Intelligent fault diagnosis method for UAVs in sample-label noise coupling scenarios

Abstract

The present invention relates to the technical field of drone fault diagnosis, and specifically to a drone intelligent fault diagnosis method for sample-label noise coupling scenarios. The steps include: acquiring signals during drone operation, injecting Gaussian white noise into the acquired original signals to simulate sample noise; constructing an LWPT subnetwork to denoise the signals containing sample noise; using a sliding window to divide and expand the samples of the denoised signals; dividing the training set and the test set for fault diagnosis tasks, and injecting label noise into the training set; constructing an SGLE subnetwork to reduce the influence of label noise during training; and verifying the fault diagnosis performance of the LWPT-SGLE model under the influence of sample-label noise coupling. The present invention has excellent fault diagnosis performance in experiments under the influence of sample and label noise coupling in industrial scenarios, and provides an effective solution to the problem of intelligent fault diagnosis of drones affected by complex noise coupling.

Description

Technical Field

[0001] The present invention relates to the technical field of drone fault diagnosis, and in particular to an intelligent drone fault diagnosis method for sample-label noise coupling scenarios. Background Art

[0002] Drones (UAVs), a crucial enabler of modern intelligent systems, have played a crucial role in military reconnaissance, environmental monitoring, emergency rescue, and commercial services. Reliable UAV operation is crucial to system safety and mission accomplishment. A UAV malfunction can lead to mission failure, equipment crashes, or even serious personal injury, resulting in significant emergency losses and safety incidents. Therefore, intelligent UAV fault diagnosis is crucial in engineering applications.

[0003] Existing drone fault diagnosis methods leverage flight data and advanced deep learning methods to achieve rapid and accurate diagnosis. While these methods have achieved some progress, they often face the challenge of low-quality data in practical drone applications. Complex flight environments and sensor measurement errors introduce sample noise, causing the model to learn erroneous features or trends. These erroneous features can trigger false positives or negatives in the fault detection system, leading to unnecessary repairs (increasing maintenance costs) or missed critical fault points (causing equipment damage or even crashes). Furthermore, label noise, caused by factors such as fuzzy label concepts or time-varying variations in labeling standards, can cause the model to receive erroneous feedback during training and learn incorrect mapping relationships. This directly impacts fault diagnosis accuracy and reduces system reliability. In practical applications, this manifests as increased average detection latency, prolonged drone mission interruptions, and significantly reduced operational efficiency. Furthermore, these combined effects can cause model training to oscillate, hinder convergence, extend algorithm development cycles, and increase R&D costs. These issues pose significant challenges to reliable drone fault diagnosis, impacting not only the lifespan and safety of the equipment but also the economic benefits and market competitiveness of drone operations. Therefore, there is an urgent need 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 fault diagnosis method for drones in sample-label noise coupling scenarios, which is used to solve the problem that data samples are interfered with by sample noise and label information is inaccurate in complex flight environments, so as to overcome the degradation of model diagnosis performance caused by the coupling of sample noise and label noise.

[0005] The basic solution provided by the present invention is an intelligent fault diagnosis method for UAVs in sample-label noise coupling scenarios, which includes the following steps:

[0006] Step 1: Use sensors to obtain signals of the drone in normal operation and in motor failure, propeller failure, low voltage failure, load loss failure, accelerometer failure, gyroscope failure, magnetometer failure, barometer failure, and GPS failure states, and inject Gaussian white noise into the collected original signals to simulate sample noise;

[0007] Step 2: Construct a learnable wavelet packet transform (LWPT) sub-network to denoise the noisy signal; the LWPT sub-network is an autoencoder architecture based on wavelet transform;

[0008] The encoder part is used to decompose each node of the original signal into low-frequency and high-frequency parts, apply a learnable threshold at all nodes to suppress the coefficients related to noise, and learn parameters to adapt to the frequency characteristics of the input signal, effectively removing noise by learning an appropriate dilution representation;

[0009] The encoder obtains clean coefficients. Based on these clean coefficients, the decoder uses learnable transposed convolution to adaptively optimize the parameters in the reconstruction process, gradually reconstruct the signal, and perform signal denoising.

[0010] Step 3: Use a sliding window to divide and expand the samples of the denoised signal;

[0011] Step 4: Divide the training set and test set for fault diagnosis tasks, and inject label noise into the training set;

[0012] The method of injecting label noise into the training set in step 4 is: To describe the probability of a clean label flipping to a noise label, 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 With the same probability, a certain class of clean label samples is flipped to the rest of the class labels. The asymmetric label noise is based on the asymmetric label noise transfer matrix , so that the samples of a certain category have a high similarity with another specific category;

[0013] Step 5: Construct a scale graph to enhance the SGLE sub-network to reduce the impact of label noise during training. In step 5, the SGLE sub-network includes the following iterative steps:

[0014] 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 minimum loss criterion;

[0015] S5.2. Sample Relabeling: For unselected samples, a graph embedding learning network component is constructed based on label propagation theory. The embedding learning network component uses clean label samples to perform label correction on the unselected samples.

[0016] Step 6: Test the trained LWPT-SGLE model to test its fault diagnosis results under the influence of sample-label noise coupling.

[0017] Furthermore, in step 1, the method of injecting Gaussian white noise is expressed as:

[0018]

[0019] in is the signal after injecting white noise is the original signal, represents the noise intensity, is Gaussian white noise that obeys the standard normal distribution.

[0020] Furthermore, in step 2, the LWPT sub-network includes an encoder and a decoder;

[0021] 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:

[0022]

[0023] in, Refers to the learnable bias; after decomposition, the decomposition coefficient of the jth node in the i-th layer is calculated by the following formula: :

[0024]

[0025] in represents inverse convolution, represents the convolution operation, Indicates getting the corresponding input from the previous node;

[0026] In the decoder part, a learnable transposed convolution layer with a step size of 2 is used to gradually reconstruct the signal, and the clean coefficient is calculated by the following formula :

[0027]

[0028] in is the deconvolution kernel, Represents an upsampling operation;

[0029] Perform deconvolution on the clean coefficients:

[0030]

[0031] Will Add together to get the complete reconstructed signal :

[0032]

[0033] Where L represents the number of LWPT decomposition layers.

[0034] Furthermore, in step three, a sliding window is used to divide the denoised signal into samples, the size of the sliding window is 128, and the step size is 16.

[0035] Furthermore, the S5.1 specifically includes the following steps:

[0036] S5.1.1, the signal after LWPT denoising is expressed as , downsample the signal:

[0037]

[0038] in is the downsampling rate and belongs to between, The larger the value is, the larger the data scale of the original signal is. The sample embeddings in the embedding space are fused according to the scale from fine-grained to coarse-grained:

[0039]

[0040] in is the learned single-scale embedding, for Embedded, Refers to the concatenation operation of two vectors, Represents a nonlinear neural network used to fuse information at different scales;

[0041] S5.1.2. Calculate the loss for each sample :

[0042]

[0043] in represents the predicted label, is the actual label, is the number of categories;

[0044] S5.1.3. Sort the damage of each sample from small to large, and set the label noise rate to , corresponding to the loss Some samples are selected as clean samples.

[0045] Furthermore, the S5.2 specifically includes the following steps:

[0046] S5.2.1. Construct a graph embedding learning network and represent the fused embedding as , M represents the batch size;

[0047] S5.2.2. Perform momentum updates early in the training process:

[0048]

[0049] in Refers to the number of rounds of current training, represents the momentum update parameter;

[0050] S5.2.3. Use the Gaussian similarity function to calculate the edge weights of the nearest neighbor graph:

[0051]

[0052] in Indicates the calculation of Euclidean distance, is a fixed parameter;

[0053] S5.2.4. Perform graph Laplacian operations on edge weights E:

[0054]

[0055] Where D is a diagonal matrix;

[0056] S5.2.5. According to label propagation theory, the following formula is used to calculate The pseudo label of each node in :

[0057]

[0058] Where F refers to the predicted pseudo label, is a hyperparameter between 0 and 1, and the corrected labels of the unselected samples are:

[0059]

[0060] Where argmax represents the pseudo label with the maximum value.

[0061] Furthermore, in step 6, testing the fault diagnosis result 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.

[0062] The principles and advantages of the present invention are:

[0063] 1. This solution constructs a fault diagnosis system framework for coupled noise. Through the cascade mechanism of the LWPT sub-network and the SGLE sub-network, it achieves the coordinated optimization of sample noise elimination and label noise suppression, and improves the diagnostic robustness of drones in complex noise-affected scenarios at the system level.

[0064] 2. In this scheme, the LWPT subnetwork integrates deep learning methods based on traditional signal methods, uses convolutional layers to replace the filter part of traditional wavelet packet transform, applies learnable thresholds at all nodes to suppress noise-related coefficients, realizes dynamic threshold adjustment, and improves sample noise denoising performance.

[0065] 3. This solution simulates label noise scenarios in coupled noise scenarios and injects noisy labels into the training set based on the concept of a noise transfer matrix. In addition to mapping original labels to noisy labels of other categories with equal probability using a symmetric label noise transfer matrix, it also considers special scenarios where similar features are more likely to be labeled as similar categories. This approach injects label noise into the training set using an asymmetric label noise transfer matrix.

[0066] 4. The SGLE subnetwork in this solution mitigates the impact of label noise during training through two steps: sample selection and sample relabeling. This involves downsampling the signal data to obtain data at different scales, integrating this data and selecting clean samples based on a minimal loss criterion. Based on label propagation theory, graph embedding learning uses the clean samples to correct the labels of unselected samples.

[0067] 5. In this solution, when verifying the fault diagnosis results under the influence of coupled noise, the noise removal effect of the samples was qualitatively analyzed (visualizing the denoised signal and the original signal) and quantitatively analyzed (calculating evaluation indicators). In addition, different types and levels of label noise were set and evaluated based on classification accuracy, fully verifying the model performance of LWPT-SGLE.

[0068] 6. The advantage of this solution is that most existing fault diagnosis methods for noisy data study sample noise and label noise independently, without considering the impact of both noise types on fault diagnosis. This solution aims to address this issue by simultaneously removing sample noise and reducing the impact of label noise during training, thereby improving diagnostic results under complex noise conditions.

[0069] In the real-world scenario of drone fault diagnosis, sensor signals are susceptible to environmental interference, leading to sample noise. Furthermore, manual annotation processes are prone to label noise. The combination of these two factors makes it challenging to use this training data for high-quality fault diagnosis. This solution addresses these challenges and aims to achieve ideal fault diagnosis results in complex, noisy data. Compared to mainstream methods, it is more applicable to real-world data scenarios and has significant application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 This is a flow chart of an embodiment of the intelligent fault diagnosis method for drones in a sample-label noise coupling scenario according to the present invention;

[0071] Figure 2 This is a principle block diagram of an embodiment of the UAV intelligent fault diagnosis method for sample-label noise coupling scenarios of the present invention;

[0072] Figure 3 This 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;

[0073] Figure 4 This 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;

[0074] Figure 5 This 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 intelligent fault diagnosis method for drones in a sample-label noise coupling scenario of the present invention;

[0075] Figure 6 This is a diagram showing the gain effect of time-scale integration learning when suppressing label noise in an embodiment of the intelligent fault diagnosis method for drones in a sample-label noise coupling scenario of the present invention. DETAILED DESCRIPTION

[0076] The following is further described in detail through specific implementation methods:

[0077] The embodiment is basically as shown in the attached Figure 1 As shown:

[0078] Intelligent fault diagnosis method for UAVs in sample-label noise coupling scenarios, such as Figure 1 As shown: Sample noise injection is mainly achieved by adding Gaussian white noise to the obtained original signal of the drone. By utilizing the learnable characteristics of deep learning, deep learning is combined with traditional signal processing methods to construct a learnable wavelet packet transform subnetwork, and learnable threshold adjustment is achieved at each node of the wavelet packet decomposition tree to achieve efficient denoising of sample noise. When the denoised signal is used as the input of the diagnostic model, label noise is introduced by exchanging sample labels. The original signal data is downsampled through a scale integration learning strategy, and data at different scales are fused and learned, and clean samples are selected based on the small loss criterion. Based on the label propagation theory, a graph embedding learning network is constructed to use clean samples to perform label correction on unselected samples. Robust diagnosis of drone faults under the influence of sample-label noise coupling is achieved, which specifically includes the following steps:

[0079] Step 1: Use sensors to obtain signals in the normal state of the drone during operation, as well as 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. Inject Gaussian white noise into the collected original signal to simulate sample noise.

[0080] In step 1, the method of injecting Gaussian white noise is expressed as:

[0081]

[0082] in is the signal after injecting white noise is the original signal, represents the noise intensity, is Gaussian white noise that obeys the standard normal distribution.

[0083] Step 2: Construct a learnable wavelet packet transform LWPT sub-network to denoise the noisy signal; Figure 2 As shown in FIG, the LWPT sub-network is an autoencoder architecture based on wavelet transform.

[0084] 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 at all nodes to suppress the coefficients related to noise, and learn parameters to adapt to the frequency characteristics of the input signal, effectively removing noise by learning an appropriate dilution representation.

[0085] After the encoder obtains clean coefficients, the decoder uses learnable transposed convolution to adaptively optimize the parameters in the reconstruction process based on the clean coefficients, gradually reconstructing the signal and performing signal denoising. The specific implementation process of step 2 is as follows:

[0086] The LWPT sub-network includes an encoder and a decoder.

[0087] 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:

[0088]

[0089] in, Refers to the learnable bias; after decomposition, the decomposition coefficient of the jth node in the i-th layer is calculated by the following formula: :

[0090]

[0091] in represents inverse convolution, represents the convolution operation, Indicates getting the corresponding input from the previous node.

[0092] In the decoder part, a learnable transposed convolution layer with a step size of 2 is used to gradually reconstruct the signal, and the clean coefficient is calculated by the following formula :

[0093]

[0094] in is the deconvolution kernel, Represents an upsampling operation.

[0095] Perform deconvolution on the clean coefficients:

[0096]

[0097] Will Add together to get the complete reconstructed signal :

[0098]

[0099] Where L represents the number of LWPT decomposition layers.

[0100] Step 3: Use a sliding window to divide and expand the sample of the denoised signal. The size of the sliding window is 128 and the step size is 16.

[0101] Step 4: Divide the training set and the test set for the fault diagnosis task. The method of dividing the training set is to divide it into the training set and the test set in a ratio of 8:2 according to the total number of samples, 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 a clean label flipping to a noise label, 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 With the same probability, a certain class of clean label samples is flipped to the rest of the class labels. The asymmetric label noise is based on the asymmetric label noise transfer matrix , so that the samples of a certain category have a high similarity with another specific category.

[0102] Step 5: Construct a scale graph to enhance the SGLE sub-network to reduce the impact of label noise during training. In step 5, the SGLE sub-network includes the following iterative steps:

[0103] 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 minimum loss criterion.

[0104] S5.1 specifically includes the following steps:

[0105] S5.1.1, the signal after LWPT denoising is expressed as , downsample the signal:

[0106]

[0107] in is the downsampling rate and belongs to between, The larger the value is, the larger the data scale of the original signal is. The sample embeddings in the embedding space are fused according to the scale from fine-grained to coarse-grained:

[0108]

[0109] in is the learned single-scale embedding, for Embedded, Refers to the concatenation operation of two vectors, Represents a nonlinear neural network used to fuse information at different scales.

[0110] S5.1.2. Calculate the loss for each sample :

[0111]

[0112] in represents the predicted label, is the actual label, is the number of categories;

[0113] S5.1.3. Sort the damage of each sample from small to large, and set the label noise rate to , corresponding to the loss Some samples are selected as clean samples.

[0114] S5.2. Sample relabeling: For unselected samples, a graph embedding learning network component is constructed based on label propagation theory. The clean label samples are used to perform label correction on the unselected samples through the embedding learning network component.

[0115] S5.2 specifically includes the following steps:

[0116] S5.2.1. Construct a graph embedding learning network and represent the fused embedding as , M represents the batch size.

[0117] S5.2.2. Perform momentum updates early in the training process:

[0118]

[0119] in Refers to the number of rounds of current training, represents the momentum update parameter.

[0120] S5.2.3. Use the Gaussian similarity function to calculate the edge weights of the nearest neighbor graph:

[0121]

[0122] in Indicates the calculation of Euclidean distance, is a fixed parameter;

[0123] S5.2.4. Perform graph Laplacian operations on edge weights E:

[0124]

[0125] Where D is a diagonal matrix;

[0126] S5.2.5. According to label propagation theory, the following formula is used to calculate The pseudo label of each node in :

[0127]

[0128] Where F refers to the predicted pseudo label, is a hyperparameter between 0 and 1, and the corrected labels of the unselected samples are:

[0129]

[0130] Where argmax represents the pseudo label with the maximum value.

[0131] The impact of label noise can be reduced by steps 5.1 and 5.2. In addition, a learning rate warmup is performed at the beginning of training. After the warmup phase, several additional training rounds are continued. After these rounds of training, the label correction phase begins.

[0132] 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 means verifying the comparison between the denoised signal and the original signal, and verifying the fault diagnosis classification accuracy under the influence of label noise.

[0133] Using a publicly available drone fault dataset as an example, hardware simulation data was used for testing. The data collection process was based on the Zowee X450 drone hardware, using RflySim as the core simulation platform and integrated with the PX4 flight control system. The RflySim platform sequentially executes control programs. When a fault is injected, the multirotor's abnormal behavior during the fault injection process confirms the successful simulation of the fault condition. Nine different fault types and normal operating conditions were simulated, including motor fault, propeller fault, low voltage fault, load loss fault, accelerometer fault, gyroscope fault, magnetometer fault, barometer fault, and GPS fault.

[0134] Figure 3 Figure 2 shows the effectiveness of LWPT-SGLE in removing sample noise. Traditional methods, such as the Kalman filter (KF), reconstruct signals that deviate significantly from the original signal, resulting in an oversmoothed effect. Deep learning methods, such as convolutional neural networks (CNNs), restore the original signal's trend but fail to effectively remove high-frequency noise, leading to overlapping reconstructed signals. Long short-term memory networks (LSTMs) experience memory decay when processing long sequences, resulting in periodic bumps in the reconstructed signal and suboptimal denoising. In contrast, LWPT-SGLE performs best across different signal segments, particularly in regions with sudden changes. LWPT-SGLE leverages the nonlinear properties of deep learning models to effectively handle complex noise patterns, resulting in denoised signals that closely resemble the original signal, accurately preserving important features and turning points.

[0135] In addition, the denoising effect of LWPT-SGLE on sample noise is quantitatively analyzed, such as Figure 4 As shown in the figure, by comparing the evaluation metrics (R², MSE, and MAE) of four different signal denoising methods, the LWPT-SGLE method achieves the best performance across all evaluation dimensions, with the R² value closest to 1 and the smallest MSE and MAE values, indicating that this method has the best signal fit and the lowest error level. The traditional KF method also achieves basic denoising results, but it is significantly lower than the other three methods in all metrics. The performance of the LSTM method and the CNN method is relatively close, and both outperform the traditional KF method, but still lags behind 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.

[0136] Figure 5The comparison of diagnostic accuracy 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 under various symmetry parameters. The SREA method has been significantly improved compared to CE, and the overall performance curve is parallel to but higher than CE. Mixup-BMM shows better robustness, its performance degradation curve is relatively gentle, and the degradation scale is slightly smaller. The performance advantage is more obvious when the level of symmetric label noise is large. Finally, the LWPT-SGLE method shows the best performance. Not only does it maintain the highest accuracy under various symmetric label noise parameters, but its performance degradation is also the gentlest, indicating that this method has significant advantages in dealing with problems of different degrees of symmetry. It is worth noting that in the case of asymmetric label noise, the performance ranking of each method remains consistent, and LWPT-SGLE is still better than other methods. This is due to the fact that the model achieves the highest accuracy through 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 As shown in the figure, under conditions of low symmetric label noise, the performance gap between single-scale learning and scale-integrated learning is relatively small. However, as the noise intensity increases, the gap gradually widens. This trend demonstrates the importance of the scale-integrated strategy, which effectively captures local and global features at multiple scales through the integrated learning process, thereby mitigating the interference effect of label noise during training.

[0137] The LWPT-SGLE model in this embodiment addresses the challenge of UAV fault diagnosis under the coupled interference of sample noise and label noise by proposing a two-stage collaborative optimization solution. In the first stage, a learnable wavelet packet transform (LWPT) subnetwork is constructed by injecting Gaussian white noise into the original flight signal to simulate real-world interference. Leveraging its ability to dynamically adjust thresholds at each node in 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, the denoised signal is first segmented using a multi-scale sliding window to increase sample diversity. Asymmetric label noise is then injected into the training set to simulate manual labeling errors. A graph embedding learning subnetwork (SGLE) is then constructed to implement noise-immune training. This approach generates multi-scale representations through downsampling and fusion learning. High-confidence samples are selected based on a minimum loss criterion. Finally, a graph structure is constructed to propagate label semantics, achieving dynamic correction of noisy labels and joint optimization of the feature space. This method forms a closed-loop optimization link from signal denoising to label enhancement through cross-level interaction between LWPT and SGLE, providing key technical support for the reliable operation and maintenance of UAVs in complex operating environments and rapid response scenarios.

[0138] The above are only embodiments of the present invention. Common knowledge such as the known specific structures and characteristics in the scheme are not described in detail here. Ordinary technicians in the field are aware of all common technical knowledge in the technical field of the invention before the application date or priority date, can obtain all existing technologies in the field, and have the ability to apply conventional experimental means before that date. Ordinary technicians in the field can improve and implement this scheme in combination with their own abilities under the inspiration given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several variations and improvements can be made, which should also be regarded as the scope of protection of the present invention. These will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. An intelligent UAV fault diagnosis method for sample-label noise coupling scenarios, characterized by: The following steps are involved: Step 1: Use sensors to obtain signals of the drone in normal operation and in motor failure, propeller failure, low voltage failure, load loss failure, accelerometer failure, gyroscope failure, magnetometer failure, barometer failure, and GPS failure states, 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 signal; the LWPT sub-network is an autoencoder architecture based on wavelet transform; The encoder part is used to decompose each node of the original signal into low-frequency and high-frequency parts, apply a learnable threshold at all nodes to suppress the coefficients related to noise, and learn parameters to adapt to the frequency characteristics of the input signal, removing noise by learning an appropriate dilution representation; The encoder obtains clean coefficients. Based on these clean coefficients, the decoder uses learnable transposed convolution to adaptively optimize the parameters in the reconstruction process, gradually reconstruct the signal, and perform signal denoising. Step 3: Use a sliding window to divide and expand the samples of the denoised signal; Step 4: Divide the training set and test set for fault diagnosis tasks, and inject label noise into the training set; The method of injecting label noise into the training set in step 4 is: To describe the probability of a clean label flipping to a noise label, 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 With the same probability, a certain class of clean label samples is flipped to the rest of the class labels. The asymmetric label noise is based on the asymmetric label noise transfer matrix , so that the samples of a certain category have a high similarity with another specific category; Step 5: Construct a scale graph to enhance the SGLE sub-network to reduce the impact of label noise during training. 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 minimum loss criterion; S5.

2. Sample Relabeling: For unselected samples, a graph embedding learning network component is constructed based on label propagation theory. The embedding learning network component uses clean label samples to perform label correction on the unselected samples. Step 6: Test the trained LWPT-SGLE model to test its fault diagnosis results under the influence of sample-label noise coupling.

2. The intelligent fault diagnosis method for unmanned aerial vehicle (UAV) in the sample-label noise coupling scenario according to claim 1 is characterized by: In step 1, the method of injecting Gaussian white noise is expressed as: in is the signal after injecting white noise is the original signal, represents the noise intensity, is Gaussian white noise that obeys the standard normal distribution.

3. The intelligent fault diagnosis method for unmanned aerial vehicle (UAV) in the sample-label noise coupling scenario according to claim 1 is characterized by: 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: in, Refers to the learnable bias; after decomposition, the decomposition coefficient of the jth node in the i-th layer is calculated by the following formula: : in represents inverse convolution, represents the convolution operation, Indicates getting the corresponding input from the previous node; In the decoder part, a learnable transposed convolution layer with a step size of 2 is used to gradually reconstruct the signal, and the clean coefficient is calculated by the following formula : in is the deconvolution kernel, Represents an upsampling operation; Perform deconvolution on the clean coefficients: Will Add together to get the complete reconstructed signal : Where L represents the number of LWPT decomposition layers.

4. The intelligent fault diagnosis method for unmanned aerial vehicle (UAV) in the sample-label noise coupling scenario according to claim 1 is characterized by: In the step 3, a sliding window is used to divide the denoised signal into samples, the size of the sliding window is 128, and the step size is 16.

5. The intelligent fault diagnosis method for unmanned aerial vehicle (UAV) in the sample-label noise coupling scenario according to claim 1 is characterized by: The S5.1 specifically includes the following steps: S5.1.1, the signal after LWPT denoising is expressed as , downsample the signal: in is the downsampling rate and belongs to between, The larger the value is, the larger the data scale of the original signal is. The sample embeddings in the embedding space are fused according to the scale from fine-grained to coarse-grained: in is the learned single-scale embedding, for Embedded, Refers to the concatenation operation of two vectors, Represents a nonlinear neural network used to fuse information at different scales; S5.1.

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

3. Sort the damage of each sample from small to large, and set the label noise rate to , corresponding to the loss Some samples are selected as clean samples.

6. The intelligent fault diagnosis method for unmanned aerial vehicle (UAV) in the sample-label noise coupling scenario according to claim 5 is characterized by: The S5.2 specifically includes the following steps: 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 updates early in the training process: in Refers to the number of rounds of current training, represents the momentum update parameter; S5.2.

3. Use the Gaussian similarity function to calculate the edge weights of the nearest neighbor graph: in Indicates the calculation of Euclidean distance, is a fixed parameter; S5.2.

4. Perform graph Laplacian operations on edge weights E: Where D is a diagonal matrix; S5.2.

5. According to label propagation theory, the following formula is used to calculate The pseudo label of each node in : Where F refers to the predicted pseudo label, is a hyperparameter between 0 and 1, and the corrected labels of the unselected samples are: Where argmax represents the pseudo label with the maximum value.

7. The intelligent fault diagnosis method for unmanned aerial vehicle (UAV) in the sample-label noise coupling scenario according to claim 1 is characterized by: In step 6, the fault diagnosis result of LWPT-SGLE is tested to verify the comparison between the denoised signal and the original signal, and to verify the fault diagnosis classification accuracy under the influence of label noise.

Images