Cross-domain fault diagnosis method for UAV clusters based on multi-source domain deep migration network

Abstract

The present invention discloses a cross-domain fault diagnosis method for drone clusters based on a multi-source domain deep migration network. First, in terms of drone cluster fault feature extraction, a spatiotemporal Bayesian graph wavelet convolution transformer is proposed in combination with the drone cluster communication topology to extract spatiotemporal coupled fault features from existing flight status monitoring data under different operating conditions. The existing different operating conditions are regarded as different source domains, and the new operating conditions are regarded as the target domain. Then, a multi-source domain cross-convolution transformer is constructed to learn the relationship between the fault features of different source domains in a segment-by-segment manner to obtain multi-source domain fusion features. Finally, a weight-sharing three-branch cross-domain alignment convolution transformer is proposed to align the multi-source domain fusion features with the target domain fault features in a segment-by-segment manner, extract domain-invariant features, and utilize the source domain fault features to improve the fault diagnosis performance under new operating conditions. The present invention solves the fault diagnosis problem of drone clusters based on limited flight data in the target domain, and improves the accuracy and generalization of cross-domain fault diagnosis.

Description

Technical Field

[0001] The present invention belongs to the field of UAV cluster fault diagnosis, and specifically relates to a UAV cluster cross-domain fault diagnosis method based on a multi-source domain deep migration network. Background Art

[0002] Compared to single drones, drone swarm systems, which operate collaboratively across multiple drones, offer significant advantages in terms of mission completion time, fault tolerance, and operational range. This has led to their widespread application in real-world scenarios such as small-scale precision agriculture, coordinated transportation, and commercial performances. However, due to the interconnectedness of drones within a swarm, failures in one or more drones can not only affect their own flight but also degrade the performance of neighboring drones, negatively impacting the entire swarm system and even causing the risk of system collapse. Therefore, research on swarm fault diagnosis technology is crucial to promptly identify swarm system failures, facilitate timely maintenance decisions, and ensure reliable swarm flight operations.

[0003] With the continuous advancement of industrial data transmission and computing technologies, a large number of flight status monitoring signals from drone swarm systems are being collected, leading to the widespread development of data-driven fault diagnosis methods. Data-driven fault diagnosis models can learn potential fault features from a large number of flight status monitoring signals and establish a nonlinear mapping relationship between features and fault classification labels, thereby achieving fault classification prediction. Data-driven fault diagnosis methods generally include two types: shallow machine learning methods and deep learning methods. Shallow machine learning methods include support vector machines, relevance vector machines, and hidden Markov models. However, these methods can only learn shallow features, resulting in limited fault diagnosis performance and requiring complex manual feature engineering, which is very time-consuming and labor-intensive. With the rapid development of deep learning technologies with powerful deep representation learning capabilities, such as recurrent neural networks (RNNs), convolutional neural networks (CNNs), and their various variants such as long short-term memory (LSTM) and gated recurrent units (GRUs), deep learning models have become more effective fault diagnosis technologies. Deep learning models can perform end-to-end fault diagnosis directly from raw flight status monitoring data without any manual feature engineering. These deep learning methods have achieved excellent fault diagnosis performance. However, due to structural limitations such as the gradient vanishing and exploding problems of recurrent neural networks, and the limited receptive field of convolutional neural networks, these CNN, RNN and their variant models are not yet fully capable of modeling the global dependencies of long-term flight status monitoring data.

[0004] In practical engineering, drone swarms are required to perform diverse missions. Depending on the mission requirements, drone swarms must fly under different operating conditions (i.e., different domains), such as different communication topologies, formation shapes, and flight speeds. Under these new operating conditions, the availability of fault data for drone swarms is very limited, hindering the adequate training of data-driven models. This leads to insufficient learning of fault signatures, high model uncertainty, poor generalization performance, and consequently, low fault diagnosis accuracy. However, performing extensive flights under these new operating conditions to acquire sufficient fault data incurs high flight costs. Considering that drone swarms accumulate a large amount of historical fault data from completed missions, and that the same faults occurring in different missions exhibit similar fault signatures, it is of engineering significance to study how to leverage existing fault data to solve the problem of fault diagnosis under new operating conditions with limited data. However, due to differences in communication topologies, flight formation shapes, and speeds across different operating conditions, fault signatures exhibit data distribution discrepancies, which in turn leads to failures in fault diagnosis algorithms. Transfer learning can address this data distribution discrepancy across domains. Therefore, developing a transfer learning approach to solve the problem of fault diagnosis under new operating conditions with limited data is of engineering significance. Summary of the Invention

[0005] Purpose of the invention: The present invention provides a cross-domain fault diagnosis method for UAV clusters based on a multi-source domain deep migration network, which overcomes the problems of difficult extraction of UAV cluster fault features, difficult fusion of multi-source domain features, and difficult migration of cross-domain features. It solves the fault diagnosis problem of UAV clusters based on limited flight data under new operating conditions, and improves the accuracy and generalization of cross-domain fault diagnosis.

[0006] Technical solution: The present invention provides a cross-domain fault diagnosis method for drone clusters based on a multi-source domain deep migration network, comprising the following steps:

[0007] (1) Collect flight status monitoring data corresponding to different faults occurring under different operating conditions of the UAV cluster, perform data preprocessing, divide the data into source domain and target domain, and construct training and test sets;

[0008] (2) Constructing and initializing a multi-source domain deep migration network model, the model includes a domain feature extractor based on spatiotemporal Bayesian graph wavelet convolutional transformer, a multi-source domain cross convolutional transformer, a weight-sharing three-branch cross-domain aligned convolutional transformer, and source domain and target domain classifiers;

[0009] (3) Input N source domain samples and one target domain sample in the training set into the domain feature extractor to extract fault features in different domains;

[0010] (4) Input N source domain fault features into the multi-source domain cross convolution Transformer, learn the relationship between different source domain fault features in a segment-by-segment manner, and obtain multi-source domain fusion features;

[0011] (5) Input the N source domain fault features into the corresponding N source domain classifiers to obtain the fault classification results of the N source domains;

[0012] (6) The multi-source domain fusion features obtained in step (4) and the target domain fault features obtained in step (3) are simultaneously input into the weight-sharing three-branch cross-domain alignment convolutional Transformer, and feature alignment is performed in a segment-by-segment manner to obtain high-level source domain and target domain fault features as well as cross-domain related features;

[0013] (7) Input the fault features obtained in step (6) into the target domain classifier to obtain the fault classification results corresponding to the three branches;

[0014] (8) Repeat steps (3) to (7) until the maximum number of predictions N is reached. p , get N for each training sample p fault classification results in different domains, and calculate the average value of the fault classification results in different domains as the final classification result;

[0015] (9) Calculate the loss function L that integrates multiple factors total , including the hierarchical multi-metric domain difference loss L dis , classification loss and distillation loss L dtl , the back-propagation algorithm is used to update the network parameters to minimize the overall loss value;

[0016] (10) Repeat steps (2) to (9) to traverse all training samples until the maximum number of training rounds M is reached. epoch , obtain the optimal multi-source domain deep migration network model;

[0017] (11) Activate the network in the target domain data flow direction of the optimal multi-source domain deep migration network model, input the test set samples into the optimal multi-source domain deep migration network model, and calculate the corresponding fault classification results;

[0018] (12) Repeat step (11) until the maximum number of predictions N is reached p , get N for each test sample p fault classification results and calculate the average value of the fault classification results as the final classification result, calculate N p The prediction variance V corresponding to the fault classification result p , quantifying the uncertainty of classification predictions.

[0019] Furthermore, the different operating conditions of the UAV cluster in step (1) include the flight of the UAV cluster in different flight formation shapes and different flight speeds under different communication topologies; each operating condition is regarded as a domain; flight status monitoring data under N existing operating conditions are collected as N source domain data sets, and limited flight status monitoring data under a new operating condition is collected as a target domain data set.

[0020] Furthermore, the different faults described in step (1) include sensor and actuator failures of a single UAV, simultaneous sensor and actuator failures of two or more UAVs, communication interruption failures between two UAVs, and propagation failures of non-communicating UAVs.

[0021] Furthermore, the flight status monitoring data in step (1) includes the absolute value of the relative distance change between the current UAV and its neighboring communicating UAV in the lateral and longitudinal directions, the speed deviation and the average value of the overall speed, the lateral and longitudinal speed, climb angle, yaw angle, bank angle and throttle opening of the current UAV.

[0022] Furthermore, the implementation process of step (2) is as follows:

[0023] Given N source domain input samples and a target domain input sample First, the samples of the source domain and the target domain are respectively extracted through the domain feature extractor to extract the fault features in each domain. Afterwards, the N source domain fault features are obtained Perform two operations: First, input the fault features of each domain into the respective source domain classifier to obtain the classification prediction results, i.e. j=1,...,N; secondly, divide the N source domain fault features into N pairs of feature combinations in the ratio of 1:N-1, that is, These combinations are then input into the multi-source domain cross convolution Transformer F mcct (·) to obtain multi-source domain fusion features Next, the multi-source domain fusion features and target domain fault characteristics Input into the weight-sharing three-branch cross-domain alignment convolutional Transformer to perform feature alignment in a segment-by-segment manner to obtain high-level source domain fault features. and target domain fault characteristics and cross-domain related features Finally, the obtained features are input into the target domain classifier Obtain the target domain fault classification results.

[0024] Furthermore, the source domain and target domain classifiers in step (2) have the same network structure, including two Bayesian linear layers, ReLU activation functions and Softmax function layers; the source domain fault features are established through the classifier j=1,...,N, high-level source domain fault characteristics and target domain fault characteristics and cross-domain related features The nonlinear mapping relationship between the fault classification label and the

[0025] Furthermore, the domain feature extractor in step (3) includes a common feature extractor F co (·) and domain-specific feature extractors The common feature extractor consists of a weight-sharing spatiotemporal Bayesian graph wavelet convolutional transformer; the domain-specific feature extractors consist of independent spatiotemporal Bayesian graph wavelet convolutional transformers.

[0026] The spatiotemporal Bayesian graph wavelet convolution transformer includes multi-head spatiotemporal convolution attention, Bayesian linear layer, residual connection and layer normalization; the multi-head spatiotemporal convolution attention includes H parallel spatiotemporal convolution attentions, each of which extracts spatiotemporal coupling fault features from different scales; the input sample passes through the H spatiotemporal convolution attentions to obtain H spatiotemporal coupling fault features and splices them, and then passes through two layers of Bayesian linear layers with residual connection and layer normalization in sequence to obtain the final spatiotemporal coupling fault features;

[0027] The spatiotemporal convolutional attention includes multi-layer multi-scale adaptive graph wavelet convolution, local and global scaling dot product attention, dilated causal convolution network layer and dimension adjustment layer; first, multi-layer multi-scale adaptive graph wavelet convolution processes the input sample Get the corresponding query vector Q h , key vector K h Sum value vector V h , and input into the local scaled dot product attention to obtain the spatial feature sequence Then the dimension adjustment layer adjusts the dimension of the spatial feature sequence to obtain where d model =Ne d ×N f ; Then, two parallel dilated causal convolutional network layers learn the spatial feature sequence The local context dependency relationship is used to obtain the corresponding query vector Q h ′, key vector K h '; Similarly, the dimension adjustment layer adjusts the input samples Dimensions obtained and use it as the value vector Vh '; the query vector Q h ′, key vector K h ′ and value vector V h ′Input into the global scaled dot product attention to adaptively integrate the local temporal features of the input samples into the modeling of global dependencies to simultaneously model local and global spatiotemporal coupling features and extract spatiotemporal coupling fault features;

[0028] The multi-layer multi-scale adaptive graph wavelet convolution is composed of K layers of multi-scale adaptive graph wavelet convolution connected in series; each multi-scale adaptive graph wavelet convolution includes |S| scale adaptive graph wavelet convolutions to extract spatial features from input samples at different scales; each adaptive graph wavelet convolution is implemented by graph wavelet transform; the graph wavelet transform uses a set of appropriate wavelet bases provided by wavelet transform to project the graph signal from the vertex domain to the spectral domain and extract fault features in the spectral domain; the wavelet base is defined as Where Ne is the number of drones, each wavelet basis ψ s,i Corresponding to the signal of the i-th UAV at scale s, the graph wavelet with scale parameter s on the i-th UAV can be written in the following matrix multiplication form:

[0029] Ψ s =UG s U T

[0030] in, is the scaling matrix, λ l is the normalized graph Laplacian matrix The lth eigenvalue of , U is Eigenvector matrix of ; normalized graph Laplacian matrix It is obtained from the adjacency matrix A corresponding to the UAV cluster communication topology, which is defined as Among them I n is the identity matrix of the corresponding dimension, is the angle matrix The graph wavelet basis is approximated by Chebyshev polynomials as follows:

[0031]

[0032] c i,s =2e -s J i (-s)

[0033] in, is the i-th order Chebyshev polynomial, Q is the number of Chebyshev polynomials, J i(-s) is the first kind of Bessel function; for input samples with graph structure The wavelet transform of the graph is The corresponding inverse transformation is The adaptive graph wavelet convolution with scale s is defined as:

[0034]

[0035] Among them, s is a graph wavelet basis with scale s, W s (i,j) are learnable parameters, is the output signal, σ(·) is the nonlinear activation function; the adaptive graph wavelet convolutions with different scales are connected in parallel and combined with the residual connection as an adaptive component to form a multi-scale adaptive graph wavelet convolution, and then multiple multi-scale adaptive graph wavelet convolutions are connected in series to form a multi-layer multi-scale adaptive graph wavelet convolution; in the domain d∈{d s1 ,...,d sN ,d t The corresponding output of the s-th scale of the K-th layer adaptive graph wavelet convolution in} is:

[0036]

[0037] in, is the initial input of the network, and are the graph wavelet basis and its inverse at scale s, Θ d,s is the diagonal matrix of the graph convolution kernel, and is a trainable parameter matrix, σ(·) is a nonlinear activation function; in the domain d∈{d s1 ,...,d sN ,d t In

[15] , the final output of the multi-layer multi-scale adaptive graph wavelet convolution is the average value of the output features of |S| parallel adaptive graph wavelet convolutions in the K-th layer of multi-scale adaptive graph wavelet convolution, which is defined as follows:

[0038]

[0039] Furthermore, the multi-source domain cross convolution Transformer in step (4) includes N multi-source cross convolution attention layers, residual connections, layer normalization, and linear layers; it is used to learn the relationship between different source domain features and extract multi-source domain fusion features;

[0040] The multi-source cross convolution attention includes an expanded causal convolutional network layer, a repeated layer, and a scaled dot product attention; N source domain fault features are sequentially regarded as value vectors, and the remaining N-1 source domain fault features are regarded as key vectors and value vectors; specifically, the N source domain fault features are sequentially divided into N pairs of feature combinations according to the ratio of 1:N-1, namely Input N pairs of feature combinations into N multi-source cross convolution attentions respectively, It corresponds to The splicing features of the fault features of the remaining source domains except ; After the dilated causal convolutional network layer, the value vector V is obtained cn , After processing by the dilated causal convolutional network layer, the query vector Q is obtained cn , key vector K cn , the repeated layer passes through the value vector V cn Perform a copy operation to make its dimension the same as the corresponding query vector Q cn , key vector K cn Consistent; query vector Q cn , key vector K cn Sum value vector V cn Input scaled dot product attention, calculate attention, learn the relationship between each source domain feature and the other source domain features in a segment-by-segment manner, and obtain N multi-source domain interaction features {Attn c1 ,...,Attn cN}; Further, the N multi-source domain interaction features are spliced ​​and then further fused through two layers of linear layers with residual connections and layer normalization to obtain the final multi-source domain fusion feature

[0041] Furthermore, the weight-sharing three-branch cross-domain aligned convolutional Transformer in step (6) includes a source branch, a source-target branch, and a target branch; the multi-source domain fusion features and the target domain fault features are aligned in a segment-by-segment manner, and domain-invariant features are extracted to achieve the improvement of the fault diagnosis performance in the target domain by utilizing the source domain fault features; the source branch, the source-target branch, and the target branch are all composed of a weight-sharing convolutional Transformer; the multi-source domain fusion features Input to the source branch, and processed by the dilated causal convolutional network layer to obtain the query vector Q s , key vector K s Sum value vector V s , and then through multi-head self-attention, two-layer residual connection, layer normalization and linear layer to obtain high-level source domain fault features Similarly, the target domain fault characteristics Input into the target branch, and processed by the dilated causal convolutional network layer to obtain the query vector Qt , key vector K t Sum value vector V t , and then through multi-head self-attention, two-layer residual connection, layer normalization and linear layer to obtain high-level target domain fault features The query vector Q in the source branch s and the key vector K of the target branch t Sum value vector V t At the same time, the multi-head cross-domain attention of the source-target branch is input to learn the relationship between the source domain and target domain features to obtain cross-domain related features; then the multi-source domain fusion features are and target domain high-level features Splicing to obtain cross-domain combined features F cd , and input it into the residual connection and layer normalization in the source-target branch to fuse with the cross-domain alignment feature, and then pass through the linear layer and residual connection and layer normalization to obtain the cross-domain related features Next, the cross-domain related features obtained by the three branches are and They are respectively input into the weight-shared target domain classifier and output the corresponding fault classification value.

[0042] Furthermore, the hierarchical multi-metric domain difference loss L in step (9) dis Used to optimize the distribution difference of fault features at different levels of each pair of target domains and source domains; it includes the low-level domain distribution loss and the high-level domain distribution loss obtained by weighted addition of the maximum average difference loss and the CORAL distance, which are used to measure the distribution difference between the fault features of the source domain samples after passing through the common feature extractor and the specific domain feature extractor and the corresponding target domain fault features;

[0043] The classification loss Used to optimize the difference between all classifier output values ​​and fault classification label values; cross entropy loss is selected as the classification loss, and the total classification loss is in represents the classification loss value corresponding to the N source domain classifier, The classification loss values ​​of the source branch and the target branch corresponding to the target domain classifier;

[0044] The distillation loss L dtl It is used to improve the fault diagnosis performance in the target domain by using the fault features in the source domain; specifically, the cross-domain related features output by the source-target branch in the weight-sharing three-branch cross-domain alignment convolutional Transformer are The fault classification probability distribution output in the target domain classifier is used as a soft label to supervise the training of the target branch through the following distillation loss to improve the classification performance of the diagnosis model on the target domain samples.

[0045]

[0046] Among them, q k and p k are the probabilities of the classification category k corresponding to the outputs of the source-target branch and the target branch, respectively;

[0047] The comprehensive multi-factor loss function L total for,

[0048]

[0049] Among them, λ and β are weight parameters.

[0050] Beneficial effects: Compared with the prior art, the present invention has the following beneficial effects:

[0051] 1. In terms of fault feature extraction, the proposed spatiotemporal Bayesian graph wavelet convolution transformer can integrate physically meaningful wavelet transforms into graph convolutional neural networks to design multi-scale adaptive graph wavelet convolutions to extract multi-scale spatial fault features. Global scaled dot product self-attention then adaptively integrates the local temporal features of the input samples into the modeling of global dependencies, allowing for simultaneous modeling of local and global temporal features. This allows for the extraction of spatiotemporal coupled features, with high generalization and interpretability.

[0052] 2. In terms of multi-source domain feature fusion, the present invention proposes a multi-source domain cross-convolution Transformer. By dividing N fault features into N pairs of feature combinations in a ratio of 1:N-1, and inputting them into N multi-source cross-convolution attentions, the relationship between different source domain features is learned in a segment-by-segment manner. This can fully fuse multi-source domain features and extract more valuable multi-source domain fusion features.

[0053] 3. In terms of cross-domain feature alignment, a weight-sharing three-branch cross-domain alignment convolutional Transformer is proposed. Through three weight-sharing source branches, source-target branches, and target branches, multi-source domain fusion features are aligned with target domain fault features in a segment-by-segment manner, extracting domain-invariant features and leveraging source domain features to improve fault diagnosis performance in the target domain.

[0054] 4. In terms of loss function optimization, this paper proposes a comprehensive multi-factor loss function that includes hierarchical multi-metric domain difference loss, classification loss, and distillation loss. This function can optimize feature distribution differences and classification losses in different domains from multiple perspectives, thereby improving fault diagnosis performance in the target domain.

[0055] In summary, the present invention overcomes the problems of difficulty in extracting fault features of drone clusters, difficulty in fusing multi-source domain features, and difficulty in migrating cross-domain features. It solves the problem of fault diagnosis of drone clusters based on limited flight data under new operating conditions, and improves the accuracy and generalization of cross-domain fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is a flow chart of the present invention;

[0057] Figure 2 This is a schematic diagram of the multi-source domain deep migration network structure;

[0058] Figure 3 Schematic diagram of the spatiotemporal Bayesian graph wavelet convolution Transformer structure;

[0059] Figure 4 Schematic diagram of the multi-source domain cross convolution Transformer structure;

[0060] Figure 5 Four different operating conditions of drone swarms: (a) square flight formation and communication topology; (b) V-shaped flight formation and communication topology; (c) pentagonal flight formation and communication topology; (d) cross-shaped flight formation and communication topology;

[0061] Figure 6 These are the fault diagnosis results of the UAV cluster based on the method of the present invention under four cross-domain diagnosis tasks. DETAILED DESCRIPTION

[0062] The present invention will be further described in detail below with reference to the accompanying drawings.

[0063] like Figure 1 As shown, the present invention provides a cross-domain fault diagnosis method for UAV clusters based on a multi-source domain deep migration network. The specific implementation process is as follows:

[0064] Step 1: Collect flight status monitoring data corresponding to different faults occurring under different operating conditions (different domains) of the drone cluster, perform data preprocessing, divide the data into source domain and target domain, and construct training and test sets.

[0065] Different operating conditions of drone clusters include drone clusters flying in different flight formation shapes and at different flight speeds under different communication topologies; each operating condition is regarded as a domain; flight status monitoring data under N operating conditions are collected as N source domain data sets, and the flight status monitoring data under one operating condition is collected as a target domain data set.

[0066] Different failures that can occur in a drone swarm include sensor and actuator failures of a single drone, simultaneous sensor and actuator failures of two or more drones, communication interruption failures between two drones, and propagation failures between non-communicating drones.

[0067] The flight status monitoring data includes the absolute value of the relative distance change between the current UAV and its neighboring communicating UAVs in the lateral and longitudinal directions, the speed deviation and the average value of the overall speed, the lateral and longitudinal speed of the current UAV, the climb angle, yaw angle, bank angle and throttle opening.

[0068] The flight status monitoring data is normalized to the interval [0, 1] using the minimum-maximum normalization formula. A sliding window approach is used, with a step size of 1 and a sliding window length of 10 sliding from the beginning to the end of the flight status monitoring data to segment the data. Samples are generated per sliding window, with each sample corresponding to a fault classification label value, ultimately generating a sample set that can be used for network model training and testing.

[0069] N labeled source domain samples and some labeled target domain samples are used as training sets, and unused labeled target domain samples are used as test sets.

[0070] Step 2: Figure 2 As shown in the figure, a multi-source domain deep migration network model is constructed and initialized, including a domain feature extractor based on spatiotemporal Bayesian graph wavelet convolutional Transformer, a multi-source domain cross-convolutional Transformer, a weight-sharing three-branch cross-domain aligned convolutional Transformer, and source domain and target domain classifiers.

[0071] The data flow of the multi-source domain deep migration network model is: given N source domain input samples and a target domain input sample First, the samples of the source domain and the target domain are respectively extracted through the domain feature extractor to extract the fault features in each domain. Afterwards, the N source domain fault features are obtained Perform two operations: First, input the fault features of each domain into the respective source domain classifier to obtain the classification prediction results, i.e. j=1,...,N; secondly, divide the N source domain fault features into N pairs of feature combinations in the ratio of 1:N-1, that is, These combinations are then input into the multi-source domain cross convolution Transformer F mcct (·) to obtain multi-source domain fusion features Next, the multi-source domain fusion features and target domain fault characteristics Input into the weight-sharing three-branch cross-domain alignment convolutional Transformer to perform feature alignment in a segment-by-segment manner to obtain high-level source domain fault features. and target domain fault characteristics and cross-domain related features Finally, the obtained features are input into the target domain classifier Obtain the target domain fault classification results.

[0072] Step 3: Input N source domain samples and one target domain sample in the training set into the domain feature extractor to extract fault features in different domains.

[0073] The domain feature extractor based on spatiotemporal Bayesian graph wavelet convolution transformer includes a common feature extractor F co (·) and domain-specific feature extractors The common feature extractor consists of a weight-sharing spatiotemporal Bayesian graph wavelet convolutional Transformer, and the specific domain feature extractors consist of independent spatiotemporal Bayesian graph wavelet convolutional Transformers.

[0074] like Figure 3 As shown in the figure, the spatiotemporal Bayesian graph wavelet convolutional transformer includes multi-head spatiotemporal convolutional attention, Bayesian linear layers, residual connections, and layer normalization. The multi-head spatiotemporal convolutional attention consists of H parallel spatiotemporal convolutional attentions. The spatiotemporal Bayesian graph wavelet convolutional transformer works as follows: the input sample passes through H spatiotemporal convolutional attentions to obtain H spatiotemporal coupled fault features, which are then concatenated. The input sample then passes through two layers of Bayesian linear layers with residual connections and layer normalization to obtain the final spatiotemporal coupled fault features.

[0075] like Figure 3 As shown in the figure, the multi-head spatiotemporal convolutional attention is a key core module for extracting spatiotemporal coupling features of drone cluster systems. It consists of H parallel spatiotemporal convolutional attentions, each of which extracts spatiotemporal coupling fault features from different scales. Each spatiotemporal convolutional attention includes multiple layers of multi-scale adaptive graph wavelet convolution, local and global scaling dot product attention, dilated causal convolution network layer and dimension adjustment layer. Its working principle is as follows: First, multiple layers of multi-scale adaptive graph wavelet convolution extract input samples The spatial features of different scales at each time point in the quantizer are aggregated to obtain the corresponding query vector Q. h , key vector K h Sum value vector V h , and then input it into the local scaled dot product attention to establish the local dependency of spatial features to obtain the spatial feature sequence Then adjust the dimension of the spatial feature sequence through the dimension adjustment layer to where d model =Ne d ×N f ; Then, two parallel dilated causal convolutional network layers process the spatial feature sequence Extract local temporal features and output them as query vector Q h ′ and key vector K h ′Input to the global scaled dot product self-attention; input sample at the same time After the dimension adjustment layer, the dimension is adjusted to and as the value vector V h ′Input the global scaled dot product self-attention; finally, the global scaled dot product self-attention adaptively incorporates the local temporal features of the input sample into the modeling of the global dependency to simultaneously model the local and global spatiotemporal coupling features, thereby realizing the extraction of spatiotemporal coupling features, which can effectively improve the generalization of feature learning.

[0076] Multi-layer multi-scale adaptive graph wavelet convolution consists of K layers of multi-scale adaptive graph wavelet convolution connected in series; each multi-scale adaptive graph wavelet convolution includes |S| scale adaptive graph wavelet convolutions, extracting spatial features from input samples at different scales; each adaptive graph wavelet convolution is implemented by graph wavelet transform. Graph wavelet transform uses a set of suitable wavelet bases provided by wavelet transform to project the graph signal from the vertex domain to the spectral domain and extract fault features in the spectral domain; the wavelet base is defined as Where Ne is the number of drones, each wavelet basis ψ s,i Corresponding to the signal of the i-th UAV at scale s, the graph wavelet with scale parameter s on the i-th UAV can be written in the following matrix multiplication form:

[0077] Ψ s =UG s U T

[0078] in, is the scaling matrix, λ l is the normalized graph Laplacian matrix The lth eigenvalue of , U is The eigenvector matrix of the normalized graph Laplacian matrix It is obtained from the adjacency matrix A corresponding to the UAV cluster communication topology, which is defined as Among them I n is the identity matrix of the corresponding dimension, is the angle matrix The graph wavelet basis is approximated by Chebyshev polynomials as follows:

[0079]

[0080] c i,s =2e -s J i (-s)

[0081] in, is the i-th order Chebyshev polynomial, Q is the number of Chebyshev polynomials, J i (-s) is the first kind of Bessel function; for input samples with graph structure The wavelet transform of the graph is The corresponding inverse transformation is The adaptive graph wavelet convolution with scale s is defined as:

[0082]

[0083] Among them, s is a graph wavelet basis with scale s, W s (i,j) are learnable parameters, is the output signal, σ(·) is the nonlinear activation function. The adaptive graph wavelet convolutions with different scales are connected in parallel and combined with the residual connection as an adaptive component to form a multi-scale adaptive graph wavelet convolution. Then, multiple multi-scale adaptive graph wavelet convolutions are connected in series to form a multi-layer multi-scale adaptive graph wavelet convolution. In the domain d∈{d s1 ,...,d sN ,d t The corresponding output of the s-th scale of the K-th layer adaptive graph wavelet convolution in} is:

[0084]

[0085] in, is the initial input of the network, and are the graph wavelet basis and its inverse at scale s, Θ d,s is the diagonal matrix of the graph convolution kernel, and is a trainable parameter matrix, σ(·) is a nonlinear activation function. s1 ,...,d sN ,d t In

[15] , the final output of the multi-layer multi-scale adaptive graph wavelet convolution is the average value of the output features of |S| parallel adaptive graph wavelet convolutions in the K-th layer of multi-scale adaptive graph wavelet convolution, which is defined as follows:

[0086]

[0087] In summary, the entire processing process of the spatiotemporal Bayesian graph wavelet convolution Transformer can be described by the formula:

[0088] F d =LayerNorm(MCdropout(ReLU(F out1 W h2 +b h2 ))+F out1 )

[0089]

[0090] Fea st =Concat(Attn1,...,Attn H )

[0091]

[0092] in, Represents the output of multi-layer multi-scale adaptive graph wavelet convolution in the h-th spatiotemporal convolutional attention, and the weight of the expanded causal convolution layer and and bias b mf , Attn h represents the h-th spatiotemporal convolution attention output, W h1 and W h2 is the weight of the Bayesian linear layer, b h1 and b h2 is the bias of the Bayesian linear layer, MCdropout(·) is the random dropout function, and F d is the output feature.

[0093] Step 4: Input N source domain fault features into Figure 4 The multi-source domain cross-convolution Transformer shown learns the relationship between fault features of different source domains in a segment-by-segment manner to obtain multi-source domain fusion features.

[0094] The multi-source domain cross convolution Transformer includes N multi-source cross convolution attentions, residual connections, layer normalization, and linear layers; it is used to learn the relationship between different source domain features and extract multi-source domain fusion features. The multi-source cross convolution attention includes an expanded causal convolutional network layer, a repeated layer, and a scaled dot product attention; each of the N source domain fault features is taken as a value vector, and the remaining N-1 source domain fault features are taken as key vectors and value vectors. Specifically, the N source domain fault features are divided into N pairs of feature combinations in a ratio of 1:N-1, namely Input N pairs of feature combinations into N multi-source cross convolution attentions respectively, It corresponds to The splicing features of the fault features of the remaining source domains except ; After the dilated causal convolutional network layer, the value vector V is obtained cn , After processing by the dilated causal convolutional network layer, the query vector Q is obtained cn , key vector K cn , the repeated layer passes through the value vector V cn Perform a copy operation to make its dimension the same as the corresponding query vector Q cn , key vector K cn Consistent; query vector Q cn , key vector K cn Sum value vector V cn Input scaled dot product attention, calculate attention, learn the relationship between each source domain feature and the other source domain features in a segment-by-segment manner, and obtain N multi-source domain interaction features {Attn c1 ,...,Attn cN Furthermore, the N multi-source domain interaction features are spliced ​​and then further fused through two layers of linear layers with residual connections and layer normalization to obtain the final multi-source domain fusion feature. The entire processing process of the multi-source domain cross convolution Transformer can be described as:

[0095]

[0096] Fea ca =Concat(Attn c1 ,...,Attn cN )

[0097]

[0098] Among them, Attn cn represents the n-th multi-source domain interaction feature, ★ Cross-correlation calculation, dilating the weights of the causal convolutional network layer and and bias b cn , is the fault feature of the nth source domain, For the corresponding The splicing features of the fault features of the remaining source domains except , Repeat(·) is the repeated layer, It is the final multi-source domain fusion feature.

[0099] Step 5: Input the N source domain fault features into the corresponding N source domain classifiers to obtain the fault classification results of the N source domains.

[0100] The source domain and target domain classifiers have the same network structure, including two Bayesian linear layers, ReLU activation function and Softmax function layer; the source domain fault features are established through the classifier j=1,...,N, high-level source domain fault characteristics and target domain fault characteristics and cross-domain related features The nonlinear mapping relationship between the source domain classifier and the fault classification label can be described as follows:

[0101]

[0102] in, is the corresponding source domain classifier, is the corresponding fault classification result, d s ∈{d s1 ,...,d sN}.

[0103] Step 6: Input the multi-source domain fusion features obtained in step 4 and the target domain fault features obtained in step 3 into the weight-sharing three-branch cross-domain alignment convolutional Transformer at the same time, perform feature alignment in a segment-by-segment manner, and obtain high-level source domain and target domain fault features as well as cross-domain related features.

[0104] like Figure 2 As shown in the figure, the weight-sharing three-branch cross-domain alignment convolutional Transformer includes a source branch, a source-target branch, and a target branch. The multi-source domain fusion features are aligned with the target domain fault features in a segment-by-segment manner to extract domain-invariant features, thereby utilizing the source domain fault features to improve the fault diagnosis performance in the target domain; the source branch, the source-target branch, and the target branch are all composed of a weight-sharing convolutional Transformer. Multi-source domain fusion features Input to the source branch, and processed by the dilated causal convolutional network layer to obtain the query vector Q s , key vector K s Sum value vector V s , and then through multi-head self-attention, two-layer residual connection, layer normalization and linear layer to obtain high-level source domain fault features Similarly, the target domain fault characteristics Input into the target branch, and processed by the dilated causal convolutional network layer to obtain the query vector Q t , key vector K t Sum value vector V t , and then through multi-head self-attention, two-layer residual connection, layer normalization and linear layer to obtain high-level target domain fault features The query vector Q in the source branch sand the key vector K of the target branch t Sum value vector V t At the same time, the multi-head cross-domain attention is input into the source-target branch to learn the relationship between the source domain and target domain features, and obtain cross-domain related features; the multi-head cross-domain attention is realized by the multi-head self-attention; then the multi-source domain fusion features are and target domain high-level features Splicing to obtain cross-domain combined features F cd , and input it into the residual connection and layer normalization in the source-target branch to fuse with the cross-domain alignment feature, and then pass through the linear layer and residual connection and layer normalization to obtain the cross-domain related features Next, the cross-domain related features obtained by the three branches are and The three branches of the weight-sharing cross-domain aligned convolutional transformer are respectively input into the target domain classifier and output the corresponding fault classification value. The processing process of the different branches of the weight-sharing three-branch cross-domain aligned convolutional transformer can be described as follows. The source branch is:

[0105]

[0106] The target branch is:

[0107]

[0108] The source-target branches are:

[0109]

[0110] Among them, W s q , W s k , and W s v is the weight of the dilated causal convolutional network layer, b s is the bias of the dilated causal convolutional network layer, H m is the number of attentions in multi-head self-attention, W s1 and W s2 is the weight of the linear layer, b s1 and b s2 is the bias of the linear layer.

[0111] Step 7: Input the fault features obtained in step 6 into the target domain classifier to obtain the fault classification results corresponding to the three branches.

[0112] like Figure 2 As shown in Figure 2, the target domain classifier has the same structure as the source domain classifier mentioned above. A target domain classifier can be described as follows:

[0113]

[0114] in, is the corresponding target domain classifier, is the corresponding fault classification result, d t ∈{d ts ,d tst ,d tt}.

[0115] Step 8: Repeat steps 3 to 7 until the maximum number of predictions N is reached. p , get N for each training sample p The fault classification results in different domains are calculated and the average value of the fault classification results in different domains is taken as the final classification result.

[0116] Step 9: Calculate the loss function L that integrates multiple factors total Including hierarchical multi-metric domain difference loss L dis , classification loss and distillation loss L dtl , the back-propagation algorithm is used to update the network parameters to minimize the overall loss value.

[0117] Hierarchical multi-metric domain difference loss L dis It is used to optimize the distribution difference of different-level features of each pair of target domain and source domain, helping the network learn domain-invariant features. It consists of two parts: low-level domain distribution loss and high-level domain distribution loss based on maximum average difference loss and CORAL distance; they are used to reduce the distribution distance between the features of source domain samples after passing through the common feature extractor and the specific domain feature extractor and the corresponding target domain features; the two metric functions of maximum average difference and CORAL distance are defined as follows. Given two domains and The maximum mean difference can be obtained by the unbiased estimate of the following formula:

[0118]

[0119] Where H is the reproducing kernel Hilbert space with characteristic kernel K, and φ(·) represents the mapping function that can map the original x to the reproducing kernel Hilbert space. The CORAL distance is the distance between the covariances of the data in the two domains and is defined as follows:

[0120]

[0121] in, represents the Frobenius norm of the square matrix, C s and C s Represents the covariance matrix of the source domain and target domain features respectively. Based on the above definition of the maximum mean difference and CORAL distance, taking the jth source domain as an example, the low- and high-level domain distribution loss and high-level domain distribution loss and hierarchical multi-metric domain difference loss L dis It can be defined as follows:

[0122]

[0123] Among them, ε is the weight parameter, F MMD (·) represents the maximum average difference loss, F CORAL (·) is the CORAL distance, F co (·) is the common feature extractor, is a domain-specific feature extractor. Therefore, we can minimize minL dis The parameters of the common feature extractor and the domain-specific feature extractor are optimized to narrow the data distribution difference between the source domain features and the target domain features.

[0124] Classification loss d∈{d s1 ,...,d sN ,d tt ,d ts}Use the cross entropy loss defined as follows, for a given dataset

[0125]

[0126] in, is the true label value, is the output value of the classifier. The total classification loss is in represents the classification loss value corresponding to the N source domain classifier, The classification loss values ​​of the source and target branches correspond to the target domain classifiers. The classification loss is used to optimize the difference between all classifier output values ​​and the fault classification label values.

[0127] Distillation loss It is used to use source domain features to assist fault diagnosis in the target domain, so as to improve the target domain fault diagnosis performance under small sample conditions; specifically, the cross-domain related features output by the source-target branch in the three-branch cross-domain alignment convolutional Transformer are The fault classification probability distribution output by the target domain classifier is used as a soft label to supervise the training of the target branch through the distillation loss defined as follows to improve the classification performance of the model on target domain samples:

[0128]

[0129] Among them, q k and p kare the probabilities of classification category k corresponding to the outputs from the source-target branch and the target branch, respectively.

[0130] Therefore, the loss function L that integrates multiple factors total It can be defined as:

[0131]

[0132] Among them, λ and β are weight parameters.

[0133] Step 10: Repeat steps 2 to 9 to traverse all training samples until the maximum number of training rounds M is reached. epoch , and obtain the optimal multi-source domain deep migration network model.

[0134] Step 11: Activate the network in the target domain data flow direction of the optimal multi-source domain deep migration network model, input the test set samples into the optimal multi-source domain deep migration network model, and calculate the corresponding fault classification results.

[0135] Step 12: Repeat step 11 until the maximum number of predictions N is reached. p , get N for each test sample p fault classification results and calculate the average value of the fault classification results as the final classification result, calculate N p The prediction variance V corresponding to the fault classification result p , quantifying the uncertainty of classification predictions.

[0136] Prediction variance V p It is defined as,

[0137]

[0138] Among them, σ 2 Yes N p The variance of the classification prediction results, Softplus(·) is the softplus activation function used to ensure the effectiveness of the variance. The higher the prediction variance V p Indicates N p The random prediction results are more discrete and have greater uncertainty. On the contrary, the lower the prediction variance V p Indicates N p The results of sub-random predictions are more stable, with less uncertainty, and more reliable.

[0139] A semi-physical simulation platform for drone swarm systems collects flight status monitoring data corresponding to different fault conditions in drone swarms. The semi-physical simulation platform primarily consists of a Raspberry Pi cluster, a server, a switch, a router, and a data analysis terminal (PC). Each pair of Raspberry Pis in the cluster simulates a drone. One Raspberry Pi serves as the flight controller system, while the other simulates the drone's dynamics, actuators, sensors, and environment. The Raspberry Pis transmit information via the switch to the router, and then to the database server and neighboring computing server for information exchange. The database server, connected to the router, stores drone swarm status information. The neighboring computing server accesses the database, calculates each drone's neighboring drone information based on the communication topology, and sends this data to the flight controller. The flight controller generates control commands based on the received sensor data and neighboring drone information, and sends them to the dynamic environment component to control the drone swarm's flight. The data analysis terminal (PC) retrieves the drone swarm status information stored in the database and displays the swarm's flight status using the situation display terminal software.

[0140] N = 4 operating conditions were selected, namely square, V-shaped, pentagonal and cross flight formations and the corresponding communication topologies and flight speeds, as shown in Table 1 and Figure 5 As shown in the figure, (a) is a square flight formation and communication topology; (b) is a V-shaped flight formation and communication topology; (c) is a pentagonal flight formation and communication topology; and (d) is a cross-shaped flight formation and communication topology. Under each operating condition, 12 types of drone cluster faults are set as shown in Table 2, and flight status monitoring data is collected as shown in Table 3. To verify the fault diagnosis performance of the present invention based on limited data under the new operating conditions of drone clusters, four cross-domain diagnostic tasks for drone clusters are set, as shown in Table 4.

[0141] Table 1 Description of four flight conditions of UAV swarm

[0142] Working conditions Communication topology Formation shape Flight speed (m / s) Flight time (s) FT1 square square 35 600 FT2 V-shaped V-shaped 25 600 FT3 pentagon pentagon 30 600 FT4 Cross Cross 20 600

[0143] Table 2 UAV cluster failure types

[0144] Fault Label Fault type Health - 0 Health status Fault-1 Multiplicative failure of a single sensor Fault-2 Additive fault of single sensor Fault-3 Single-machine sensor time-varying fault Fault-4 Single engine multiplication failure Fault-5 Single-machine control surface additive failure Fault-6 Combined failure of two machines Fault-7 Combined failure of two machines Fault-8 Single-machine propagation failure outside the communication topology Fault-9 Two-machine propagation failure outside the communication topology Fault-10 Three-machine propagation failure outside the communication topology Fault-11 Communication failure between the two machines Fault-12 Communication failure between the two machines

[0145] Table 3 UAV cluster flight status monitoring data

[0146]

[0147] Table 4 Four cross-domain diagnostic tasks of drone clusters

[0148]

[0149] In order to evaluate the fault diagnosis performance of different models, three evaluation indicators are selected: Accuracy, Recallrate, and F1-score, which are defined as follows: Among them, TP, TN, FP and FN are the number of correctly classified positive samples, the number of correctly classified negative samples, the number of incorrectly classified negative samples, and the number of incorrectly classified positive samples, respectively.

[0150] like Figure 6 Table 5 shows the fault diagnosis results of the proposed method for drone swarms in four cross-domain diagnostic tasks. The experimental results show that the proposed method can obtain accurate fault diagnosis results in various cross-domain fault diagnosis tasks, demonstrating its effectiveness. Furthermore, diagnostic uncertainty remains low across all tasks, with minimal fluctuations, verifying the reliability of the fault diagnosis results.

[0151] Table 5 Diagnosis results of drone clusters in four cross-domain fault diagnosis tasks

[0152] Task Accuracy Recall F1 score uncertainty A 94.58％ 94.53％ 94.51％ 0.023 B 93.96％ 93.96％ 93.99％ 0.015 C 94.41％ 94.40％ 94.38％ 0.020 D 94.58％ 94.58％ 94.54％ 0.018

[0153] To further verify the superiority of our method, we selected existing ensemble transfer learning methods, a multi-source domain adaptation network, and a multi-source weighted deep transfer network as comparative analysis methods. Table 6 shows the corresponding fault diagnosis results for Task A using different methods.

[0154] Table 6 Diagnosis results of different methods under task A

[0155] method Accuracy Recall F1 score Method of the present invention 94.58％ 94.53％ 94.51％ Ensemble transfer learning methods 89.63％ 89.07％ 89.25％ Multi-source domain adaptation network 87.36％ 87.42％ 87.38％ Multi-source partial migration network 86.78％ 86.57％ 86.74％

[0156] According to the comparative diagnostic results of the method of the present invention and the integrated transfer learning method, the multi-source domain adaptive network and the multi-source partial transfer network in Table 6, it can be seen that the cross-domain fault diagnosis method of the drone cluster based on the multi-source domain deep transfer network proposed in the present invention obtains the best values ​​in all three diagnostic performance evaluation indicators, which indicates that this method can provide more accurate and reliable cross-domain fault diagnosis results of the drone cluster.

[0157] The method of the present invention can extract more accurate, rich and complete spatiotemporal coupling fault characteristics and uncertainty information from the UAV cluster flight status monitoring data, and has high generalization and robustness.

[0158] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A cross-domain fault diagnosis method for drone clusters based on a multi-source domain deep migration network, characterized by: The following steps are involved: (1) Collect flight status monitoring data corresponding to different faults occurring under different operating conditions of the UAV cluster, perform data preprocessing, divide the data into source domain and target domain, and construct training and test sets; (2) Constructing and initializing a multi-source domain deep migration network model, the model includes a domain feature extractor based on spatiotemporal Bayesian graph wavelet convolutional transformer, a multi-source domain cross convolutional transformer, a weight-sharing three-branch cross-domain aligned convolutional transformer, and source domain and target domain classifiers; (3) Input N source domain samples and one target domain sample in the training set into the domain feature extractor to extract fault features in different domains; (4) Input N source domain fault features into the multi-source domain cross convolution Transformer, learn the relationship between different source domain fault features in a segment-by-segment manner, and obtain multi-source domain fusion features; (5) Input the N source domain fault features into the corresponding N source domain classifiers to obtain the fault classification results of the N source domains; (6) The multi-source domain fusion features obtained in step (4) and the target domain fault features obtained in step (3) are simultaneously input into the weight-sharing three-branch cross-domain alignment convolutional Transformer, and feature alignment is performed in a segment-by-segment manner to obtain high-level source domain and target domain fault features as well as cross-domain related features; (7) Input the fault features obtained in step (6) into the target domain classifier to obtain the fault classification results corresponding to the three branches; (8) Repeat steps (3) to (7) until the maximum number of predictions N is reached. p , get N for each training sample p fault classification results in different domains, and calculate the average value of the fault classification results in different domains as the final classification result; (9) Calculate the loss function L that integrates multiple factors total , including the hierarchical multi-metric domain difference loss L dis , classification loss And distillation loss L, use the back propagation algorithm to update the network parameters to minimize the overall loss value; dtl (10) Repeat steps (2) to (9) to traverse all training samples until the maximum number of training rounds M is reached. epoch , obtain the optimal multi-source domain deep migration network model; (11) Activate the network in the target domain data flow direction of the optimal multi-source domain deep migration network model, input the test set samples into the optimal multi-source domain deep migration network model, and calculate the corresponding fault classification results; (12) Repeat step (11) until the maximum number of predictions N is reached p , get N for each test sample p fault classification results and calculate the average value of the fault classification results as the final classification result, calculate N p The prediction variance V corresponding to the fault classification result p , quantifying the uncertainty of classification predictions.

2. The cross-domain fault diagnosis method for drone clusters based on a multi-source domain deep migration network according to claim 1 is characterized in that: The different operating conditions of the UAV cluster in step (1) include the flight of the UAV cluster in different flight formation shapes and different flight speeds under different communication topologies; each operating condition is regarded as a domain; flight status monitoring data under N existing operating conditions are collected as N source domain data sets, and limited flight status monitoring data under a new operating condition is collected as a target domain data set.

3. The cross-domain fault diagnosis method for drone clusters based on a multi-source domain deep migration network according to claim 1 is characterized in that: The different faults described in step (1) include sensor and actuator failures of a single UAV, simultaneous sensor and actuator failures of two or more UAVs, communication interruption failures between two UAVs, and propagation failures of non-communicating UAVs.

4. The cross-domain fault diagnosis method for drone clusters based on a multi-source domain deep migration network according to claim 1 is characterized in that: The flight status monitoring data in step (1) includes the absolute value of the relative distance change between the current UAV and its neighboring communicating UAV in the horizontal and vertical directions, the speed deviation and the average value of the overall speed, the horizontal and vertical speed of the current UAV, the climb angle, the yaw angle, the bank angle and the throttle opening.

5. The cross-domain fault diagnosis method for drone clusters based on a multi-source domain deep migration network according to claim 1 is characterized in that: The implementation process of step (2) is as follows: Given N source domain input samples and a target domain input sample First, the samples of the source domain and the target domain are respectively extracted through the domain feature extractor to extract the fault features in each domain. Afterwards, the N source domain fault features are obtained Perform two operations: First, input the fault features of each domain into the respective source domain classifier to obtain the classification prediction results, i.e. The second is to divide the N source domain fault features into N pairs of feature combinations in the ratio of 1:N-1. These feature combinations are then input into the multi-source domain cross convolution Transformer F mcct (·) to obtain multi-source domain fusion features Next, the multi-source domain fusion features and target domain fault characteristics Input into the weight-sharing three-branch cross-domain alignment convolutional Transformer to perform feature alignment in a segment-by-segment manner to obtain high-level source domain fault features. and target domain fault characteristics and cross-domain related features Finally, the obtained features are input into the target domain classifier Obtain the target domain fault classification results.

6. The cross-domain fault diagnosis method for drone clusters based on a multi-source domain deep migration network according to claim 1 is characterized in that: The source domain and target domain classifiers in step (2) have the same network structure, including two Bayesian linear layers, ReLU activation function and Softmax function layer; the source domain fault features are established through the classifier High-level source domain fault characteristics and target domain fault characteristics and cross-domain related features The nonlinear mapping relationship between the fault classification label and the 7. The cross-domain fault diagnosis method for drone clusters based on a multi-source domain deep migration network according to claim 1 is characterized in that: The domain feature extractor in step (3) includes a common feature extractor F co (·) and domain-specific feature extractors The common feature extractor consists of a weight-sharing spatiotemporal Bayesian graph wavelet convolutional transformer; the domain-specific feature extractors consist of independent spatiotemporal Bayesian graph wavelet convolutional transformers. The spatiotemporal Bayesian graph wavelet convolution transformer includes multi-head spatiotemporal convolution attention, Bayesian linear layer, residual connection and layer normalization; the multi-head spatiotemporal convolution attention includes H parallel spatiotemporal convolution attentions, each of which extracts spatiotemporal coupling fault features from different scales; the input sample passes through the H spatiotemporal convolution attentions to obtain H spatiotemporal coupling fault features and splices them, and then passes through two layers of Bayesian linear layers with residual connection and layer normalization in sequence to obtain the final spatiotemporal coupling fault features; The spatiotemporal convolutional attention includes multi-layer multi-scale adaptive graph wavelet convolution, local and global scaling dot product attention, dilated causal convolution network layer and dimension adjustment layer; first, multi-layer multi-scale adaptive graph wavelet convolution processes the input sample Get the corresponding query vector Q h , key vector K h Sum value vector V h , and input into the local scaled dot product attention to obtain the spatial feature sequence Then the dimension adjustment layer adjusts the dimension of the spatial feature sequence to obtain where d model =Ne d ×N f ; Then, two parallel dilated causal convolutional network layers learn the spatial feature sequence The local context dependency relationship is used to obtain the corresponding query vector Q h ′, key vector K h '; Similarly, the dimension adjustment layer adjusts the input samples Dimensions obtained and use it as the value vector V h '; the query vector Q h ′, key vector K h ′ and value vector V h ′Input into the global scaled dot product attention to adaptively integrate the local temporal features of the input samples into the modeling of global dependencies to simultaneously model local and global spatiotemporal coupling features and extract spatiotemporal coupling fault features; The multi-layer multi-scale adaptive graph wavelet convolution is composed of K layers of multi-scale adaptive graph wavelet convolution connected in series; each multi-scale adaptive graph wavelet convolution includes adaptive graph wavelet convolutions of |S| scales, extracting spatial features from input samples at different scales; each adaptive graph wavelet convolution is implemented by graph wavelet transform; the graph wavelet transform uses a set of appropriate wavelet bases provided by wavelet transform to project the graph signal from the vertex domain to the spectral domain, and extract fault features in the spectral domain; The wavelet basis is defined as Where Ne is the number of drones, each wavelet basis ψ s,i Corresponding to the signal of the i-th UAV at scale s, the graph wavelet with scale parameter s on the i-th UAV can be written in the following matrix multiplication form: Ψ s =UG s AT T in, is the scaling matrix, λ l is the normalized graph Laplacian matrix The lth eigenvalue of , U is Eigenvector matrix of ; normalized graph Laplacian matrix It is obtained from the adjacency matrix A corresponding to the UAV cluster communication topology, which is defined as Among them I n is the identity matrix of the corresponding dimension, is the angle matrix The graph wavelet basis is approximated by Chebyshev polynomials as follows: c i,s =2e -s J i (-s) in, is the i-th order Chebyshev polynomial, Q is the number of Chebyshev polynomials, J i (-s) is the first kind of Bessel function; for input samples with graph structure The wavelet transform of the graph is The corresponding inverse transformation is The adaptive graph wavelet convolution with scale s is defined as: Among them, s is a graph wavelet basis with scale s, W s (i,j) are learnable parameters, is the output signal, σ(·) is the nonlinear activation function; the adaptive graph wavelet convolutions with different scales are connected in parallel and combined with the residual connection as an adaptive component to form a multi-scale adaptive graph wavelet convolution, and then multiple multi-scale adaptive graph wavelet convolutions are connected in series to form a multi-layer multi-scale adaptive graph wavelet convolution; in the domain d∈{d s1 ,...,d sN ,d t The corresponding output of the s-th scale of the K-th layer adaptive graph wavelet convolution in} is: in, is the initial input of the network, and are the graph wavelet basis and its inverse at scale s, Θ d,s is the diagonal matrix of the graph convolution kernel, and is a trainable parameter matrix, σ(·) is a nonlinear activation function; in the domain d∈{d s1 ,...,d sN ,d t In [15], the final output of the multi-layer multi-scale adaptive graph wavelet convolution is the average value of the output features of |S| parallel adaptive graph wavelet convolutions in the K-th layer of multi-scale adaptive graph wavelet convolution, which is defined as follows:

8. The cross-domain fault diagnosis method for UAV clusters based on a multi-source domain deep migration network according to claim 1 is characterized in that: The multi-source domain cross convolution Transformer in step (4) includes N multi-source cross convolution attention layers, residual connections, layer normalization, and linear layers; it is used to learn the relationship between different source domain features and extract multi-source domain fusion features; The multi-source cross convolution attention includes an expanded causal convolutional network layer, a repeated layer, and a scaled dot product attention; N source domain fault features are sequentially regarded as value vectors, and the remaining N-1 source domain fault features are regarded as key vectors and value vectors; specifically, the N source domain fault features are sequentially divided into N pairs of feature combinations according to the ratio of 1:N-1, namely Input N pairs of feature combinations into N multi-source cross convolution attentions respectively, It corresponds to The splicing features of the fault features of the remaining source domains except ; After the dilated causal convolutional network layer, the value vector V is obtained cn , After processing by the dilated causal convolutional network layer, the query vector Q is obtained cn , key vector K cn , the repeated layer passes through the value vector V cn Perform a copy operation to make its dimension the same as the corresponding query vector Q cn , key vector K cn consistent; The query vector Q cn , key vector K cn Sum value vector V cn Input scaled dot product attention, calculate attention, learn the relationship between each source domain feature and the other source domain features in a segment-by-segment manner, and obtain N multi-source domain interaction features {Attn c1 ,...,Attn cN }; Further, the N multi-source domain interaction features are spliced ​​and then further fused through two layers of linear layers with residual connections and layer normalization to obtain the final multi-source domain fusion feature 9. The cross-domain fault diagnosis method for drone clusters based on a multi-source domain deep migration network according to claim 1 is characterized in that: Step (6) The weight-sharing three-branch cross-domain alignment convolutional Transformer includes a source branch, a source-target branch, and a target branch; the multi-source domain fusion features and the target domain fault features are aligned in a segment-by-segment manner to extract domain-invariant features, thereby realizing the use of source domain fault features to improve the fault diagnosis performance in the target domain; the source branch, the source-target branch, and the target branch are all composed of a weight-sharing convolutional Transformer; the multi-source domain fusion features are aligned with the target domain fault features ... Input to the source branch, and processed by the dilated causal convolutional network layer to obtain the query vector Q s , key vector K s Sum value vector V s , and then through multi-head self-attention, two-layer residual connection, layer normalization and linear layer to obtain high-level source domain fault features Similarly, the target domain fault characteristics Input into the target branch, and processed by the dilated causal convolutional network layer to obtain the query vector Q t , key vector K t Sum value vector V t , and then through multi-head self-attention, two-layer residual connection, layer normalization and linear layer to obtain high-level target domain fault features The query vector Q in the source branch s and the key vector K of the target branch t Sum value vector V t At the same time, the multi-head cross-domain attention input is used to learn the relationship between the source domain and target domain features, and obtain cross-domain related features; Then the multi-source domain fusion features and target domain high-level features Splicing to obtain cross-domain combined features F cd , and input it into the residual connection and layer normalization in the source-target branch to fuse with the cross-domain alignment feature, and then pass through the linear layer and residual connection and layer normalization to obtain the cross-domain related features Next, the cross-domain related features obtained by the three branches are and They are respectively input into the weight-shared target domain classifier and output the corresponding fault classification value.

10. The cross-domain fault diagnosis method for drone clusters based on a multi-source domain deep migration network according to claim 1 is characterized in that: The hierarchical multi-metric domain difference loss L in step (9) is used to optimize each pair of dis The distribution difference of fault features at different levels in the target domain and source domain; this includes the low-level domain distribution loss and the high-level domain distribution loss obtained by weighted addition of the maximum average difference loss and the CORAL distance, which are used to measure the distribution difference between the fault features of the source domain samples after passing through the common feature extractor and the specific domain feature extractor and the corresponding target domain fault features; The classification loss Used to optimize the difference between all classifier output values ​​and fault classification label values; cross entropy loss is selected as the classification loss, and the total classification loss is in Represents the classification loss value corresponding to N source domain classifiers, The classification loss values ​​of the source branch and the target branch corresponding to the target domain classifier; The distillation loss L is used to improve the fault diagnosis performance in the target domain by utilizing the fault characteristics of the source domain; dtl Specifically, the weight-shared three-branch cross-domain alignment convolution Transformer outputs cross-domain related features of the source-target branch The fault classification probability distribution output in the target domain classifier is used as a soft label to supervise the training of the target branch through the following distillation loss to improve the classification performance of the diagnosis model on the target domain samples. Among them, q k and p k are the probabilities of the classification category k corresponding to the outputs of the source-target branch and the target branch, respectively; The comprehensive multi-factor loss function L is: total Among them, λ and β are weight parameters.

Images