Unmanned aerial vehicle motor bearing fault diagnosis method based on small sample data
By integrating mode decomposition and wavelet threshold function denoising, and combining deep belief network transfer learning and support vector machine, the problem of bearing fault diagnosis under small sample data of UAV motors is solved, achieving efficient fault identification and accurate classification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-28
- Publication Date
- 2026-05-19
AI Technical Summary
Under the condition of small sample vibration data of UAV motors, existing deep learning models have low generalization ability due to insufficient training data, making it difficult to effectively diagnose bearing faults. Moreover, noise in the vibration signal masks the fault characteristics, and existing signal denoising methods such as EEMD may cause signal distortion.
Signal denoising is performed using ensemble mode decomposition combined with wavelet thresholding to extract time-domain, frequency-domain, and time-frequency hybrid features. Fault diagnosis is then performed using deep belief network transfer learning and support vector machine. Specific steps include: signal denoising, feature extraction, deep autoencoder network pre-training, parameter transfer, target domain data fine-tuning, and final SVM classification.
It effectively handles bearing fault diagnosis with small sample data, improves diagnostic accuracy, overcomes the influence of noise, and achieves efficient identification of bearing faults in UAV motors.
Smart Images
Figure CN122065137A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of UAV motor fault diagnosis, specifically relating to a method for diagnosing UAV motor bearing faults based on small sample data. Background Technology
[0002] Deep learning has been frequently applied in the field of fault diagnosis due to its complex structure, powerful nonlinear capabilities, and multi-layered network structure, enabling it to extract features from data at a deep level, map the characteristics of objects, and even extract features from unlabeled data. However, training such models requires a large amount of fault data as the basis for determining model parameters, and the training and testing data must be marginally distributed. When the amount of data is insufficient, deep learning networks will not train sufficiently, resulting in low generalization ability, insufficient feature extraction, and difficulty in capturing the main features of the data. With the increasing reliability of UAV motor operation, collecting labeled fault data for motors is very difficult. Therefore, in practical applications, UAV motor fault diagnosis encounters the problem of insufficient labeled data, i.e., the small sample problem.
[0003] Vibration signals can reveal conditions such as impacts, oscillations, speed fluctuations, and structural deformations caused by faults. Vibrations caused by bearing faults may be masked by noise. The sources of noise are complex, including those from the signal acquisition system itself and noise from the coupling between various components under complex operating conditions. Excessive noise can obscure fault characteristics in the data. Ensemble Empirical Mode Decomposition (EEMD) algorithms are widely used in signal denoising due to their advantages of being adaptive, multi-resolution, and not requiring pre-selection of basis functions. However, EEMD reconstruction discards some intrinsic mode functions (IMFs), leading to signal distortion. Summary of the Invention
[0004] (a) Technical problems to be solved The technical problem to be solved by the present invention is: a method for diagnosing the bearing fault of a drone motor based on small sample data, which is used to diagnose the bearing fault of a drone motor under the condition of small sample vibration data of drone motor.
[0005] (II) Technical Solution To address the aforementioned technical problems, this invention provides a method for diagnosing UAV motor bearing faults based on small sample data, comprising the following steps: Step 1: Denoise the source domain data, extract time domain, frequency domain, and time-frequency domain hybrid features, and normalize the feature data before using it as input to the network model; Step 2: Use source domain feature data to pre-train a DBN, training only one RBM at a time, and using the output of the previous RBM as the input of the next RBM; Step 3: After all the autoencoder networks have been trained, they are connected into a deep autoencoder network. The entire deep autoencoder network is then fine-tuned using a small amount of training sample data to finally obtain the source domain DBN network model. Step 4: Transfer the DBN network parameters pre-trained in the source domain to the target domain as the initial parameters of the target domain network; Step 5: Denoise the target domain data and extract time domain, frequency domain, and time-frequency domain hybrid features. Fine-tune the target domain network model parameters using source domain data and target domain data. Step 6: Use the data after feature extraction by the DBN network as input for SVM classification to perform fault classification and identification.
[0006] The noise reduction process in step 1 specifically includes: S1: For signal Perform integrated mode decomposition to obtain IMF components; S2: Perform autocorrelation function calculation on the IMF; S3: Determine the noise-dominant component based on the calculated value of the autocorrelation function in step 2; S4: For the dominant noise component The wavelet threshold function is used to improve the algorithm for noise reduction; S5: Reconstruct the denoised components to obtain the denoised signal.
[0007] In step 1, the time-domain features are analyzed by analyzing the time-domain waveform of the acquired bearing signal and using time-domain indicators for fault diagnosis. The time-domain statistical indicators include mean, peak value, root mean square value, kurtosis, impulse index, waveform index, margin index, kurtosis, and skewness. The frequency domain features are used to represent the frequency structure of the signal. By extracting frequency domain information, the bearing fault frequency is discovered. The signal is then transformed to the frequency domain using Fourier transform, and information is obtained and features are extracted using frequency domain indices. The frequency domain indices and their calculation formulas are as follows: Center of gravity frequency:
[0008] Mean square frequency:
[0009] Root mean square frequency:
[0010] Frequency variance:
[0011] Where i is the sample number and N is the total number of samples, , The sampling frequency; The energy values of the 16 frequency bands after the wavelet packet is decomposed into 4 levels are used as time-frequency domain features.
[0012] Data normalization To facilitate the convergence of deep learning algorithms, the constructed time-frequency domain features need to be normalized using the following formula:
[0013] in These are the normalized eigenvalues. These are the eigenvalues before normalization. The minimum value among the feature values in the experimental data sample. It represents the maximum value among the feature values in the experimental data sample.
[0014] Improved feature dimensionality reduction mapping This method employs multi-core MMD (MKMMD) to further improve accuracy and reduce the relative error between transferable features. Typically, convex combinations... Used to provide efficient mapping estimates. The multi-kernel function is defined as follows.
[0015]
[0016] in These represent the trade-off parameters for different kernels. Based on multi-core functions, MK-MMD can be defined as follows:
[0017] in express The distance. This represents a multi-kernel function. A Gaussian kernel is chosen as the kernel function, and the number of kernels is set to 5. Following the median heuristic, the bandwidth of each Gaussian kernel is set to the median of the pairwise distances on the training data. It is trained by ensembling two objectives: (1) Cross-entropy loss between the predicted label and the true label of the sample; (2) The distance between the transferable features of the two domains. Minimizing the equation. Defined as the overall optimization objective.
[0018]
[0019]
[0020]
[0021] Where L represents the total loss function, The cross-entropy loss function of the classification layer is represented. Indicates the first The corresponding DBN network output for each data point For its corresponding tag, Represents the multi-kernel domain loss function. For the current training round, This represents the maximum value after training epochs. These are the weight parameters.
[0022] Deep Belief Network Learning Deep Belief Networks (DBNs) consist of multiple Restricted Bolzmann Machines (RBMs) stacked together. DBNs primarily train each RBM layer individually, mapping feature vectors to different feature spaces through mappings between each RBM.
[0023] SVM classification and recognition failure Support Vector Machines (SVMs) are linear classifiers defined on a feature space that maximize margin. For linearly inseparable cases, SVMs use kernel tricks to address non-linearity in classification problems. The strategy of SVMs is to solve a convex quadratic programming problem, namely, maximizing the margin. SVMs construct the optimal classifier plane by maximizing the margin, thereby achieving the goal of classifying the data.
[0024] (III) Beneficial Effects Compared with existing technologies, this invention has the following advantages: First, it proposes a vibration signal denoising method based on EEMD combined with an improved wavelet threshold function. By utilizing the adaptive features of EEMD decomposition, it can effectively process signal decomposition. Combined with the wavelet threshold denoising algorithm for processing frequency localization features, it can compensate for the shortcomings of the EEMD denoising algorithm and avoid signal distortion caused by discarding some IMFs during EEMD reconstruction. The improved algorithm first uses EEMD to decompose the noisy signal, obtaining multiple sets of IMFs. Then, it performs autocorrelation calculations on the decomposed IMFs, and determines whether the IMF is noise-dominated or signal-dominated based on the calculated autocorrelation coefficients. The noise-dominated IMFs are then denoised using the improved wavelet threshold function, and reconstructed with the signal-dominated IMFs to obtain the denoised signal.
[0025] This paper proposes a fault diagnosis method based on deep belief network transfer learning. In practical applications, the reliability of UAV motor operation is constantly improving, resulting in a limited amount of collected motor bearing fault label data, which cannot meet the requirement of a large amount of fault data for deep learning-based fault diagnosis methods. Transfer learning does not require the training and testing data to have the same distribution; therefore, it can learn knowledge from related domains and complete classification tasks, transferring the knowledge learned in the source domain to the target domain. However, the inconsistent distribution of data features across domains prevents the direct transfer of model parameters from the source domain to the target domain. This method, based on feature transfer, utilizes the distribution difference between the target and source domain data to fine-tune the network model parameters in the source domain. This allows the DBN network to map data features from high-dimensional to low-dimensional identically distributed features. The difference in the maximum mean of multiple kernels is used to reduce the distance between the two domains, ensuring that the same fault features mapped to the low-dimensional source and target domain data tend to have the same distribution. Attached Figure Description
[0026] Figure 1 Here is a flowchart of the noise reduction algorithm; Figure 2 Time-domain plots of the original signal and the noisy signal Figure 3 This is an overall flowchart of the present invention; Figure 4 This is a diagram showing the direct migration classification results of the present invention. Figure 5 This is a diagram showing the direct SVM classification results of this invention; Figure 6 This is a diagram showing the classification results of the present invention. Detailed Implementation
[0027] To make the objectives, contents, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.
[0028] Example 1 This embodiment presents a method for diagnosing UAV motor bearing faults based on small sample data, including the following steps: Step 1: Denoise the source domain data, extract time domain, frequency domain, and time-frequency domain hybrid features, and normalize the feature data before using it as input to the network model; Step 2: Use source domain feature data to pre-train a DBN, training only one RBM at a time, and using the output of the previous RBM as the input of the next RBM; Step 3: After all the autoencoder networks have been trained, they are connected into a deep autoencoder network. The entire deep autoencoder network is then fine-tuned using a small amount of training sample data to finally obtain the source domain DBN network model. Step 4: Transfer the DBN network parameters pre-trained in the source domain to the target domain as the initial parameters of the target domain network; Step 5: Denoise the target domain data and extract time domain, frequency domain, and time-frequency domain hybrid features. Fine-tune the target domain network model parameters using source domain data and target domain data. Step 6: Use the data after feature extraction by the DBN network as input for SVM classification to perform fault classification and identification.
[0029] The noise reduction process in step 1 specifically includes: S1: For signal Perform integrated mode decomposition to obtain IMF components; S2: Perform autocorrelation function calculation on the IMF; S3: Determine the noise-dominant component based on the calculated value of the autocorrelation function in step 2; S4: For the dominant noise component The wavelet threshold function is used to improve the algorithm for noise reduction; S5: Reconstruct the denoised components to obtain the denoised signal.
[0030] To demonstrate the effectiveness of this method in denoising bearing fault signals, simulations of bearing outer ring damage were used for verification. When an outer ring fault occurs in a bearing, the damage caused by the fault leads to a higher fault frequency. The continuous generation of these pulses will cause high-frequency damped vibrations. Assume the natural frequency caused by the pulse is... Fault characteristic frequency That is, the period of pulse train generation sampling frequency Therefore, the damped vibration caused by a single pulse can be derived as follows:
[0031] In the formula, The natural angular frequency of vibration. , The number of sampling points for simulating outer ring faults is set to 2048.
[0032] To simulate the noise content in real-world bearing vibration signals, Gaussian white noise with a signal-to-noise ratio of 6 was added to the simulated bearing fault characteristic signals. In wavelet denoising, a decomposition level of 4 was selected for the wavelet transform, and the wavelet basis was chosen as "sym4". The original simulated bearing signal and the signal after adding noise are shown below. Figure 2 As shown.
[0033] As can be seen from Table 1, this method has the highest signal-to-noise ratio, indicating that this method can have a good noise reduction effect on vibration signals.
[0034] Table 1 Signal-to-noise ratio
[0035] In step 1, the time-domain features are analyzed by analyzing the time-domain waveform of the acquired bearing signal and using time-domain indicators for fault diagnosis. The time-domain statistical indicators include mean, peak value, root mean square value, kurtosis, impulse index, waveform index, margin index, kurtosis, and skewness. The frequency domain features are used to represent the frequency structure of the signal. By extracting frequency domain information, the bearing fault frequency is discovered. The signal is then transformed to the frequency domain using Fourier transform, and information is obtained and features are extracted using frequency domain indices. The frequency domain indices and their calculation formulas are as follows: Center of gravity frequency:
[0036] Mean square frequency:
[0037] Root mean square frequency:
[0038] Frequency variance:
[0039] Where i is the sample number and N is the total number of samples, , The sampling frequency; The energy values of the 16 frequency bands after the wavelet packet is decomposed into 4 levels are used as time-frequency domain features.
[0040] Data normalization To facilitate the convergence of deep learning algorithms, the constructed time-frequency domain features need to be normalized using the following formula:
[0041] in These are the normalized eigenvalues. These are the eigenvalues before normalization. The minimum value among the feature values in the experimental data sample. It represents the maximum value among the feature values in the experimental data sample.
[0042] Improved feature dimensionality reduction mapping This method employs multi-core MMD (MKMMD) to further improve accuracy and reduce the relative error between transferable features. Typically, convex combinations... Used to provide efficient mapping estimates. The multi-kernel function is defined as follows.
[0043]
[0044] in These represent the trade-off parameters for different kernels. Based on multi-core functions, MK-MMD can be defined as follows:
[0045] in express The distance. This represents a multi-kernel function. A Gaussian kernel is chosen as the kernel function, and the number of kernels is set to 5. Following the median heuristic, the bandwidth of each Gaussian kernel is set to the median of the pairwise distances on the training data. It is trained by ensembling two objectives: (1) Cross-entropy loss between the predicted label and the true label of the sample; (2) The distance between the transferable features of the two domains. Minimizing the equation. Defined as the overall optimization objective.
[0046]
[0047]
[0048]
[0049] Where L represents the total loss function, The cross-entropy loss function of the classification layer is represented. Indicates the first The corresponding DBN network output for each data point For its corresponding tag, Represents the multi-kernel domain loss function. For the current training round, This represents the maximum value after training epochs. These are the weight parameters.
[0050] Deep Belief Network Learning Deep Belief Networks (DBNs) consist of multiple Restricted Bolzmann Machines (RBMs) stacked together. DBNs primarily train each RBM layer individually, mapping feature vectors to different feature spaces through mappings between each RBM.
[0051] SVM classification and recognition failure Support Vector Machines (SVMs) are linear classifiers defined on a feature space that maximize margin. For linearly inseparable cases, SVMs use kernel tricks to address non-linearity in classification problems. The strategy of SVMs is to solve a convex quadratic programming problem, namely, maximizing the margin. SVMs construct the optimal classifier plane by maximizing the margin, thereby achieving the goal of classifying the data.
[0052] Deep Belief Networks (DBNs) consist of multiple Restricted Bolzmann Machines (RBMs) stacked together. DBNs primarily train each RBM layer individually, mapping feature vectors to different feature spaces through mappings between each RBM. The following sections introduce RBMs and DBN networks separately.
[0053] Restricted Boltzmann Machine Boltzmann machines (BMs) are a type of neural network invented in 1985 by computer scientists Jeffrey Hinton and Metsrišeženowski. According to numerous studies, feedback neural networks can lead to pattern recognition errors if they get stuck in local optima. Boltzmann machines avoid the drawbacks of local optima by allowing each unit to follow a probability distribution, thus resolving local optima after a state change. The connection weights between units in a Boltzmann machine are symmetric and there are no self-connections. Furthermore, the output of each unit must be either 0 or 1. These characteristics are similar to Hopfield neural networks, but the biggest difference is that the output of a Hopfield network is determined by a certain determinism, while the output of a Boltzmann machine is determined by a probability distribution, as shown in the equation:
[0054] k and T represent specific coefficients, which increase as they grow. The probability of 1 doesn't change significantly, but as it decreases, the slope of the curve near 0 increases rapidly. As T approaches infinity, regardless of... How much is it equal to? The probability of a value being 1 or 0 is equal. Like Hopfield networks, Boltzmann machines (BMs) are interconnected networks; their units are connected via connection weights, and the training process is the same as for Hopfield networks. Training Boltzmann machines with hidden variables is very difficult. Therefore, Hinton proposed adding the condition of "no connection between units within a layer" to Boltzmann machines, resulting in Restricted Boltzmann Machines (RBMs). A Restricted Boltzmann Machine has a two-layer structure: a visible layer and a hidden layer, composed of visible and hidden variables, respectively. The visible and hidden layers are connected, but units within the same layer are not connected. A Restricted Boltzmann Machine is a special type of Boltzmann machine, a probabilistic undirected graphical model composed of visible and hidden layers, characterized by symmetric connectivity and no self-feedback.
[0055] Based on the structural characteristics of Boltzmann Machines (RBMs), it is known that RBMs are suitable for dimensionality reduction, classification, collaborative filtering, feature extraction, and other applications. Their overall structure is simple, yet highly versatile. They can choose between supervised or unsupervised learning methods depending on the task. RBMs play two important roles: first, encoding data, which can be understood as feature extraction; and second, obtaining the initial weights for training the neural network.
[58] .
[0056] Deep Belief Network The training method of DBN differs significantly from other multilayer neural networks. Other neural networks begin training by defining the network structure and updating weights and biases based on the error at the top layer. The core idea is to propagate the error to the next layer using backpropagation, adjusting all weights and biases accordingly. DBN, however, trains and adjusts layer by layer, using a contrastive divergence algorithm to adjust the weights and biases of each layer. Specifically, it trains the connection weights between the input and hidden layers first. After training a layer, its output is used as the input for the next layer, and the connection weights are adjusted again. This iterative process completes the DBN training. This layer-by-layer training process is called pre-training. DBN trains each Restricted Boltzmann Machine layer individually and unsupervised. Through layer-to-layer mapping, feature vectors can be mapped from low-dimensional to high-dimensional space, and vice versa, preserving as many uncorrelated features as possible. DBN can be used as both a generative and discriminative model. When used as a generative model, the network generates training data according to a certain probability distribution. When used as a discriminative model, a classification network layer needs to be added to the top of the DBN network to achieve classification functionality. A DBN network cannot function as a discriminative model independently; a classification network must be added at its top for data classification. A supervised DBN network typically consists of several layers of Restricted Block Models (RBMs) and a classification network. The hidden layers of the last RBM layer are connected to the classification network, using the feature vector output by the last RBM layer as the input feature vector for the classification network. Classification training is then performed in the classification network. Each RBM layer is trained to ensure that the mapping of its feature vectors is optimal, but this does not guarantee that the feature mapping of the entire DBN network is optimal because the backpropagation algorithm transmits error information from top to bottom to each RBM layer; therefore, fine-tuning is necessary. Thus, the RBM training process can be viewed as the parameter initialization process in a feedforward neural network. This pre-training method overcomes the problems of excessively long training cycles and models getting trapped in local optima due to random parameters. When using a DBN network as a discriminative model, a classification layer is added at the top. Classification requires training samples and the desired output; this is the supervised training mode. Except for the classification layer, all other layers are trained using unsupervised learning. The trained parameters are used as initial values, and the backpropagation algorithm is used to train the classification network, which includes the top layer. Verification test Data source: The source domain data comes from the bearing data of a certain type of UAV motor in our company. Bearing data with a sampling frequency of 12000 Hz was selected, and 500 points were extracted as one sample. As shown in Table 2, a total of 160 samples were selected.
[0057] Table 2 Source Domain Data
[0058] The target domain data comes from the bearing data of another type of drone motor in the company. Vibration signals collected by an accelerometer are used as the target domain dataset, with a sampling frequency of 1000 Hz. In the experiment, single-point corrosion damage caused by electrical discharge machining is used to simulate bearing failure. 500 data points are selected as one sample, and the data selection is shown in Table 3.
[0059] Table 3 Target Domain Data
[0060] Parameter details: This method employs a deep feature extraction network (DBN) composed of three layers of Restricted Baseline Model (RBM), with a network structure of 29-11-9-3. In the experiments, the pre-training phase had 100 iterations, the fine-tuning phase had 200 iterations, the learning rate was 0.001, and the momentum term learning rate was 0.5. Of the 160 datasets, 120 were training data and 40 were test data.
[0061] Experimental results: To verify the superiority of the transfer method presented in this chapter, comparative experiments were conducted with SVM classification using small sample data from the target domain and methods based on parameter transfer. The results are as follows: Figure 4 , 5 As shown in Figure 6 from Figure 4 , 5 The experimental results in section 6 show that when using actual UAV motor data for fault diagnosis, the first two methods have lost their ability to identify faults and can hardly distinguish between faulty and normal data. However, the improved method presented in this paper maintains a high diagnostic accuracy. The above experiments verify that the fault diagnosis method based on feature dimensionality reduction mapping of improved DBN parameters using MK-MMD, as proposed in this paper, has higher diagnostic accuracy and better classification performance.
[0062] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A method for diagnosing unmanned aerial vehicle (UAV) motor bearing faults based on small sample data, characterized in that, Includes the following steps: Step 1: Denoise the source domain data, extract time domain, frequency domain, and time-frequency domain hybrid features, and normalize the feature data before using it as input to the network model; Step 2: Use source domain feature data to pre-train a DBN, training only one RBM at a time, and using the output of the previous RBM as the input of the next RBM; Step 3: After all the autoencoder networks have been trained, they are connected into a deep autoencoder network. The entire deep autoencoder network is then fine-tuned using a small amount of training sample data to finally obtain the source domain DBN network model. Step 4: Transfer the DBN network parameters pre-trained in the source domain to the target domain as the initial parameters of the target domain network; Step 5: Denoise the target domain data and extract time domain, frequency domain, and time-frequency domain hybrid features. Fine-tune the target domain network model parameters using source domain data and target domain data. Step 6: Use the data after feature extraction by the DBN network as input for SVM classification to perform fault classification and identification.
2. The method for diagnosing UAV motor bearing faults based on small sample data as described in claim 1, characterized in that, The noise reduction process in step 1 specifically includes: S1: For signal Perform integrated mode decomposition to obtain IMF components; S2: Perform autocorrelation function calculation on the IMF; S3: Determine the noise-dominant component based on the calculated value of the autocorrelation function in step 2; S4: For the dominant noise component The wavelet threshold function is used to improve the algorithm for noise reduction; S5: Reconstruct the denoised components to obtain the denoised signal.
3. The method for diagnosing UAV motor bearing faults based on small sample data as described in claim 1, characterized in that, In step 1, the time-domain features are analyzed by analyzing the time-domain waveform of the acquired bearing signal and using time-domain indicators for fault diagnosis. The time-domain statistical indicators include mean, peak value, root mean square value, kurtosis, impulse index, waveform index, margin index, kurtosis, and skewness. The frequency domain features are used to represent the frequency structure of the signal. By extracting frequency domain information, the bearing fault frequency is discovered. The signal is then transformed to the frequency domain using Fourier transform, and information is obtained and features are extracted using frequency domain indices. The frequency domain indices and their calculation formulas are as follows: Center of gravity frequency: Mean square frequency: Root mean square frequency: Frequency variance: Where i is the sample number and N is the total number of samples, , The sampling frequency.