Unmanned aerial vehicle intelligent fault diagnosis method for sample-label noise coupling scene

Abstract

The invention relates to the technical field of unmanned aerial vehicle fault diagnosis, in particular to an unmanned aerial vehicle intelligent fault diagnosis method for a sample-label noise coupling scene. The method comprises the following steps: acquiring a signal in the operation process of an unmanned aerial vehicle, and injecting Gaussian white noise into the collected original signal to simulate sample noise; constructing an LWPT sub-network to carry out denoising on a signal containing sample noise; performing sample division and expansion on the denoised signal by using a sliding window; dividing a training set and a test set for a fault diagnosis task, and injecting label noise into the training set; constructing an SGLE sub-network to weaken the influence of label noise in the training process; and verifying the fault diagnosis performance of the LWPT-SGLE model under the influence of sample-label noise coupling. The method has excellent fault diagnosis performance in an experiment under the influence of sample and label noise coupling in an industrial scene, and an effective solution is provided for solving the problem of unmanned aerial vehicle intelligent fault diagnosis under the influence of complex noise coupling.

Description

Technical Field

[0001] The present invention relates to the technical field of UAV fault diagnosis, and particularly relates to an intelligent UAV fault diagnosis method for a sample-label noise coupling scenario. Background Art

[0002] Unmanned aerial vehicles (UAVs) are important carriers of modern intelligent systems and have played a key role in fields such as military reconnaissance, environmental monitoring, emergency rescue, and commercial services. The reliable operation of UAVs is crucial for system safety and mission completion. If a UAV malfunctions, it may lead to mission failure, equipment crash, and even serious personal injuries, thereby causing significant emergency losses and safety accidents. Therefore, developing intelligent fault diagnosis for UAVs is of great significance in engineering applications.

[0003] Existing UAV fault diagnosis methods use flight data and advanced deep learning methods to achieve fast and accurate diagnosis. Although these methods have made certain progress, they often face the challenge of low-quality data in actual UAV application scenarios. On the one hand, complex flight environments and sensor measurement errors can introduce sample noise, causing the model to learn incorrect features or trends. These incorrect features can trigger false alarms or missed detections in the fault detection system, resulting in unnecessary maintenance of the UAV (increasing maintenance costs) or missing critical fault points (leading to equipment damage or even crash accidents). On the other hand, due to reasons such as ambiguous label concepts or changes in label annotation standards over time, label noise is caused, making the model receive incorrect feedback during the training process and thus learning incorrect mapping relationships. This directly affects the accuracy of fault diagnosis, reduces system reliability, and in actual applications, it is manifested as an increase in the average detection delay, resulting in an extended UAV mission interruption time and a significant decline in operational efficiency. Furthermore, their coupled effects can also cause oscillations in the model training process, making it difficult to converge, extending the algorithm development cycle, and increasing the R & D investment cost. These problems pose a huge challenge to reliable UAV fault diagnosis, not only affecting the service life and safety of the equipment, but also directly related to the economic benefits and market competitiveness of UAV operations. Therefore, it is urgent to explore intelligent fault diagnosis methods that can handle low-quality data. Summary of the Invention

[0004] The technical problem solved by the present invention is to provide an intelligent UAV fault diagnosis method for a sample-label noise coupling scenario, which is used to solve the problem that data samples are interfered by sample noise and label information is inaccurate in a complex flight environment, so as to overcome the decline in the diagnostic performance of the model caused by the coupled effects of sample noise and label noise.

[0005] The basic solution provided by the present invention: An intelligent UAV fault diagnosis method for a sample-label noise coupling scenario includes the following steps: Step 1: Use sensors to obtain signals during the normal operation of the drone and in the states of motor failure, propeller failure, low voltage failure, load loss failure, accelerometer failure, gyroscope failure, magnetometer failure, barometer failure, and GPS failure respectively, and inject Gaussian white noise into the collected original signals to simulate sample noise; Step 2: Construct a learnable wavelet packet transform (LWPT) sub-network to denoise the noisy signals; the LWPT sub-network is an autoencoder architecture based on wavelet transform; Among them, the encoder part is used to decompose each node of the original signal into a low-frequency part and a high-frequency part, apply learnable thresholds to all nodes to suppress noise-related coefficients, and learn parameters to adapt to the frequency characteristics of the input signal, and effectively remove noise by learning appropriate sparse representations; After obtaining the clean coefficients through the encoder, based on the clean coefficients, the decoder part adaptively optimizes the parameters during the reconstruction process using learnable transposed convolution, gradually reconstructs the signal, and performs signal denoising; Step 3: Use a sliding window to divide and expand the denoised signals to obtain samples; Step 4: Divide the training set and the test set for the fault diagnosis task, and inject label noise into the training set; The method of injecting label noise into the training set in Step 4 is: through the noise transfer matrix to describe the probability of clean labels flipping to noise labels. The injected label noise includes symmetric label noise and asymmetric label noise. The symmetric label noise is based on the symmetric label noise transfer matrix to flip clean label samples of a certain class to the labels of the remaining classes with the same probability. The asymmetric label noise is based on the asymmetric label noise transfer matrix to make samples of a certain class have a high similarity with another specific class; Step 5: Construct a scale graph learning enhancement (SGLE) sub-network to weaken the influence of label noise during the training process. In Step 5, the SGLE sub-network includes the following iterative steps: S5.1: Sample selection: Downsample the sample signals to obtain signals of different scales, fuse the signals of different scales and perform learning, and select samples with clean labels based on the small loss criterion; S5.2: Sample relabeling: For the unselected samples, based on the label propagation theory, construct a graph embedding learning network component, and use the clean label samples through the embedding learning network component to correct the labels of the unselected samples; Step 6: Test the trained LWPT-SGLE model and test its fault diagnosis results under the influence of sample-label noise coupling.

[0006] Further, in the first step, the method of injecting Gaussian white noise is expressed as:

[0007] where is the signal after injecting white noise is the original signal, represents the noise intensity, is Gaussian white noise that follows a standard normal distribution.

[0008] Further, in the second step, the LWPT sub-network includes an encoder and a decoder; The encoder uses a convolutional layer with a stride of 2 to replace the filter part of the traditional WPT, and uses a learnable denoising activation function η for feature extraction and noise suppression:

[0009] where, refers to the learnable bias; after decomposition, the decomposition coefficient of the j-th node in the i-th layer is calculated by the following formula :

[0010] where represents the transposed convolution, represents the convolution operation, represents obtaining the corresponding input from the previous node; In the decoder part, a learnable transposed convolutional layer with a stride of 2 is used to gradually reconstruct the signal, and the clean coefficient is calculated by the following formula :

[0011] where is the transposed convolution kernel, represents the upsampling operation; Perform the transposed convolution operation on the clean coefficient:

[0012] Add to obtain the completely reconstructed complete signal :

[0013] where L represents the number of LWPT decomposition layers.

[0014] Further, in the third step, a sliding window is used to divide the denoised signal into samples. The size of the sliding window is 128 and the stride is 16.

[0015] Further, the specific steps of S5.1 are as follows: S5.1.1. The signal after LWPT denoising is expressed as , and downsample the signal:

[0016] where is the downsampling rate and belongs to between, The larger

[0017] where is the learned single-scale embedding, is 's embedding, refers to the concatenation operation of two vectors, represents a non-linear neural network for fusing information of different scales; S5.1.2. Calculate the loss of each sample :

[0018] where represents the predicted label, is the actual label, is the number of classes; S5.1.3. Sort from small to large based on the damage of each sample, and set the label noise rate to , and the part of the samples corresponding to the loss are selected as clean samples.

[0019] Furthermore, the specific steps of S5.2 are as follows: S5.2.1. Construct a graph embedding learning network, and represent the fused embedding as , M represents the batch size; S5.2.2. Perform momentum update in the early stage of the training process:

[0020] where refers to the current training round number, represents the momentum update parameter; S5.2.3. Use the Gaussian similarity function to calculate the edge weights of the nearest neighbor graph:

[0021] where represents calculating the Euclidean distance, is a fixed parameter; S5.2.4. Perform graph Laplacian operation on the edge weight E:

[0022] where D is a diagonal matrix; S5.2.5. According to the label propagation theory, calculate the pseudo-labels of each node in through the following formula:

[0023] where F refers to the predicted pseudo-label, is a hyperparameter between 0 and 1, and the corrected label of the unselected sample is:

[0024] where argmax represents the pseudo-label of taking the maximum value.

[0025] Furthermore, in the sixth step, the fault diagnosis result of testing LWPT-SGLE refers to verifying the comparison between the denoised signal and the original signal, and verifying the fault diagnosis classification accuracy under the influence of label noise.

[0026] The principle and advantages of the present invention are as follows: 1. This solution constructs a fault diagnosis system framework for the influence of coupled noise. Through the cascading mechanism of the LWPT sub-network and the SGLE sub-network, it realizes the collaborative optimization of sample noise elimination and label noise suppression, and improves the diagnostic robustness in the scenario of complex noise influence on UAVs at the system level.

[0027] 2. In this solution, the LWPT sub-network integrates deep learning methods on the basis of traditional signal methods. It uses convolutional layers to replace the filter part of the traditional wavelet packet transform, and applies learnable thresholds at all nodes to suppress noise-related coefficients, realizing dynamic threshold adjustment and improving the sample noise denoising performance.

[0028] 3. In this solution, a label noise scenario under a simulated coupled noise scenario is simulated, and noise labels are injected into the training set based on the idea of a noise transfer matrix. In addition to mapping the original label to a noise label of other categories with the same probability based on a symmetric label noise transfer matrix, the case where features are similar in special scenarios and are more likely to be marked as similar categories is also considered, and label noise is injected into the training set based on an asymmetric label noise transfer matrix.

[0029] 4. In this solution, the SGLE sub-network part weakens the influence of label noise during the training process through two steps: sample selection and sample relabeling. Different-scale data is obtained by downsampling the signal data, and these data are fused and learned. Clean samples are selected based on the small-loss criterion, and the labels of the unselected samples are corrected using the clean samples through graph embedding learning based on the label propagation theory.

[0030] 5. When verifying the fault diagnosis results under the influence of coupled noise in this solution, the removal effects of sample noise are qualitatively (visualizing the denoised signal and the original signal) and quantitatively analyzed (calculating evaluation indicators). In addition, different types and levels of label noise are set, and the classification accuracy is used for evaluation, fully verifying the model performance of LWPT-SGLE.

[0031] 6. The advantage of this solution is that most existing fault diagnosis methods for noisy data study sample noise and label noise independently and do not consider the influence of the simultaneous occurrence of the two types of noise on fault diagnosis. This solution aims to address this point, weakening the influence of label noise during the training process while removing sample noise to improve the diagnosis results under complex noise.

[0032] In the actual scenario of UAV fault diagnosis, sensor signals are vulnerable to environmental interference, resulting in sample noise. At the same time, the manual annotation process is prone to generate label noise. The superposition of the two makes it challenging to use these training data for high-quality fault diagnosis. Starting from solving the above problems, this solution aims to achieve ideal fault diagnosis results in the case of complex noise coupling data, and is more applicable to real data scenarios than mainstream methods, with good application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a schematic flowchart of an embodiment of the UAV intelligent fault diagnosis method for the sample-label noise coupling scenario of the present invention; Figure 2 It is a principle block diagram of an embodiment of the UAV intelligent fault diagnosis method for the sample-label noise coupling scenario of the present invention; Figure 3 It is a qualitative analysis diagram of the sample noise denoising effect in an embodiment of the UAV intelligent fault diagnosis method for the sample-label noise coupling scenario of the present invention; Figure 4 It is a quantitative analysis diagram of the sample noise denoising effect in an embodiment of the UAV intelligent fault diagnosis method for the sample-label noise coupling scenario of the present invention; Figure 5 It is a schematic diagram of the fault diagnosis accuracy under the influence of different types and degrees of label noise in an embodiment of the UAV intelligent fault diagnosis method for the sample-label noise coupling scenario of the present invention; Figure 6This is the gain effect diagram of scale integration learning when suppressing label noise in the embodiment of the UAV intelligent fault diagnosis method for the sample-label noise coupling scenario of the present invention. Detailed implementation manners

[0034] The following is a further detailed description through specific implementation manners: The embodiment is basically as shown in the appendix Figure 1 as follows: The UAV intelligent fault diagnosis method for the sample-label noise coupling scenario is as shown in Figure 1 as follows: Mainly, sample noise injection is realized by adding Gaussian white noise to the obtained original UAV signal. Utilizing the learnable characteristics of deep learning, deep learning is combined with traditional signal processing methods to construct a learnable wavelet packet transform sub-network, realizing learnable threshold adjustment at each node of the wavelet packet decomposition tree, and achieving efficient denoising of sample noise. When using the denoised signal as the input of the diagnosis model, label noise is introduced by swapping sample labels. The original signal data is downsampled through a scale integration learning strategy, data at different scales is fused and clean samples are selected based on a small loss criterion. Based on the label propagation theory, a graph embedding learning network is constructed to correct the labels of the unselected samples using the clean samples. Robust diagnosis of UAV faults under the influence of sample-label noise coupling is realized, which specifically includes the following steps: Step 1: Use sensors to respectively obtain the signals of the UAV in the normal state and the states of motor faults, propeller faults, low voltage faults, load loss faults, accelerometer faults, gyroscope faults, magnetometer faults, barometer faults, and GPS faults during the operation of the UAV, and inject Gaussian white noise into the collected original signals to simulate sample noise.

[0035] In the above Step 1, the method of injecting Gaussian white noise is expressed as:

[0036] where is the signal after injecting white noise is the original signal, represents the noise intensity, is the Gaussian white noise obeying the standard normal distribution.

[0037] Step 2: Construct a learnable wavelet packet transform LWPT sub-network to denoise the noisy signal; as shown in Figure 2 the LWPT sub-network is an auto-encoder architecture based on wavelet transform.

[0038] Among them, the encoder part is used to decompose each node of the original signal into a low-frequency part and a high-frequency part, apply a learnable threshold to all nodes to suppress noise-related coefficients, and learn parameters to adapt to the frequency characteristics of the input signal, and effectively remove noise by learning an appropriate sparse representation.

[0039] After obtaining the clean coefficients through the encoder, based on the clean coefficients, the decoder part adaptively optimizes the parameters in the reconstruction process using learnable transposed convolutions, gradually reconstructs the signal, and performs signal denoising. The specific implementation process of step two is as follows: The LWPT sub-network includes an encoder and a decoder.

[0040] The encoder uses a convolutional layer with a stride of 2 to replace the filter part of the traditional WPT, and uses a learnable denoising activation function η for feature extraction and noise suppression:

[0041] Among them, refers to the learnable bias; after decomposition, the decomposition coefficient of the j-th node in the i-th layer is calculated by the following formula :

[0042] Among them represents the transposed convolution, represents the convolution operation, represents obtaining the corresponding input from the previous node.

[0043] In the decoder part, a learnable transposed convolutional layer with a stride of 2 is used to gradually reconstruct the signal, and the clean coefficients are calculated by the following formula :

[0044] Among them is the transposed convolution kernel, represents the upsampling operation.

[0045] Perform a transposed convolution operation on the clean coefficients:

[0046] Add to obtain the completely reconstructed complete signal :

[0047] where L represents the number of LWPT decomposition layers.

[0048] Step three: Use a sliding window to partition and augment the samples of the denoised signal. The size of the sliding window is 128 and the stride is 16.

[0049] Step 4: Divide the training set and the test set for the fault diagnosis task. The method for dividing the training set is to divide it into a training set and a test set according to the ratio of 8:2 of the total number of samples, and inject label noise into the training set. The way to inject label noise in Step 4 is: through the noise transfer matrix to describe the probability of clean labels flipping to noise labels. The injected label noise includes symmetric label noise and asymmetric label noise. The symmetric label noise is based on the symmetric label noise transfer matrix to flip a certain type of clean label sample to the labels of the remaining categories with the same probability. The asymmetric label noise is based on the asymmetric label noise transfer matrix to make the samples of a certain category have a high similarity with another specific category.

[0050] Step 5: Construct a scale graph to enhance the SGLE sub-network to weaken the influence of label noise during the training process. In Step 5, the SGLE sub-network includes the following iterative steps: S5.1. Sample selection: Downsample the sample signal to obtain signals of different scales, fuse the signals of different scales and perform learning, and select samples with clean labels based on the small loss criterion.

[0051] S5.1 specifically includes the following steps: S5.1.1. The signal after LWPT denoising is expressed as , and downsample the signal:

[0052] where is the downsampling rate and belongs to between, the larger it is, the larger the data scale in the original signal. Fuse the sample embeddings in the embedding space according to the scale from fine-grained to coarse-grained:

[0053] where is the learned single-scale embedding, is the embedding of , refers to the concatenation operation of two vectors, represents a non-linear neural network for fusing information of different scales.

[0054] S5.1.2. Calculate the loss of each sample :

[0055] where represents the predicted label, is the actual label, is the number of categories; S5.1.3. Sort the damages of each sample from smallest to largest, and set the label noise rate to , and the samples corresponding to the loss are selected as clean samples.

[0056] S5.2. Sample relabeling: For the unselected samples, based on the label propagation theory, construct a graph embedding learning network component, and use the clean label samples to correct the labels of the unselected samples through the embedding learning network component.

[0057] S5.2 specifically includes the following steps: S5.2.1. Construct a graph embedding learning network, and represent the fusion embedding as , where M represents the batch size.

[0058] S5.2.2. Perform momentum update in the early stage of the training process:

[0059] where refers to the current training epoch, represents the momentum update parameter.

[0060] S5.2.3. Use the Gaussian similarity function to calculate the edge weights of the nearest neighbor graph:

[0061] where represents calculating the Euclidean distance, is a fixed parameter; S5.2.4. Perform graph Laplacian operation on the edge weights E:

[0062] where D is a diagonal matrix; S5.2.5. According to the label propagation theory, calculate the pseudo-labels of each node in through the following formula:

[0063] where F refers to the predicted pseudo-label, is a hyperparameter between 0 and 1, and the corrected label of the unselected sample is:

[0064] where argmax represents taking the pseudo-label of the maximum value.

[0065] The influence of label noise can be reduced through Steps 5.1 and 5.2. Additionally, learning rate warm-up was performed at the beginning of training. After the warm-up phase, several additional training sessions were carried out, and the label correction phase began only after these rounds of training were completed.

[0066] Step 6: Test the fault diagnosis results of the LWPT-SGLE model under the influence of sample-label noise coupling. Testing the fault diagnosis results of LWPT-SGLE refers to verifying the comparison between the denoised signal and the original signal, and verifying the fault diagnosis classification accuracy under the influence of label noise.

[0067] Taking the publicly available UAV fault dataset as an example, hardware simulation data was used for testing. The data acquisition process was based on the hardware of the Zhuoyi X450 UAV, with RflySim as the core simulation platform, which was integrated with the PX4 flight control. The RflySim platform executed the control program in sequence. When a fault was injected, the abnormal performance of the multi-rotor aircraft during the fault injection process would confirm the successful simulation of the fault conditions. Nine different fault types and normal operating conditions were simulated, and the fault categories included motor faults, propeller faults, low voltage faults, load loss faults, accelerometer faults, gyroscope faults, magnetometer faults, barometer faults, and GPS faults.

[0068] Figure 3 Figure shows the effect of LWPT-SGLE in removing sample noise. Signals reconstructed by traditional methods such as Kalman Filter (KF) have deviated significantly from the original signal, showing an over-smoothed effect. Deep learning methods such as Convolutional Neural Network (CNN) restored the trend of the original signal but could not effectively remove the high-frequency noise in the signal, resulting in overlapping reconstructed signals. Long Short-Term Memory Network (LSTM) experienced memory decay when processing long sequences, and periodic bumps appeared in the reconstructed signal, making the denoising effect less than ideal. In contrast, LWPT-SGLE performed best in different signal segments, especially in regions with mutations. It utilized the non-linear characteristics of the deep learning model to effectively handle complex noise patterns. The denoised signal was very close to the original signal, accurately retaining important features and turning points.

[0069] In addition, the denoising effect of LWPT-SGLE on sample noise was quantitatively analyzed, such as Figure 4As shown, by comparing the evaluation metrics (R², MSE, and MAE) of four different signal denoising methods, the results show that the LWPT-SGLE method achieved the best performance in all evaluation dimensions. Its R² value is closest to 1, while the MSE and MAE values are the smallest, indicating that this method has the best signal fitting degree and the lowest error level. The traditional KF method can also achieve a basic denoising effect, but it is significantly lower than the other three methods in terms of various metrics. The performance of the LSTM method and the CNN method is relatively close. Both are better than the traditional KF method, but there is still a certain gap compared with the LWPT-SGLE method. These quantitative evaluation results corroborate the previous waveform analysis results, further verifying the superiority of the LWPT-SGLE method in signal denoising tasks.

[0070] Figure 5 The comparison of the diagnostic accuracies of various methods under different types and levels of label noise is shown. The cross-entropy (CE) method performs the most basic and is at the bottom in all symmetry parameters. The SREA method has a significant improvement compared to CE, and its overall performance curve is parallel to but higher than CE. Mixup-BMM shows better robustness, with a relatively gentle performance degradation curve and a slightly smaller degradation scale, and its performance advantage is more obvious when the symmetric label noise level is large. Finally, the LWPT-SGLE method shows the best performance. It not only maintains the highest accuracy under all symmetric label noise parameters, but also has the gentlest degree of performance degradation, indicating that this method has significant advantages in dealing with different degrees of symmetry problems. It is worth noting that in the case of asymmetric label noise, the performance ranking of each method remains the same, and LWPT-SGLE is still superior to other methods. This is due to the fact that the model achieves the highest accuracy through a robust sample selection and label correction mechanism, showing good robustness and adaptability. When dealing with label noise, the scale integration learning strategy plays a key role. As Figure 6 shown, under low symmetric label noise conditions, the performance gap between single-scale learning and scale integration learning is relatively small. However, as the noise intensity increases, the gap gradually expands. This trend indicates the importance of the scale integration strategy, which effectively captures local and global features at multiple scales through the integration learning process, thereby reducing the interference effect of label noise during the training process.

[0071] In this embodiment, the LWPT-SGLE model proposes a two-stage collaborative optimization solution for the problem of fault diagnosis of drones under the coupled interference of sample noise and label noise. In the first stage, by injecting Gaussian white noise into the original flight signal to simulate the actual environmental interference, a learnable wavelet packet transform subnet (LWPT) is constructed. Using its characteristic of dynamically adjusting the threshold at each node of the wavelet packet decomposition tree, the fixed threshold mechanism of the traditional wavelet packet filter is upgraded to a deep learning-based adaptive noise suppression method, significantly improving the denoising accuracy for non-stationary noise. In the second stage, first, the denoised signal is segmented by a multi-scale sliding window to expand sample diversity. Subsequently, asymmetric label noise is synchronously injected into the training set to simulate manual annotation errors, and a graph embedding learning subnet (SGLE) is constructed to implement noise-immune training - generating multi-scale representations through downsampling and fusing and learning, screening high-confidence samples based on the small-loss criterion, and then constructing a graph structure to propagate label semantics to achieve dynamic correction of noise labels and joint optimization of the feature space. Through the cross-level interaction between LWPT and SGLE, this method forms a closed-loop optimization link from signal denoising to label enhancement, providing key technical support for the reliable operation and maintenance of drones in complex operating environments and fast response scenarios.

[0072] The above are only embodiments of the present invention. Specific structures and common knowledge such as characteristics that are well-known in the art are not described in detail here. Those of ordinary skill in the art know all the common general technical knowledge in the technical field to which the invention pertains before the application date or the priority date, can know all the existing technologies in this field, and have the ability to apply conventional experimental means before this date. Those of ordinary skill in the art can, under the inspiration given in this application, complete and implement this solution in combination with their own abilities. Some typical well-known structures or well-known methods should not become an obstacle for those of ordinary skill in the art to implement this application. It should be noted that for those skilled in the art, without departing from the structure of the present invention, several deformations and improvements can still be made, and these should also be regarded as the protection scope of the present invention, which will not affect the implementation effect of the present invention and the practicality of the patent. The protection scope required by this application should be subject to the content of its claims, and the specific implementation manners described in the specification can be used to interpret the content of the claims.

Claims

1. An intelligent fault diagnosis method for drones in the sample-label noise coupling scenario, characterized in that: It includes the following steps: Step 1: Use sensors to obtain signals during the normal operation of the drone and in the states of motor failure, propeller failure, low voltage failure, load loss failure, accelerometer failure, gyroscope failure, magnetometer failure, barometer failure, and GPS failure respectively, and inject Gaussian white noise into the collected original signals to simulate sample noise; Step 2: Construct a learnable wavelet packet transform (LWPT) sub-network to denoise the noisy signals; the LWPT sub-network is an autoencoder architecture based on wavelet transform; Among them, the encoder part is used to decompose each node of the original signal into a low-frequency part and a high-frequency part, apply learnable thresholds to all nodes to suppress noise-related coefficients, and learn parameters to adapt to the frequency characteristics of the input signal, and remove noise by learning appropriate sparse representations; After obtaining the clean coefficients through the encoder, based on the clean coefficients, the decoder part adaptively optimizes the parameters during the reconstruction process using learnable transposed convolutions, gradually reconstructs the signal, and performs signal denoising; Step 3: Use a sliding window to divide and expand the denoised signals into samples; Step 4: Divide the training set and the test set for the fault diagnosis task, and inject label noise into the training set; The method of injecting label noise into the training set in the fourth step is as follows: through the noise transition matrix to describe the probability of clean labels flipping to noisy labels. The injected label noise includes symmetric label noise and asymmetric label noise. The symmetric label noise is based on the symmetric label noise transition matrix to flip a certain type of clean label sample to the remaining category labels with the same probability. The asymmetric label noise is based on the asymmetric label noise transition matrix to make samples of one category have a high similarity to another specific category; Step 5: Construct a scale graph enhancement (SGLE) sub-network to weaken the influence of label noise during the training process. In Step 5, the SGLE sub-network includes the following iterative steps: S5.1: Sample selection: Downsample the sample signals to obtain signals of different scales, fuse the signals of different scales and perform learning, and select samples with clean labels based on the small-loss criterion; S5.2: Sample relabeling: For the unselected samples, based on the label propagation theory, construct a graph embedding learning network component, and use the clean-label samples through the embedding learning network component to correct the labels of the unselected samples; Step 6: Test the trained LWPT-SGLE model and test its fault diagnosis results under the influence of sample-label noise coupling.

2. The intelligent UAV fault diagnosis method for the sample-label noise coupling scenario according to claim 1, wherein: In Step 1, the method of injecting Gaussian white noise is expressed as: Among them is the signal after injecting white noise is the original signal represents the noise intensity is Gaussian white noise that follows a standard normal distribution 3. The UAV intelligent fault diagnosis method for the sample-label noise coupling scenario according to claim 1, wherein: In Step 2, the LWPT sub-network includes an encoder and a decoder; The encoder uses a convolutional layer with a stride of 2 to replace the filter part of the traditional WPT, and uses a learnable denoising activation function η for feature extraction and noise suppression: Among them, refers to the learnable bias; after decomposition, the decomposition coefficient of the j-th node in the i-th layer is calculated by the following formula : Among them represents transposed convolution represents convolution operation represents obtaining the corresponding input from the previous node The decoder part uses a learnable transposed convolutional layer with a stride of 2 to gradually reconstruct the signal, and calculates the clean coefficients through the following formula :[[]]END]] wherein is the transposed convolution kernel, represents the upsampling operation; Perform an inverse convolution operation on the clean coefficients: Add to obtain the completely reconstructed complete signal : Where L represents the number of LWPT decomposition layers.

4. The UAV intelligent fault diagnosis method for the sample-label noise coupling scenario according to claim 1, characterized in that: In Step 3, use a sliding window to divide the denoised signals into samples. The size of the sliding window is 128 and the stride is 16.

5. The method for intelligent fault diagnosis of an unmanned aerial vehicle for a sample-label noise coupling scenario according to claim 1, wherein: S5.1 specifically includes the following steps: S5.1.

1. The signal after LWPT denoising is expressed as , and the signal is downsampled: wherein is the downsampling rate and belongs to between The larger it is, the larger the data scale of the data in the original signal. The samples embedded in the embedding space are fused according to the scale from fine-grained to coarse-grained: where is the learned single-scale embedding, is 's embedding, refers to the concatenation operation of two vectors, represents a non-linear neural network for fusing information of different scales; S5.1.

2. Calculate the loss of each sample : wherein represents the predicted label, is the actual label, is the number of categories; S5.1.

3. Sort in ascending order of damage for each sample, and set the label noise rate to , and select the samples in the part corresponding to the loss as clean samples.

6. The intelligent UAV fault diagnosis method for the sample-label noise coupling scenario according to claim 5, characterized in that: S5.2 specifically includes the following steps: S5.2.

1. Construct a graph embedding learning network and represent the fusion embedding as , where M represents the batch size; S5.2.2: Perform momentum update in the early stage of the training process: where refers to the current training round number, represents the momentum update parameter; S5.2.3: Use a Gaussian similarity function to calculate the edge weights of the nearest neighbor graph: Among them represents calculating the Euclidean distance is a fixed parameter S5.2.4: Perform a graph Laplacian operation on the edge weights E: Where D is a diagonal matrix; S5.2.

5. Calculate the pseudo-labels of each node in according to the label propagation theory as follows: The pseudo-labels of each node in: where F refers to the predicted pseudo-label, is a hyperparameter between 0 and 1, and the corrected label for the unselected samples is: Where argmax represents the pseudo-label of taking the maximum value.

7. The intelligent UAV fault diagnosis method for the sample-label noise coupling scenario according to claim 1, characterized in that: In Step 6, testing the fault diagnosis results of LWPT-SGLE is to verify the comparison between the denoised signal and the original signal, and verify the fault diagnosis classification accuracy under the influence of label noise.

Images