distribution.

[0292] S5.9.2, the loss of the instance graph is obtained through the Euclidean distance between the predicted distribution and the true label:

[0293]

[0294] Among them, y i is the true label of node i, S labelRepresents the set that supports all the marked nodes in the set.

[0295] The design method of the loss function of the distribution map is exactly the same as that of the instance map loss function.

[0296] S5.9.3, Combining the losses of the instance map and the distribution map, adding a regularization term to constrain the model parameters W and the edge weights E, the total loss function L (t+1) Is defined as:

[0297]

[0298] λ ins And ξ dis Are the weight parameters of the instance map and distribution map losses respectively, λ reg Is the weight coefficient of the regularization term, Is the loss of the distribution map; Is the instance map edge weight matrix at the (t + 1)-th iteration; Is the distribution map edge weight matrix at the (t + 1)-th iteration.

[0299] To prevent overfitting and improve the graph update efficiency, a dynamic early stopping mechanism based on graph synergy monitors the change rate of the loss function and the synergy distance between the two graphs during training.

[0300] S5.9.4, The change rate of the loss function ΔL (t+1) Is defined as:

[0301]

[0302] The synergy distance measures the difference in node features between the instance map and the distribution map Is defined as:

[0303]

[0304] Where, Is the instance map key vector at the (t + 1)-th iteration, Represents the feature representation obtained after iteration of the distribution center in the same cluster as the instance node i.

[0305] S5.9.5, Within the set observation window Ω, if any of the following conditions is satisfied, early stopping is triggered:

[0306] ΔL (t+1) <∈ loss Or

[0307] Where, ∈ loss Is the loss change rate threshold, ∈ coh Is the synergy distance threshold.

[0308] After the dynamic early stopping is triggered, the final features V of nodes i and j are used final,i and V final,j to perform label propagation and fault classification; the label of the unlabeled node i is obtained by weighting the Gaussian similarity with the labeled nodes in the support set:

[0309]

[0310] where, I(y j = c) indicates whether the true label y j of node j belongs to class c; The Gaussian kernel function measures the feature similarity between nodes; is the predicted label of node i, and arg max is a function that finds the parameter in the function arguments that makes the function value the largest.

[0311] Specifically, in step S6, after freezing the model parameters, the data to be detected is input into the trained few-shot dual-graph collaborative propagation network after the same preprocessing and feature extraction process to identify and diagnose the fault type or status.

[0312] Step S6 also includes the following steps:

[0313] S6.1, Fix all the parameters in the trained few-shot dual-graph collaborative propagation network to ensure that the network is no longer updated during the inference stage.

[0314] S6.2, Preprocess the input data to be detected according to the same process as in the training stage, including: data cleaning and standardization: removing invalid data and unifying the data format; normalization and alignment: normalizing the data collected by different sensors and performing temporal alignment; feature transformation: converting the temporal signal into a three-channel two-dimensional image representation through multi-order difference and Markov mapping methods.

[0315] S6.3, Input the preprocessed two-dimensional image into the feature extraction module combined with the trained residual network and multi-scale attention mechanism to generate multi-level fusion features.

[0316] S6.4, Use the extracted feature representation as node features and input them into the frozen few-shot dual-graph collaborative propagation network, and combine the feature propagation mechanisms of the instance graph and the distribution graph to identify the fault type or status.

[0317] S6.5, In the few-shot dual-graph collaborative propagation network, calculate the fault labels of nodes through the label propagation and Gaussian kernel similarity mechanism, and output the final diagnosis result of the fault type.

[0318] S6.6. Deploy the frozen model to the cross-domain vehicle system to provide real-time fault diagnosis function. Integrate it into the computing device according to the actual application requirements for on-line detection and diagnosis.

[0319] In the design of the cross-domain vehicle, in order to comprehensively monitor various faults of the rotor water hitting and provide stable operation guarantee, a variety of sensors are reasonably deployed. The strain gauge sensor is installed at the root of the rotor blade, mainly used to capture the instantaneous impact force and stress change when the rotor starts, detect whether there is deformation or fatigue damage of the blade, so as to ensure the mechanical health state of the rotor. Adjacent to the rotor support structure, a thrust sensor is arranged. This sensor monitors the output of the rotor thrust in real time, judges whether the thrust is uniform, and quickly identifies problems such as insufficient thrust or abnormal fluctuation. The associated rotational speed sensor is installed at the shaft end of the rotor motor, responsible for monitoring the rotational speed change when the rotor starts and operates, and timely detecting faults such as unstable rotational speed or abnormal output of the power system.

[0320] At the same time, current and voltage sensors are deployed at the power supply module and the motor input end of the power system to monitor the current and voltage fluctuations during the operation of the motor and diagnose whether there are phenomena such as current overload or power abnormality in the power system. In order to capture the overall vibration of the vehicle, acceleration sensors are deployed in the front and rear areas of the fuselage. This deployment can effectively identify abnormal vibration and impact signals caused by rotor startup, fluid disturbance or structural faults. In addition, in order to monitor the attitude change of the vehicle during the water exit process, an inertial detection unit is installed at the center of gravity of the fuselage to provide attitude data such as pitch angle, roll angle and yaw angle in real time, assisting the attitude control system to adjust the attitude and ensuring that the vehicle can smoothly complete the transition from underwater to air.

[0321] Through the precise deployment of these sensors, the vehicle can comprehensively collect key information such as the mechanical state of the rotor, the output of the power system, attitude change and structural vibration, thus providing strong data support for the fault diagnosis of the rotor water hitting.

[0322] Such as Figure 2As shown in the figure, the present invention first uses a sensor module to collect data. The sensor module includes: an inertial detection unit, a strain gauge sensor, a current and voltage sensor, a thrust sensor, a rotational speed sensor, and an acceleration sensor. Subsequently, the collected data is subjected to data cleaning, denoising, normalization, and standardization through a data processing module, and data alignment and data synchronization are performed. Then, feature extraction and fault diagnosis are carried out through a calculation processing module, and then the fault diagnosis result is sent to a human-machine interaction module. The human-machine interaction module includes: a monitoring screen, a diagnostic parameter adjustment panel, and an alarm system. The human-machine interaction module is used for generating a fault report. Finally, the generated fault report is sent to an execution control module. According to the generated fault report, an attitude controller, a power controller, and a safety protection unit in the execution control module control the cross-domain vehicle.

[0323] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions, or replacements made by those skilled in the art within the scope of the essence of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for diagnosing rotor water-pumping faults of a cross-domain aircraft, characterized in that: The specific steps include: S1, obtain sensor data and build multi-source heterogeneous data sets; S2, preprocessing and standardizing the data in multi-source heterogeneous datasets; S3, performing multi-order differential analysis on the data processed in step S2, and constructing a statistical pattern based on an improved Markov method to convert the one-dimensional sensor signal into a two-dimensional image representation with three-channel characteristics; S4, the two-dimensional image representation is first input into the residual network, followed by dilated convolution and multi-scale feature extraction, attention mechanism and weighted fusion, and bottom-up feature fusion to obtain the feature representation of the fault data; S5, using the representation after feature extraction, construct a few-shot dual-graph collaborative propagation network, and train the few-shot dual-graph collaborative propagation network in combination with the dynamic early stopping method based on graph synergy improvement; S6, after the data to be detected has passed through steps S1 to S4, it is input into the trained few-sample dual-graph collaborative propagation network to identify and diagnose the fault type or state.

2. A cross-domain aircraft rotor water-flying fault diagnosis method according to claim 1, characterized in that: Step S3 specifically includes the following steps: S3.1, assuming that the original signal after step S2 is x = {x1, x2, ..., x N }, the first-order difference is Δ 1 x t′ , the second-order difference is Δ 2 x t′ ,but: Δ 1 x t′ =x t′ -x t′-1 , t=2,…,N (1); D 2 x t′ =D 1 x t′ -D 1 x t′-1 , t=3,…,N (2); Among them, x t′ represents the t′th data in the original signal, and N represents a positive integer; S3.2, divide the original signal x into M bins according to the value range, each bin corresponds to a state, where m = 1, 2, ..., M, and the bin division is completed by formula (3): Among them, S m represents the range of the mth quantile bin, Q m is the mth quantile, It means that there is The data is less than or equal to the quantile of this value; Bin each value x in the signal according to the quantile t′ Mapped to the corresponding number s t : s t =m, x t′ ∈S m (4); S3.3, the state transition matrix P represents the transition from state S i Transfer to state S j The probability P ij : Among them, n ij Indicates state S i To state S j The number of transitions, J is the total number of states; The state transition matrix P is constructed as I, where the pixel value I ij Represents the state transition probability: I ij =P ij , i,j=1,2,…,J (6); Where I is a single-channel Markov transfer field; S3.4, for the original signal x, the first-order difference Δ 1 x t′ , second-order difference Δ 2 x t′ Execute the Markov method separately to obtain three single-channel transfer field images, and merge the three single-channel transfer field images into a three-channel image I 3C : I 3C =Concat(I (0) ,I (1) ,I (2) 9 (7); Among them, I (0) is the transfer field of the original signal, I (1) is the first-order difference transfer field, I (2) is the transfer field of the second-order difference, Concat represents the channel merging operation; S3.5, scaling the pixel value range of the image and normalizing it according to the mean and standard deviation.

3. The method for diagnosing rotor water-pumping faults of a cross-domain aircraft according to claim 1, characterized in that: Step S4 specifically includes the following steps: S4.1, input three-channel two-dimensional image After ResNet50 extracts multi-scale features, we get the feature map F f , where f = 2, 3, 4, 5; S4.2, for each feature map F f , where f = 2, 3, 4, 5, through the expansion convolution operation with different expansion rates k, three feature maps of different scales are obtained, and then the feature maps are spliced ​​along the channel dimension to form a multi-scale fusion feature map; S4.3, perform global average pooling on the multi-scale fusion feature map, generate an attention weight vector through a fully connected layer, and then perform channel-wise weighted fusion to generate an attention-enhanced feature map; S4.4, the deepest attention-enhanced feature map is used as the top-level feature map, and the shallow features are upsampled and fused in turn to form the final multi-scale fused feature set, which is finally converted into a feature vector through a global average pooling operation.

4. The method for diagnosing rotor water-pumping faults of a cross-domain aircraft according to claim 1, characterized in that: Step S5 specifically includes the following steps: S5.1, divide the fault feature data into 80% training set and 20% test set, extract the support set S and query set Q from the training set, where the support set contains labeled samples and unlabeled samples; for each labeled sample, encode the label of the sample by one-hot encoding Indicates that Splice and generate the node feature vector V of the labeled sample node : Among them, C and d represent the vector dimensions; For unlabeled samples, the label vector is initialized to a zero vector. Node feature V′ of unlabeled samples node It is expressed as: Among them, Concat represents the concatenation operation; S5.2, Node feature matrix in instance graph It is composed of the node features of the support set and the query set, where N ins is the total number of samples in the support set and query set; the node features of the labeled samples in the support set are given by The node features of the unlabeled samples in the support set are composed of The label vector is set to zero vector; the node features of the query set samples are composed of Composition; all node features are combined into the instance graph node feature matrix S5.3, using K-means clustering algorithm Divide into N dis distribution clusters and determine the center point C of the distribution cluster k , concatenate the center point with the initialization label vector as the node feature matrix in the distribution graph Represents the distribution information of support set and query set samples in feature space; S5.4, Instance Graph Edge Weight Matrix Initialized by feature similarity between node features, specifically expressed as: in, and They represent the features of the i-th and j-th nodes in the instance graph, σ′ is the scale parameter that controls the feature similarity; exp is the natural exponential function; Distribution graph edge weight matrix Initialize by feature similarity between distribution centers: in, and They represent the characteristics of the kth and lth nodes in the distribution graph respectively.

5. A cross-domain aircraft rotor water-pumping fault diagnosis method according to claim 4, characterized in that: Step S5 also includes: S5.5, Input instance graph node features and instance graph edge weights The instance graph node features are enhanced through the dual-graph collaborative Transformer method. The process of enhancing the instance graph node features is consistent with that of enhancing the distribution graph node features. The specific steps of enhancing the instance graph node features are as follows: S5.5.1, the instance graph node features Mapped to instance graph query vector Q respectively ins , key vector K ins Sum value vector V ins , the calculation formula is: in, is the mapping weight matrix, h is the number of heads of multi-head attention, and d a is the dimension of a single head; S5.5.2, use the instance graph mask matrix and the instance graph edge weight matrix to calculate the attention weights between nodes; S5.5.3, the instance graph attention weight α ins and the instance graph value vector V ins Multiply them together to get the instance graph attention output X att,ins , and mapped back to the original feature dimension through a linear transformation: X att,ins =Concat(α ins V ins )W Oins (13); in, is the weight matrix of the output mapping; S5.5.4, output the instance graph attention X att,ins Instance graph node features Perform residual connection and layer normalization to obtain the instance graph node feature X after the first step of enhancement. res1,ins : Among them, LayerNorm is the layer normalization operation; S5.5.5, Input X res1,ins To the two-layer feed-forward network FFN, and then perform residual connection and layer normalization again, and finally obtain the instance map enhanced feature 6. A cross-domain aircraft rotor water-flying fault diagnosis method according to claim 5, characterized in that: Step S5.5.2 specifically includes the following steps: First, the query vector Q of the instance graph node i is ins,i and the key vector K of instance graph node j ins,j The dot product between them calculates the raw attention score: Among them, α raw,ij is the attention score between node i and node j, is the scaling factor, It's K ins,j The transpose of Then, combined with the edge weight matrix between node i and node j in the instance graph and the instance image mask matrix M ins,ij , mask and adjust the attention score to get the final attention weight α ins,ij : Among them, Softmax is the activation function, and the instance image mask matrix M ins,ij Defined as:

7. The method for diagnosing rotor water-pumping faults of a cross-domain aircraft according to claim 5, characterized in that: Step S5 further includes: S5.6, using a collaborative attention mechanism to perform cross-graph interaction between the instance graph and the distribution graph; the cross-graph interaction process between the instance graph and the distribution graph is consistent; The cross-graph interaction of instance graphs specifically includes the following steps: S5.6.1, Enhance the features of the instance graph Mapped to the cross-graph interactive attention query vector Q ins2dis , the distribution map is enhanced Mapped to the cross-graph interaction attention key vector K respectively dis2ins and the cross-graph interaction attention value vector V dis2ins , the specific calculation is: in, is the linear mapping matrix of the cross-graph interactive attention, h is the number of heads of multi-head attention, and d a is the characteristic dimension of a single head; S5.6.2, by Q ins2dis and K dis2ins Calculate the cross-graph attention weight α ins←dis : Among them, α ins←dis,ij represents the cross-graph attention weight between the i-th node in the instance graph and the j-th node in the distribution graph, is the scaling factor, Q ins2dis,i Enhance features for instance graphs The query vector of the i-th node, K dis2ins,j Enhance features for distribution maps The key vector of the jth node, K dis2ins,j The transpose of S5.6.3, combine the cross-graph attention weights with the value vector V of the distribution graph dis2ins Multiply them together to get the cross-graph fusion feature V of the instance graph node ins←dis : in, It represents the representation of the i-th node in the instance graph after fusion of distribution graph features; S5.6.4, cross-graph fusion features V of instance graph nodes ins←dis Enhanced features with instance graph Perform residual connection and layer normalization to obtain instance graph cross-graph interaction features 8. A cross-domain aircraft rotor water-flying fault diagnosis method according to claim 7, characterized in that: Step S5 also includes: S5.7, fusing the features of the instance graph and the distribution graph nodes; the process of fusing the features of the instance graph and the distribution graph nodes is the same; The instance graph node feature fusion specifically includes the following steps: S5.7.1, Enhance the features of the instance graph Cross-graph interaction features with instance graphs Splice and get the intermediate feature representation Z of the instance graph ins : Concat(·) means and Splice along the feature dimension to get Z ins ; S5.7.2, gate coefficient g by example graph ins The instance graph enhancement features and the instance graph cross-graph fusion features are weighted fused to obtain the final node features of the instance graph. in, is the instance graph weight matrix of fusion gate; is the bias vector; σ(·) represents the Sigmoid activation function; where ⊙ represents element-wise multiplication.

9. A cross-domain aircraft rotor water-flying fault diagnosis method according to claim 8, characterized in that: Step S5 also includes: S5.8, updating the instance graph edge weights and the distribution graph edge weights; the instance graph edge weight update process and the distribution graph weight update process are consistent; The instance graph edge weight update specifically includes the following steps: S5.8.1, calculate the square of the feature difference of the node: in, Represents the square of the feature difference between the i-th node and the j-th node in the instance graph; S5.8.2, As input, the instance graph edge weight score h is obtained through two layers of MLP mapping. ins,ij , the calculation formula is: Among them, W e1 and W e2 is the weight matrix of MLP; b e1 and b e2 is the bias vector; ReLU(·) represents the activation function to ensure nonlinear mapping; S5.8.3, the instance graph edge weight score h ins,ij Through exponential mapping, the edge weights of the instance graph are ensured to be positive, which can be expressed as: in, Represents the edge weight of the instance graph after the lth round of update.

10. A cross-domain aircraft rotor water-pumping fault diagnosis method according to claim 9, characterized in that: Step S5 further includes: S5.9, training the few-sample dual-graph collaborative propagation network based on dual-graph collaborative iteration with improved dynamic early stopping; The specific steps include: S5.9.1, for node i that supports centralized labeling, the predicted distribution is calculated by the Gaussian kernel function between nodes and normalized by Softmax, which is expressed as follows: Among them, t represents the number of current iterations, is the predicted label distribution of node i; is the edge weight between nodes i and j in the instance graph; φ(y j ) is the one-hot encoding representation of the true label of node j; ζ s (·) is the Softmax function; S5.9.

2. Loss of Instance Graph The Euclidean distance between the predicted distribution and the true label is obtained: Among them, y i is the true label of node i, S label represents the set of all marked nodes in the support set; S5.9.3, comprehensive instance graph and distribution graph losses, adding regularization terms to constrain model parameters W and edge weights E, the total loss function L (t+1) Defined as: Among them, λ ins and dis are the weight parameters of instance graph and distribution graph loss, respectively, reg is the weight coefficient of the regularization term, is the loss of the distribution map; is the instance graph edge weight matrix at the t+1 iteration; is the edge weight matrix of the distribution graph at the t+1 iteration; S5.9.4, rate of change of loss function ΔL (t+1) Defined as: The cooperativity distance measures the difference between node features in the instance graph and the distribution graph. Defined as: in, is the instance graph key vector at iteration t+1, It represents the feature representation of the distribution center of the same cluster as the instance node i after iteration; S5.9.5: Within the set observation window Ω, if any of the following conditions is met, early stopping is triggered: Among them, ∈ loss is the loss change rate threshold, ∈ coh is the synergy distance threshold; After the dynamic early stop is triggered, the final features V of nodes i and j are used final,i and V final,i Perform label propagation and fault classification; the label of unlabeled node i is weighted by the Gaussian similarity with the labeled nodes in the support set: Among them, I(y j =c) represents the true label y of node j j Whether it belongs to category c; The Gaussian kernel function measures the feature similarity between nodes; is the predicted label of node i, and arg max is a function that finds the parameter that maximizes the function value.

Images