Unmanned aerial vehicle cluster unseen fault diagnosis method based on semantic guidance attribute migration

Abstract

The invention discloses an unmanned aerial vehicle cluster unseen fault diagnosis method based on semantic guidance attribute migration, and the method comprises the steps: firstly, building an unmanned aerial vehicle cluster mathematical model considering a fault, and generating simulation fault data consistent with a fault attribute; then, corresponding fault features are extracted based on the simulation fault data and the actually measured fault data; on the basis, a hierarchical Bayesian semantic guidance attribute migration network based on a Bayesian causal attention-gate loop unit and an attribute embedding learning network is provided, and a highly matched corresponding relation between features and attribute semantics is gradually learned from available fault data. Attribute knowledge and characteristic representation related to attributes are learned, and the information is utilized to diagnose that no fault is found. According to the method, the problem that the unmanned aerial vehicle cluster does not have fault diagnosis in the absence of corresponding fault data is solved, and the accuracy and generalization of fault diagnosis are improved.

Description

Technical Field

[0001] The present invention relates to the field of fault diagnosis of unmanned aerial vehicle (UAV) swarms, and particularly to a method for diagnosing unseen faults in UAV swarms based on semantic-guided attribute transfer. Background Art

[0002] UAV swarms have been widely used in fields such as smart agriculture, transportation, and fire rescue. However, since UAVs operate in harsh environments for a long time, they will inevitably experience faults such as engine thrust loss, pitot tube offset, and communication interruption. Considering the interconnection between UAVs in a UAV swarm, when one or more UAVs experience faults, these faults will not only affect their own normal flight but also degrade the performance of neighboring communication UAVs, thus having an adverse impact on the entire UAV swarm and even posing a risk of causing the entire system to crash. Therefore, studying UAV swarm fault diagnosis technology to timely detect faults in UAV swarms to facilitate timely formulation of maintenance decisions and ensure reliable operation of swarm flight has important practical significance.

[0003] With the continuous progress of industrial data transmission and computing technologies, a large number of UAV swarm flight state monitoring signals have been collected, which has enabled the wide development of data-driven fault diagnosis methods. Data-driven fault diagnosis models can learn potential fault features from a large number of flight state monitoring signals and establish a non-linear mapping relationship between features and fault classification labels to achieve classification prediction of faults. 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, etc. However, these methods can only learn shallow features, have limited fault diagnosis performance, and require complex manual feature engineering, which is very time-consuming and laborious. With the rapid development of deep learning technologies with powerful deep representation learning capabilities, such as recurrent neural networks, convolutional neural networks, and their various variants such as long short-term memory and gated recurrent units, deep learning models have become more effective fault diagnosis technologies. Deep learning models can directly perform end-to-end fault diagnosis from raw flight state monitoring data without any manual feature engineering. These deep learning methods have achieved good diagnostic performance in fault diagnosis.

[0004] In practical engineering, an unmanned aerial vehicle (UAV) cluster contains a large number of UAVs, resulting in a very large variety of possible faults, especially compound fault forms. However, due to reasons such as flight cost and duration, it is impossible to collect flight data corresponding to all fault categories. In addition, some faults of the UAV cluster require the design of destructive experiments such as engine damage and rudder surface jamming, which will bring safety risks and huge losses to the UAV cluster system, making it difficult to obtain fault data under these conditions. In view of this, in practical engineering, there are some faults in the UAV cluster for which there is no corresponding fault data, and these faults can be regarded as unseen faults. However, existing data-driven fault diagnosis algorithms rely on a large amount of fault data for training to identify faults. It is very difficult for existing data-driven fault diagnosis algorithms to diagnose these unseen faults without fault data. Considering that zero-shot learning can enable the model to accurately predict data that has not been explicitly trained by leveraging the semantic relationship between known and unknown categories. Therefore, researching an unseen fault diagnosis method for UAV clusters based on zero-shot learning, diagnosing unseen faults based on existing fault data and fault knowledge, and leveraging the semantic relationship between fault modes has engineering significance. Summary of the Invention

[0005] Object of the Invention: The present invention proposes an unseen fault diagnosis method for UAV clusters based on semantic-guided attribute transfer, which solves the problem of unseen fault diagnosis in UAV clusters lacking corresponding fault data and improves the accuracy and generalization of fault diagnosis.

[0006] Technical Solution: An unseen fault diagnosis method for UAV clusters based on semantic-guided attribute transfer according to the present invention specifically includes the following steps:

[0007] Step 1: Collect the actual flight fault data of the UAV cluster and generate corresponding simulated flight fault data according to the mathematical model of the UAV cluster considering unseen faults. Define the semantic description of the fault attributes of the UAV cluster and divide it into four layers of semantic layers, and construct a training set and a test set;

[0008] Step 2, construct a Bayesian graph convolutional Transformer network, a hierarchical Bayesian semantic-guided attribute transfer network, and a Bayesian classifier, and initialize the network parameters; the hierarchical Bayesian semantic-guided attribute transfer network is composed of four learning modules with the same structure connected in series; each learning module includes a Bayesian causal attention-gated recurrent unit network, an attribute embedding learning network, and a Bayesian attribute-related feature prediction network;

[0009] Step 3, through the Monte Carlo sampling method, sample the network parameters W constructed in Step 2 from the approximate variational distribution Q θ (W) of the true posterior distribution P(WD) of all network parameters W;

[0010] Step 4: Input the training set samples into the Bayesian graph convolutional Transformer network to extract initial fault features;

[0011] Step 5: Input the initial fault features obtained in Step 4 and the fault attribute semantic description matrix into the hierarchical Bayesian semantic-guided attribute transfer network to obtain hierarchical attribute-related fault features and hierarchical reconstructed attribute semantics, and calculate the weighted attribute embedding learning loss L AEL ;

[0012] Step 6: Merge and splice the hierarchical attribute-related fault features and hierarchical reconstructed attribute semantics obtained in Step 5 in accordance with the unified fault attribute semantic description order to obtain fault features with attribute information;

[0013] Step 7: Input the fault features with attribute information obtained in Step 6 into the Bayesian fault classifier to output the fault classification prediction value and calculate the classification loss L with the fault classification label CLS and the corresponding distribution loss L DS ;

[0014] Step 8: Repeat Step 3 until the maximum number of Monte Carlo sampling times M max is reached, obtain the fault classification prediction distribution of the training samples, and take the average of all prediction values of each training sample as the final fault classification prediction value

[0015] Step 9: Calculate the average weighted attribute embedding learning loss, average classification loss, and average distribution loss, and add them according to different weight values to obtain the total loss value L TOT ;

[0016] Step 10: Update the parameters θ of the approximate variational distribution Q θ (W) by the backpropagation algorithm to minimize the total loss value L TOT ;

[0017] Step 11: Repeat Steps 3 to 10 to perform E max rounds of training on the network to obtain the optimal Bayesian graph convolutional Transformer network, hierarchical Bayesian semantic-guided attribute transfer network, and Bayesian classifier;

[0018] Step 12: Input the test set samples into the trained Bayesian graph convolutional Transformer network, hierarchical Bayesian semantic-guided attribute transfer network, and Bayesian classifier in sequence to output the fault classification prediction value;

[0019] Step 13: Repeat Step 12 until the maximum number of predictions P max, obtain the fault classification prediction distribution of the test sample, and take the average of all predicted values as the final fault classification value;

[0020] Step 14, calculate the fault classification prediction variance to evaluate the uncertainty of the classification result.

[0021] Furthermore, the four-layer semantic layer described in Step 1 includes a flight state change layer, a faulty drone layer, a faulty equipment layer, and a fault mode layer.

[0022] Furthermore, the hierarchical Bayesian semantic-guided attribute transfer network described in Step 2 is based on the initial fault feature F ea and the fault attribute semantic description matrix According to the unified fault attribute semantic description order, namely flight state change, single drone system, faulty equipment, and fault mode, the hierarchical attribute-related fault features and hierarchical reconstructed attribute semantics are obtained.

[0023] Furthermore, the Bayesian causal attention-gated recurrent unit network described in Step 2 includes Bayesian attribute attention and Bayesian attribute-related attention-gated recurrent units; it can realize the interaction between the reconstructed attribute semantics of the previous layer and the initial fault feature, and capture the context relationship of attribute semantics at different layers.

[0024] Furthermore, the Bayesian attribute attention can fuse the reconstructed attribute semantics of the previous layer Capture the context relationship of attribute semantics at different layers, and assist in learning the fault feature related to the attribute semantics of this layer; first, the reconstructed attribute semantics of the previous layer As the initial attribute semantic feature A a =[a 1 ,a 2 ,...,a n , successively pass through two layers of one-dimensional Bayesian convolutional neural networks to learn the local dependence relationship of attribute semantic features, and obtain the locally enhanced attribute semantic feature A q ; through a compatibility function f c (A q ,A a ) Measure the dependence relationship between the initial attribute semantic feature A a and the locally enhanced attribute semantic feature A q , and obtain the corresponding attention weight f s ; then, use the SoftMax function to normalize the attention weight f s into a probability distribution representing the attention weight; finally, use a splicing layer to integrate C v and the initial attribute feature A a And then pass through a Bayesian linear layer with a hyperbolic tangent function activation function tanh(·) to obtain the attribute attention vector

[0025] Further, the Bayesian attribute-related attention-gated recurrent unit is used for the interaction between the reconstructed attribute semantics of the previous layer and the initial fault features, and the fused fault features with the attribute information of the previous layer are obtained; the attribute attention vector and the initial fault feature F ea are simultaneously input into the Bayesian attribute-related attention-gated recurrent unit network; first, the update gate Z l is calculated; then, the attribute attention vector and the initial fault feature F ea are respectively transformed into the query vector q l , the key vector k l and the value vector v l through the Bayesian linear layer; then, the reset gate R l with the attribute-related attention mechanism is calculated; finally, the hidden state of the Bayesian attribute-related attention-gated recurrent unit network is updated by combining the update gate Z l and the reset gate R l and used as the fused fault feature F l .

[0026] Further, the attribute embedding learning network described in step 2 includes fault feature-attribute semantic consistency learning, attribute semantic enhancement, and attribute-guided attention; the fault features related to the hierarchical attribute semantics and the reconstructed attribute semantics are extracted.

[0027] Further, the fault feature-attribute semantic consistency learning includes two multi-layer perceptrons, namely MLP F2S and MLP S2F and the weighted attribute embedding learning loss; the perceptron MLP F2S maps the fused fault feature F l to the reconstructed attribute semantics The perceptron MLP S2F maps the true attribute semantics to the reconstructed fault feature F l r ; to achieve the fault feature-attribute semantic consistency learning, a weighted attribute embedding learning loss is designed, which includes an attribute semantic reconstruction loss and a fault feature reconstruction loss The attribute semantic reconstruction loss is composed of the 1-norm and the Euclidean distance and is defined as:

[0028]

[0029] where n a1 represents the number of elements of the attribute semantic vector; the fault feature reconstruction loss Composed of the maximum average difference d MMD (·), is defined as:

[0030]

[0031] The weighted attribute embedding learning loss is composed of the four - layer attribute semantic reconstruction loss and the fault feature reconstruction loss added together according to different weights, and is defined as:

[0032]

[0033] Among them, λ FR and λ AR are the corresponding weight parameters; a larger weight is set on the attribute semantic reconstruction loss than the fault feature reconstruction loss to establish the correspondence between fault features and attribute semantic information and enhance the subsequent learning of fault attribute information.

[0034] Furthermore, the enhancement of the attribute semantics is realized through attribute semantic embedding to explicitly integrate the attribute semantic information into the fault features; specifically, the reconstructed fault feature F l r and the fused fault feature F l are weighted and added; the attribute - guided attention further enhances the fault features related to the attribute semantics; specifically, the fault feature enhanced by the attribute semantics is used as the query vector q, the fused fault feature F l is regarded as the key vector K and the value vector V, and the two are interacted through the attention mechanism, which is defined as follows:

[0035]

[0036] Among them, d is the scale coefficient, and att is the attention score representing the correspondence between the fault features and the attribute semantics.

[0037] Furthermore, the implementation process of step 5 is as follows:

[0038] The initial fault feature F ea is sequentially input into the four learning modules in the hierarchical Bayesian semantic - guided attribute transfer network to sequentially learn the attribute - related fault features and reconstruct the attribute semantics corresponding to the flight state change layer, the fault UAV layer, the fault equipment layer, and the fault mode layer; the fault attribute semantic description matrix is further split into four semantic matrices, namely the flight state change layer the fault UAV layer the fault equipment layer and the fault mode layer The construction process is as follows: First, in the learning module of the flight state change layer, the initial fault feature F ea , the initialized reconstructed attribute semantics A 0 and the initialized fused fault feature F l0 are simultaneously input into the Bayesian causal attention-gated recurrent unit network to obtain the fused fault feature F l1 of this layer. Then, this fused fault feature and the known attribute matrix of the flight state change layer are simultaneously input into the attribute embedding learning network to extract the fault features related to the attribute semantics and the reconstructed attribute semantics Then, the fault features related to the attribute semantics are input into the Bayesian attribute-related feature prediction network to obtain the predicted fault features related to the attribute semantics Next, in the learning module of the faulty UAV layer, the initial fault feature F ea , the reconstructed attribute semantics of the flight state change layer and the fused fault feature F l1 of the flight state change layer are simultaneously input into the Bayesian causal attention-gated recurrent unit network to obtain the fused fault feature F l2 of this layer; then, this fused fault feature and the known attribute matrix of the faulty UAV layer are input into the attribute embedding learning network to extract the fault features related to the attribute semantics and the reconstructed attribute semantics Then, the fault features related to the attribute semantics are input into the Bayesian attribute-related feature prediction network to obtain the predicted fault features related to the attribute semantics The learning modules of the faulty equipment layer and the fault mode layer respectively obtain the hierarchical fault features and hierarchical reconstructed attribute semantics of the corresponding layers; during testing, the input fault attribute semantic description matrix of each layer is replaced by the reconstructed attribute semantics of each layer.

[0039] Beneficial effects: Compared with the prior art, the beneficial effects of the present invention are as follows:

[0040] 1. The present invention builds a mathematical model of an unmanned aerial vehicle cluster considering unseen faults, generates simulation fault data for corresponding unseen faults, improves the matching degree between the generated sample data and the corresponding attribute semantic information, enhances the effect of attribute transfer learning, reduces attribute negative transfer, and effectively solves the domain shift problem;

[0041] 2. The present invention designs a hierarchical Bayesian semantic-guided attribute transfer network, which divides the fault attributes of the UAV swarm into four layers, gradually learns the corresponding relationship with a high degree of matching between features and attribute semantics from the available fault data, learns the context dependence of attribute semantics at different layers, so as to learn attribute knowledge and extract feature representations related to attributes, and then uses this information to diagnose unseen faults, improving the attribute transfer effect and the performance of unseen fault diagnosis;

[0042] 3. The present invention extends the designed network to a Bayesian deep learning framework. Using the Bayesian deep learning framework, it simultaneously considers the uncertainties of fault classification and fault attribute prediction, improving the generalization of the diagnostic model. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a flowchart of the present invention;

[0044] Figure 2 is a schematic diagram of the training and testing structure of the fault diagnosis network proposed by the present invention;

[0045] Figure 3 is a schematic diagram of the hierarchical Bayesian semantic-guided attribute transfer network structure;

[0046] Figure 4 is a schematic diagram of the Bayesian causal attention-gated recurrent unit network structure;

[0047] Figure 5 is a fault attribute semantic description matrix;

[0048] Figure 6 The flight formation shape and communication topology structure diagram of the UAV swarm;

[0049] Figure 7 is the fault diagnosis result of the UAV swarm based on the method of the present invention under four unseen fault diagnosis tasks. DETAILED DESCRIPTION OF THE INVENTION

[0050] The following further describes the present invention in detail with reference to the accompanying drawings.

[0051] As Figure 1 shown, the present invention proposes a method for diagnosing unseen faults of a UAV swarm based on semantic-guided attribute transfer, including the following steps:

[0052] Step 1, collect the actual flight fault data of the UAV swarm and generate corresponding simulated flight fault data according to the mathematical model under unseen faults, define the fault attribute semantics description of the UAV swarm and divide it into four semantic layers, and construct a training set and a testing set.

[0053] The UAV cluster can be a fixed-wing UAV cluster and a quadrotor UAV cluster. No-fault seen means the fault of the UAV cluster without corresponding actual flight fault data. Define the semantic description of the fault attributes of the UAV cluster and divide it into four semantic layers, specifically including: flight state change layer, faulty UAV layer, faulty equipment layer, and fault mode layer. The mathematical model is the kinematic and dynamic equations of the UAV cluster considering faults and communicates according to an undirected communication topology structure.

[0054] Step 2: Construct a Bayesian graph convolutional Transformer network, a hierarchical Bayesian semantic-guided attribute transfer network, and a Bayesian classifier, and initialize the network parameters.

[0055] As Figure 2 shown, the Bayesian graph convolutional Transformer network includes a Bayesian graph neural network and a Bayesian convolutional Transformer; the input sample first passes through the Bayesian graph neural network to extract spatial features; then the obtained spatial features are input into the Bayesian convolutional Transformer to further extract time-series related features, so as to obtain spatio-temporal fault features and use them as the initial fault feature F ea . The Bayesian graph neural network includes a graph convolutional neural network and a Bayesian linear layer. The Bayesian convolutional Transformer includes a Bayesian convolutional neural network, a Transformer encoder, and a Bayesian linear layer.

[0056] The network initialization process is as follows: randomly sample random values from the standard normal distribution N(0,1), and then randomly initialize all model parameters to these sampled random values to promote the diversity and learning ability of the network.

[0057] As Figure 3 shown, the hierarchical Bayesian semantic-guided attribute transfer network, according to the initial fault feature F ea and the fault attribute semantic description matrix in accordance with the unified fault attribute semantic description order, namely flight state change, single UAV system, faulty equipment, and fault mode, obtains the hierarchical attribute-related fault features and hierarchical reconstructed attribute semantics; this network is composed of four learning modules with the same structure connected in series. As Figure 3 shown, the learning module includes a Bayesian causal attention-gated recurrent unit network, an attribute embedding learning network, and a Bayesian attribute-related feature prediction network.

[0058] The hierarchical Bayesian semantic-guided attribute transfer network, the structure and construction process include: the initial fault feature F eaAre successively input into four learning modules in the hierarchical Bayesian semantic-guided attribute transfer network, and successively learn the attribute-related fault features and reconstructed attribute semantics corresponding to the flight state change layer, the faulty UAV layer, the faulty equipment layer, and the fault mode layer; the fault attribute semantic description matrix is further split into four semantic matrices, namely the flight state change layer the faulty UAV layer the faulty equipment layer and the fault mode layer Construction process: First, in the learning module of the flight state change layer, the initial fault feature F ea , the initialized reconstructed attribute semantics A 0 and the initialized fused fault feature F l0 are simultaneously input into the Bayesian causal attention-gated recurrent unit network to obtain the fused fault feature F l1 of this layer. Then, the fused fault feature and the known attribute matrix of the flight state change layer are simultaneously input into the attribute embedding learning network to extract the fault features related to the attribute semantics and the reconstructed attribute semantics The fault features related to the attribute semantics are then input into the Bayesian attribute-related feature prediction network to obtain the predicted fault features related to the attribute semantics Next, in the learning module of the faulty UAV layer, the initial fault feature F ea , the reconstructed attribute semantics of the flight state change layer and the fused fault feature F l1 of the flight state change layer are simultaneously input into the Bayesian causal attention-gated recurrent unit network to obtain the fused fault feature F l2 of this layer; then, the fused fault feature and the known attribute matrix of the faulty UAV layer are input into the attribute embedding learning network to extract the fault features related to the attribute semantics and the reconstructed attribute semantics The fault features related to the attribute semantics are then input into the Bayesian attribute-related feature prediction network to obtain the predicted fault features related to the attribute semantics Similarly, the learning modules of the faulty equipment layer and the fault mode layer respectively obtain the hierarchical fault features and hierarchical reconstructed attribute semantics of the corresponding layers. During testing, the fault attribute semantic description matrix input for each layer is replaced by the reconstructed attribute semantics of each layer.

[0059] As Figure 4 shown, the Bayesian causal attention-gated recurrent unit network includes Bayesian attribute attention and Bayesian attribute-related attention-gated recurrent units; it can realize the interaction between the reconstructed attribute semantics of the previous layer and the initial fault features, and capture the context relationship of the attribute semantics of different layers.

[0060] Bayesian attribute attention, which can fuse the reconstructed attribute semantics of the previous layer Capture the context relationship of attribute semantics in different layers to assist in learning the fault features related to the attribute semantics of this layer; first, the reconstructed attribute semantics of the previous layer As the initial attribute semantic feature A a =[a 1 , a 2 ,..., a n , and successively pass through two layers of one-dimensional Bayesian convolutional neural networks to learn the local dependence relationship of the attribute semantic features, and obtain the locally enhanced attribute semantic feature A q ; through a compatibility function f c (A q , A a ) to measure the dependence relationship between the initial attribute semantic feature A a and the locally enhanced attribute semantic feature A q , and obtain the corresponding attention weight f s ; then, use the softmax function to normalize the attention weight f s into a probability distribution representing the attention weight. The above calculation process is as follows:

[0061] A q =W a2 *(W a1 *A a +b a1 )+b a2

[0062]

[0063] P(m|A q , A a ) = softmax(f s )

[0064] where * represents the one-dimensional Bayesian convolutional neural network operation, W a1 and W a2 are the weights of the Bayesian convolutional neural network, b a1 and b a2 are the biases of the Bayesian convolutional neural network, Then, the initial attribute feature A a can be summarized into an attribute context representation vector C v , that is:

[0065]

[0066] Finally, use a concatenation layer for C v and the initial attribute feature Aa Integrate them, and then pass through a Bayesian linear layer with the hyperbolic tangent function activation function tanh(·) to obtain the attribute attention vector That is:

[0067]

[0068] where W a is the weight of the Bayesian linear layer.

[0069] The Bayesian attribute-related attention-gated recurrent unit is used for the interaction between the reconstructed attribute semantics of the previous layer and the initial fault features, and obtains the fused fault features with the attribute information of the previous layer; the attribute attention vector and the initial fault feature F ea are simultaneously input into the Bayesian attribute-related attention-gated recurrent unit network.

[0070] First, calculate the update gate Z l :

[0071]

[0072] where W z1 and W z2 are the weights of the Bayesian linear layer, and σ(·) is the sigmoid activation function.

[0073] Next, the attribute attention vector and the initial fault feature F ea are respectively transformed into the query vector q l , the key vector k l and the value vector v l through the Bayesian linear layer, that is:

[0074] k l =W k F ea , v l =W v F ea

[0075] where W q , W k , and W v are the weights of the Bayesian linear layer corresponding to the query vector, the key vector, and the value vector respectively; calculate the reset gate R l with the attribute-related attention mechanism according to the following formula:

[0076] R l =att l v l

[0077]

[0078] Among them, att l is the attention weight related to attributes, and d k represents the dimension of the key vector.

[0079] Finally, combining the update gate Z l and the reset gate R l update the hidden state of the Bayesian attribute-related attention-gated recurrent unit network and use it as the fused fault feature F l , and the calculation is as follows:

[0080]

[0081] As Figure 3 shown, the attribute embedding learning network includes fault feature-attribute semantic consistency learning, attribute semantic enhancement, and attribute-guided attention; it can extract fault features related to hierarchical attribute semantics and reconstruct attribute semantics.

[0082] Fault feature-attribute semantic consistency learning includes two multi-layer perceptrons, namely MLP F2S and MLP S2F , as well as the weighted attribute embedding learning loss; the perceptron MLP F2S maps the fused fault feature F l to the reconstructed attribute semantics The perceptron MLP S2F maps the true attribute semantics to the reconstructed fault feature F l r ; to achieve fault feature-attribute semantic consistency learning, a weighted attribute embedding learning loss is designed, which includes an attribute semantic reconstruction loss and a fault feature reconstruction loss The attribute semantic reconstruction loss consists of the 1-norm and the Euclidean distance and is defined as:

[0083]

[0084] Among them, n a1 represents the number of elements of the attribute semantic vector; the fault feature reconstruction loss consists of the maximum mean discrepancy d MMD (·) and is defined as:

[0085]

[0086] For two given features and the maximum mean discrepancy d MMD (·) can be obtained by the unbiased estimate of the following formula:

[0087]

[0088] Among them, \(H\) is a reproducing kernel Hilbert space with a characteristic kernel \(K\), and \(\varphi(\cdot)\) represents a mapping function that can map the original \(x\) to the reproducing kernel Hilbert space.

[0089] The weighted attribute embedding learning loss is composed of a four - layer attribute semantic reconstruction loss and a fault feature reconstruction loss added together according to different weights, and is defined as:

[0090]

[0091] Among them, \(\lambda\) FR and \(\lambda\) AR are the corresponding weight parameters. A larger weight is set on the attribute semantic reconstruction loss than on the fault feature reconstruction loss to establish the correspondence between fault features and attribute semantic information, and enhance the subsequent learning of fault attribute information.

[0092] Attribute semantic enhancement is achieved through attribute semantic embedding to explicitly incorporate attribute semantic information into fault features; specifically, the reconstructed fault feature \(F\) l r and the fused fault feature \(F\) l are weighted and added according to the following formula, that is:

[0093]

[0094] Among them, \(\gamma\) is a combination coefficient, is the fault feature with enhanced attribute semantics;

[0095] Attribute - guided attention further enhances the fault features related to attribute semantics; specifically, the fault feature with enhanced attribute semantics is used as the query vector \(q\), the fused fault feature \(F\) l is regarded as the key vector \(K\) and the value vector \(V\), and the two are interacted through the attention mechanism, which is defined as follows:

[0096]

[0097] Among them, \(d\) is a scale coefficient, and \(att\) is the attention score representing the correspondence between fault features and attribute semantics. Through this attention, the fault feature segments most relevant to attribute semantics can be effectively located, thus obtaining the fault features related to attribute semantics

[0098] The Bayesian attribute - related feature prediction network is composed of two - layer Bayesian linear layers with LeakyReLU activation functions.

[0099] The Bayesian classifier consists of two Bayesian linear layers with ReLu activation functions and a SoftMax function layer. The construction process includes:

[0100] Input the fault features with attribute information Pass through two Bayesian linear layers with ReLU activation functions and a SoftMax function layer in sequence, and output the fault classification prediction value The formula is described as follows:

[0101]

[0102] Among them, W c0 , W c1 and b c0 , b c1 are the weights and biases of the two Bayesian linear layers respectively.

[0103] Step 3, through the Monte Carlo sampling method, sample the network parameters W constructed in Step 2 from the approximate variational distribution Q θ (W) of the true posterior distribution P(W|D) of the network parameters W.

[0104] Step 4, input the training set samples into the Bayesian graph convolutional Transformer network to extract the initial fault features.

[0105] Step 5, input the initial fault features obtained in Step 4 and the fault attribute semantic description matrix into the hierarchical Bayesian semantic-guided attribute transfer network to obtain the hierarchical attribute-related fault features and the hierarchical reconstructed attribute semantics, and calculate the weighted attribute embedding learning loss L AEL .

[0106] Step 6, merge and splice the hierarchical attribute-related fault features and the hierarchical reconstructed attribute semantics obtained in Step 5 in the order of the unified fault attribute semantic description to obtain the fault features with attribute information.

[0107] The hierarchical attribute-related fault features and the hierarchical reconstructed attribute semantics are merged and spliced in the order of the attribute semantic descriptions of the flight state change layer, the faulty UAV layer, the faulty equipment layer, and the fault mode layer in sequence to obtain the fault features with attribute information

[0108] Step 7, input the fault features with attribute information obtained in Step 6 into the Bayesian fault classifier, output the fault classification prediction value and calculate the classification loss L with the fault classification label CLS and the corresponding distribution loss L DS .

[0109] Classification loss L CLS The cross-entropy loss is adopted and defined as where N n is the number of samples, y n is the true fault classification label, and is the predicted value of fault classification. The distribution loss L DS is used to optimize the Kullback-Leibler divergence KL(Q θ (W) between the approximate variational distribution Q θ (W) of the model parameter W and its true posterior distribution P(W|D); the approximate solution of KL(Q θ (W)||P(W|D)) is taken as the distribution loss value L DS That is, L DS ≈ logQ θ (W s ) - logP(W s ) - log(P(Y|X,W s ))), where W s represents the model parameter sampled s times from the approximate variational distribution Q θ (W), and θ is the parameter of the approximate variational distribution.

[0110] Step 8: Repeat Step 3 until the maximum number of Monte Carlo samplings 1000 is reached, obtain the fault classification prediction distribution of the training samples, and take the average of all predicted values of each training sample as the final fault classification prediction value

[0111] Step 9: Calculate the average weighted attribute embedding learning loss, average classification loss, average distribution loss, and add them according to different weight values to obtain the total loss value L TOT .

[0112] The total loss value is obtained by adding the average classification loss the average distribution loss and the average weighted attribute embedding learning loss according to different weight values, that is where the weight parameters are set to 0.7, 0.3, and 0.8 respectively.

[0113] Step 10: Update the parameter θ of the approximate variational distribution Q θ (W) through the backpropagation algorithm to minimize the total loss value L TOT .

[0114] The approximate variational distribution Q θ (W) is set to the Gaussian distribution N(wμ,σ 2), where the learnable distribution parameter θ = {μ, σ}; then the reparameterization trick is adopted to ensure the gradient information of the learned network parameters W = {w, b}, and the calculation formula is

[0115] Step 11, repeat Steps 3 to 10, train the network for 200 rounds to obtain the optimal Bayesian graph convolutional Transformer network, hierarchical Bayesian semantic-guided attribute transfer network, and Bayesian classifier.

[0116] Step 12, sequentially input the test set samples into the trained Bayesian graph convolutional Transformer network, hierarchical Bayesian semantic-guided attribute transfer network, and Bayesian classifier, and output the fault classification prediction values;

[0117] Step 13, repeat Step 12 until the maximum number of predictions reaches 1000 to obtain the fault classification prediction distribution of the test samples, and take the average of all prediction values as the final fault classification value;

[0118] Specifically, the final fault classification value can be calculated by the following formula:

[0119]

[0120] where, is the final fault classification value, W is the network parameter, {x * , Y *} is the sample to be predicted, and θ is the parameter of the approximate variational distribution Q θ (W).

[0121] Step 14, calculate the fault classification prediction variance to evaluate the uncertainty of the classification result.

[0122] The prediction variance V p is defined as:

[0123]

[0124] where, is the p-th classification prediction value, σ 2 is the variance of the P max -th classification prediction result, and Softplus(·) is the softplus activation function to ensure the effectiveness of the variance. A higher prediction variance V p indicates that the p max -th random prediction result has a larger degree of dispersion and greater uncertainty. On the contrary, a lower prediction variance V p indicates that the p max -th random prediction result is more stable, has less uncertainty, and the prediction result is more reliable.

[0125] To verify the effectiveness and superiority of the present invention in the unseen fault diagnosis of UAV clusters, flight data corresponding to different faults occurring in the UAV cluster under different faults is collected through a UAV cluster semi-physical flight platform. The UAV cluster semi-physical flight platform mainly 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 Raspberry Pi cluster simulates a UAV. One of the Raspberry Pis serves as the flight controller system, and the other Raspberry Pi is used to simulate the dynamics, actuators, sensors, and environment of the UAV. The Raspberry Pi sends information to the router through the switch, and finally sends it to the database server and the neighbor calculation server for information interaction. The database server is connected to the router and is used to store the status information of the UAV cluster. The neighbor calculation server accesses the database, calculates the neighbor UAV information of each UAV according to the communication topology structure, and sends data to the flight controller. The flight controller mainly calculates and generates control commands based on the received sensor data and neighbor UAV information, and sends them to the dynamics environment part to control the UAV flight cluster. The data analysis terminal (computer) obtains the status information of the UAV cluster stored in the database and displays the flight situation of the cluster UAVs through the situation display terminal software.

[0126] Taking a cluster system composed of 18 fixed-wing UAVs as an example, the UAV cluster flies according to the formation shape and communication topology structure as Figure 5 shown, and the fault diagnosis algorithm of the present invention is only deployed on the 8th and 11th UAVs. All diagnosis results are sent to the ground station for summary to facilitate the formulation of fault maintenance and repair strategies. Actual flight fault data similar to actual engineering is collected on the semi-physical flight platform of the UAV cluster. Simulation flight fault data is obtained according to the mathematical model of the UAV cluster under unseen faults in MATLAB. The flight status monitoring data shown in Table 1 is collected.

[0127] Table 1 Flight Status Monitoring Data

[0128]

[0129]

[0130] Taking the 8th UAV as an example, |Δ(Δx u )(u = 1,..., Ne), N e = 8 represents the absolute value of the change in the relative distance in the x-axis direction between all neighboring communication UAVs and the 8th UAV. |Δ(Δy u )(u = 1,..., Ne) is the same. V xi , (i = 1,..., N), N = 9 represents the speed in the x-axis direction between the UAVs communicating with the 8th UAV and the 8th UAV. The rest is the same. 15 types of fault types as shown in Table 2 are set.

[0131] Table 2 Fault Categories and Corresponding Tags

[0132]

[0133]

[0134] Specifically, Fault-1 to Fault-7 occurred on UAV No. 8. Fault-8 is the pitot tube offset fault that occurred on UAV No. 2. Fault-9 is the concurrent fault of two UAVs, namely the pitot tube offset fault on UAV No. 7 and the engine thrust loss fault on UAV No. 15. Fault-10 is the concurrent fault of three UAVs, namely the pitot tube offset fault on UAV No. 7, the engine thrust loss fault on UAV No. 8, and the rudder surface offset fault on UAV No. 14. Fault-11 is the communication interruption fault between two UAVs, that is, the communication interruption fault between UAV No. 8 and UAV No. 1. Fault-12 is the communication interruption fault of multiple UAVs, that is, UAV No. 8 simultaneously has communication interruption faults with UAV No. 3 and UAV No. 9. Fault-13 is the engine shutdown of UAV No. 13; Fault-14 is the complete jamming of the rudder surface of UAV No. 14. Fault-15 is the complete damage of the communication equipment of UAV No. 8. Taking the fault diagnosis algorithm deployed on UAV No. 8 as an example, Table 3 gives the semantic descriptions of 23 fault attributes, which are divided into four semantic layers, namely the flight state change layer, the fault UAV layer, the fault equipment layer, and the fault mode layer.

[0135] Table 3 Semantic Descriptions of Fault Attributes

[0136]

[0137]

[0138] Specifically, attributes Att#1-9 describe the fault attributes of the fault UAV layer, indicating which UAV the fault occurred on, such as a fault occurring on UAV No. 8; attributes Att#10-13 describe the attributes of the fault equipment layer, indicating which component the fault occurred on, such as the pitot tube; attributes Att#14-20 describe the attributes of the fault mode layer, indicating what specific type of fault it is, such as an offset fault; attributes Att#21-23 describe the fault attributes of the flight state change layer, indicating the impact of the fault on the flight, such as a decrease in speed, a decrease in altitude, a deviation in direction, etc.

[0139] According to Table 2, Table 3, and the UAV swarm mechanism model, Figure 6A corresponding semantic description matrix of fault attributes is given. Each fault is described by 23 physical attributes. The value 1 indicates the presence of this physical attribute in the fault, and the value 0 indicates the absence of this physical attribute in the fault. Taking the fault Fault#4 as an example, it is the engine thrust loss fault of UAV No. 8, which has 4 physical attributes: Att#5 (occurring on UAV No. 8), Att#10 (engine fault), Att#14 (engine thrust loss fault), Att#21 (speed decrease), Att#22 (altitude decrease).

[0140] To verify the diagnostic performance of the present invention for unseen faults in a UAV swarm, four unseen fault diagnosis tasks for the UAV swarm are set up, as shown in Table 4.

[0141] Table 4 Four unseen fault diagnosis tasks for the UAV swarm

[0142]

[0143] In each unseen fault diagnosis task, 11 faults are selected as training faults and 4 faults are selected as unseen faults. To evaluate the fault diagnosis performance of different models, three evaluation indicators, namely accuracy, recall rate, and F1-score, are selected and defined as follows where TP, TN, FP, and FN are the number of correctly classified positive samples, the number of correctly classified negative samples, the number of misclassified negative samples, and the number of misclassified positive samples, respectively.

[0144] Table 5 Fault diagnosis performance under different tasks

[0145] Task Accuracy Recall F1 Score Uncertainty a 93.84％ 93.93％ 93.88％ 0.010 b 92.46％ 92.58％ 92.50％ 0.005 c 91.43％ 91.37％ 91.37％ 0.008 d 93.21％ 93.05％ 93.00％ 0.011

[0146] As Figure 7 shown in and Table 5, the fault diagnosis results of the UAV swarm based on the method of the present invention under four unseen fault diagnosis tasks. According to the experimental results, it can be seen that the method of the present invention can obtain accurate fault diagnosis results under different unseen fault diagnosis tasks, demonstrating the effectiveness of the method. In addition, in all tasks, the diagnostic uncertainty remains at a low level without large fluctuations, verifying the reliability of the fault diagnosis results.

[0147] The embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those of ordinary skill in the art, various changes can be made without departing from the gist of the present invention.

Claims

1. A method for unmanned aerial vehicle cluster unseen fault diagnosis based on semantically guided attribute migration, characterized in that: The following steps are involved: Step 1: Collect the actual flight fault data of the UAV cluster and generate the corresponding simulated flight fault data based on the mathematical model of the UAV cluster considering the unseen faults, define the semantic description of the fault attributes of the UAV cluster and divide it into four semantic layers, and construct the training set and test set; Step 2, constructing a Bayesian graph convolutional Transformer network, a hierarchical Bayesian semantic-guided attribute transfer network and a Bayesian classifier, and initializing network parameters; the hierarchical Bayesian semantic-guided attribute transfer network is composed of four learning modules with the same structure connected in series; each learning module includes a Bayesian causal attention-gate recurrent unit network, an attribute embedding learning network and a Bayesian attribute-related feature prediction network; Step 3: Using the Monte Carlo sampling method, we can obtain the approximate variational distribution Q of the true posterior distribution P(WD) of all network parameters W. θ (W) is sampled to obtain the network parameter W constructed in step 2; Step 4: Input the training set samples into the Bayesian graph convolutional Transformer network to extract the initial fault features; Step 5: Input the initial fault features and fault attribute semantic description matrix obtained in step 4 into the hierarchical Bayesian semantic guided attribute migration network to obtain hierarchical attribute related fault features and hierarchical reconstruction attribute semantics, and calculate the weighted attribute embedding learning loss L AEL ; Step 6: merge and concatenate the hierarchical attribute-related fault features and hierarchical reconstruction attribute semantics obtained in step 5 according to a unified fault attribute semantic description order to obtain a fault feature with attribute information; Step 7: Input the fault features with attribute information obtained in step 6 into the Bayesian fault classifier and output the fault classification prediction value. And calculate the classification loss L with the fault classification label CLS And the corresponding distribution loss L DS ; Step 8: Repeat step 3 until the maximum number of Monte Carlo sampling times M is reached. max , obtain the fault classification prediction distribution of the training samples, and take the average of all prediction values ​​of each training sample as the final fault classification prediction value Step 9: Calculate the average weighted attribute embedding learning loss, average classification loss, and average distribution loss, and add them up according to different weight values ​​to get the total loss value L TOT ; Step 10: Update the parameters Q of the approximate variational distribution through the back-propagation algorithm θ The parameters θ of (W) are used to minimize the total loss value L TOT ; Step 11: Repeat steps 3 to 10 to perform E max After rounds of training, we obtained the optimal Bayesian graph convolutional Transformer network, hierarchical Bayesian semantic guided attribute transfer network and Bayesian classifier. Step 12: Input the test set samples into the trained Bayesian graph convolutional Transformer network, hierarchical Bayesian semantic guided attribute transfer network and Bayesian classifier in sequence, and output the fault classification prediction value; Step 13: Repeat step 12 until the maximum number of predictions P is reached. max , obtain the fault classification prediction distribution of the test sample, and take the average of all predicted values ​​as the final fault classification value; Step 14, calculate the fault classification prediction variance and evaluate the uncertainty of the classification result.

2. The method for unmanned aerial vehicle cluster unseen fault diagnosis based on semantically guided attribute migration according to claim 1 is characterized in that: The four semantic layers described in step 1 include a flight state change layer, a faulty drone layer, a faulty device layer, and a fault mode layer.

3. The method for unmanned aerial vehicle cluster unseen fault diagnosis based on semantically guided attribute migration according to claim 1 is characterized in that: Step 2: The hierarchical Bayesian semantic guided attribute migration network is based on the initial fault feature F ea and fault attribute semantic description matrix According to the unified fault attribute semantic description order, i.e., flight state change, single UAV system, faulty equipment and fault mode, the hierarchical attribute-related fault features and hierarchical reconstruction attribute semantics are obtained.

4. The method for unmanned aerial vehicle cluster unseen fault diagnosis based on semantically guided attribute migration according to claim 1 is characterized in that: The Bayesian causal attention-gated recurrent unit network described in step 2 includes Bayesian attribute attention and Bayesian attribute correlation attention-gated recurrent units; it can realize the interaction between the attribute semantics of the previous layer reconstruction and the initial fault features, and capture the contextual relationship between the attribute semantics of different layers.

5. The method for diagnosing unseen faults of drone clusters based on semantically guided attribute migration according to claim 4 is characterized in that: The Bayesian attribute attention can integrate the previous layer to reconstruct the attribute semantics Capture the contextual relationship of attribute semantics at different layers to assist in learning the fault features related to the attribute semantics of this layer; first, reconstruct the attribute semantics of the previous layer As the initial attribute semantic feature A a =[a1,a2,...,a n ], and then pass through two layers of one-dimensional Bayesian convolutional neural networks to learn the local dependency of attribute semantic features and obtain the locally enhanced attribute semantic feature A q ; through a compatibility function f c (A q ,A a ) measures the initial attribute semantic features A a and locally enhanced attribute semantic features A q The dependency relationship between them is used to obtain the corresponding attention weight f s ; Then, the attention weight f is calculated using the SoftMax function s Normalized to a probability distribution representing the attention weight; finally, a concatenation layer is used to v and the initial attribute feature A a After integration, it passes through a Bayesian linear layer with a hyperbolic tangent activation function tanh(·) to obtain the attribute attention vector 6. The method for diagnosing unseen faults of drone clusters based on semantically guided attribute migration according to claim 4 is characterized in that: The Bayesian attribute-related attention-gate recurrent unit is used to interact between the attribute semantics of the previous layer reconstruction and the initial fault feature to obtain a fused fault feature with the attribute information of the previous layer; Attribute Attention Vector and initial fault characteristics F ea At the same time, the Bayesian attribute-related attention-gate recurrent unit network is input; first, the update gate Z is calculated l ; Next, the attribute attention vector and initial fault characteristics F ea They are transformed into query vectors q through the Bayesian linear layer respectively l , key vector k l Sum value vector v l ; Then calculate the reset gate R with attribute-dependent attention mechanism l ; Finally, combined with the update gate Z l and reset gate R l Update the hidden state of the Bayesian attribute-related attention-gate recurrent unit network and use it as the fused fault feature F l .

7. The method for diagnosing unseen faults of drone clusters based on semantically guided attribute migration according to claim 1 is characterized in that: The attribute embedding learning network described in step 2 includes fault feature-attribute semantic consistency learning, attribute semantic enhancement and attribute guided attention; extracting fault features related to hierarchical attribute semantics and reconstructing attribute semantics.

8. The method for diagnosing unseen faults of drone clusters based on semantically guided attribute migration according to claim 7 is characterized in that: The fault feature-attribute semantic consistency learning includes two multi-layer perceptrons, namely MLP F2S and MLP S2F and weighted attribute embedding learning loss; Perceptron MLP F2S The fused fault feature F l Mapping to reconstruct attribute semantics A l p ; Perceptron MLP S2F The real attribute semantics A l r Mapping to reconstruct fault feature F l r ; To achieve fault feature-attribute semantic consistency learning, a weighted attribute embedding learning loss is designed, which includes an attribute semantic reconstruction loss and a fault signature reconstruction loss Attribute semantic reconstruction loss It consists of the 1-norm and the Euclidean distance, defined as: Among them, n a1 Indicates the number of attribute semantic vector elements; fault feature reconstruction loss By the maximum average difference d MMD (·) composition, defined as: The weighted attribute embedding learning loss is composed of four layers of attribute semantic reconstruction loss and fault feature reconstruction loss It is obtained by adding different weights and is defined as: Among them, λ FR and λ AR is the corresponding weight parameter; in the attribute semantic reconstruction loss The upper setting ratio fault feature reconstruction loss A larger weight is used to establish the correspondence between fault features and attribute semantic information, thus enhancing the subsequent learning of fault attribute information.

9. The method for diagnosing unseen faults of drone clusters based on semantically guided attribute migration according to claim 7 is characterized in that: The attribute semantic enhancement is achieved by embedding attribute semantics to explicitly incorporate attribute semantic information into fault features; Specifically, the fault feature F is reconstructed l r and fusion fault feature F l The attribute guides attention and further enhances the fault features related to the attribute semantics; specifically, the fault features enhanced by the attribute semantics As the query vector q, the fused fault feature F l As key vector K and value vector V, the two are interacted through the attention mechanism, which is defined as follows: Where d is the scale coefficient and att is the attention score representing the correspondence between fault features and attribute semantics.

10. The method for diagnosing unseen faults of drone clusters based on semantically guided attribute migration according to claim 1 is characterized in that: The implementation process of step 5 is as follows: Initial fault feature F ea The four learning modules in the hierarchical Bayesian semantic guided attribute transfer network are input in sequence to learn the attribute-related fault features and reconstruction attribute semantics corresponding to the flight state change layer, faulty drone layer, faulty equipment layer and fault mode layer; the fault attribute semantic description matrix Further split into four semantic matrices, namely the flight state change layer Fault Drone Layer Faulty device layer and failure mode layers The construction process is as follows: First, in the learning module of the flight state change layer, the initial fault feature F ea , the initialized reconstructed attribute semantics A0 and the initialized fusion fault feature F l0 At the same time, it is input into the Bayesian causal attention-gate recurrent unit network to obtain the fused fault feature F of this layer l1 , and then the fusion fault characteristics and the known flight state change layer attribute matrix At the same time, it is input into the attribute embedding learning network to extract fault features related to attribute semantics. and reconstruct attribute semantics Then the fault features related to the attribute semantics are input into the Bayesian attribute-related feature prediction network to obtain the predicted fault features related to the attribute semantics. Next, in the learning module of the faulty drone layer, the initial fault feature F ea , the reconstruction attribute semantics of the flight state change layer and the fusion fault feature F of the flight state change layer l1 At the same time, it is input into the Bayesian causal attention-gate recurrent unit network to obtain the fused fault feature F of this layer l2 ; Then the fusion fault features and the known fault drone layer attribute matrix Input into the attribute embedding learning network to extract fault features related to attribute semantics and reconstruct attribute semantics Then, the fault features related to the attribute semantics are input into the Bayesian attribute-related feature prediction network to obtain the fault features related to the attribute semantics of the prediction of this layer. The learning modules of the fault device layer and the fault mode layer obtain the hierarchical fault features and hierarchical reconstruction attribute semantics of the corresponding layers respectively; during testing, the fault attribute semantic description matrix of each layer input is replaced by the reconstruction attribute semantics of each layer.

Images