Unmanned aerial vehicle rotor bearing fault diagnosis method and system based on multi-source data

Abstract

The invention discloses an unmanned aerial vehicle rotor bearing fault diagnosis method and system based on multi-source data, and relates to the technical field of bearing fault diagnosis, and the method comprises the steps: obtaining a to-be-diagnosed sample during the operation of an unmanned aerial vehicle rotor bearing; extracting a multi-source fusion implicit code feature vector, inputting the multi-source fusion implicit code feature vector into a fault diagnosis learning device, and outputting a fault diagnosis result; the training method of the feature extractor and the fault diagnosis learner comprises the following steps: collecting labeled samples and unlabeled samples; constructing an integrated model; performing expert marking on K unlabeled samples with the highest comprehensive score, and supplementing the K unlabeled samples to a set of labeled samples; predicting a false label for the unmarked label-free sample and obtaining a false label sample; the pseudo label samples are incorporated into a training set, wherein the training set further comprises label samples and marked non-label samples; and training the integrated model. According to the method, the problems of insufficient utilization of multi-source features, serious waste of unlabeled samples and insufficient representativeness of the unlabeled samples are solved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of bearing fault diagnosis, in particular to a method for diagnosing faults of rotor bearings of unmanned aerial vehicles (UAVs) by using multi-source data. BACKGROUND

[0002] An unmanned aerial vehicle (UAV) is a kind of unmanned equipment frequently used in current life and production. Since no space for a pilot to sit is reserved and no safety of the pilot needs to be considered, the size of the UAV can be very small, greatly reducing the manufacturing cost and the difficulty of use.

[0003] The most common UAV at present is a multi-rotor UAV. This kind of UAV adopts the flight mode of a traditional helicopter and has a low requirement for the ground on which the UAV takes off and lands, and is very convenient to use. The rotor can generate lift when driven by a motor, and then drive the UAV to perform actions such as taking off, landing, accelerating and decelerating. The driving motor of the rotor of the UAV is generally at a high speed, usually several thousand revolutions per minute or even tens of thousands of revolutions per minute. In order to ensure the stable operation of the rotor of the UAV, a high requirement is put forward for the stability of the rotor bearing. Regardless of the structure or material of the rotor bearing, it is inevitable that the rotor bearing will fail. At the initial stage of the failure, the failure is not obvious and is easy to be ignored, and then serious loss is caused. Therefore, it is very important to accurately and timely diagnose the failure of the rotor bearing. CN120234570A provides a failure prediction method for a rotor bearing of an unmanned aerial vehicle. The patent embeds unlabeled training samples into labeled training samples after the unlabeled training samples are manually labeled, and realizes the use of the unlabeled samples.

[0004] However, the above patent has the following three problems: (1) The multi-source characteristics of the data are not fully utilized. When the rotor bearing fails, multi-source data such as vibration and temperature will change at the same time. There is a potential relationship between these changes. The above patent does not consider the correlation between these multi-source data, resulting in insufficient use of the data.

[0005] (2) There is a serious waste of unlabeled samples. The above patent only labels the top K unlabeled samples, and the remaining unlabeled samples are discarded, resulting in a serious waste of data.

[0006] (3) The representative of the selected unlabeled samples is insufficient. The above patent only scores by a ranker, and the selected unlabeled samples may have the problems of repetition, similarity or low value. The representative of these data is insufficient, which easily leads to the redundancy of the workload of the experts when labeling. SUMMARY

[0007] The embodiment of the present application provides a method and system for diagnosing faults of a rotor bearing of a UAV based on multi-source data, to solve the above three problems existing in the prior art.

[0008] In one aspect, the embodiment of the present application provides a method for diagnosing faults of a rotor bearing of a UAV based on multi-source data, comprising: Obtaining multi-source data of the rotor bearing of the UAV during operation, preprocessing the multi-source data, and obtaining corresponding to-be-diagnosed samples; Using a feature extractor to extract a multi-source fusion hidden code feature vector of the to-be-diagnosed sample, inputting the multi-source fusion hidden code feature vector into a fault diagnosis learner, and outputting a fault diagnosis result; The training method of the feature extractor and the fault diagnosis learner is as follows: Collecting multi-dimensional data of the rotor bearing of the UAV during operation, and dividing the multi-dimensional data into labeled samples and unlabeled samples according to whether the fault condition of the multi-dimensional data is known; Building an integrated model, the integrated model comprising a feature extractor, a ranker and a fault diagnosis learner; Calculating a correlation feature vector of the unlabeled sample, supplementing the correlation feature vector to the original features of the unlabeled sample, and forming an enhanced feature vector; Inputting the enhanced feature vector into the ranker to obtain a corresponding ranking value; inputting the unlabeled sample into the feature extractor and the fault diagnosis learner in turn to obtain a corresponding prediction entropy; clustering the enhanced feature vector to obtain a clustering representative; calculating the similarity between the unlabeled sample and a correlation mode template to determine multi-source correlation significance; Weighted summing the ranking value, the prediction entropy, the clustering representative and the multi-source correlation significance to obtain a comprehensive score, marking the K unlabeled samples with the highest comprehensive score by an expert, and supplementing the labeled unlabeled samples to the set of labeled samples; Predicting a pseudo label for the unlabeled sample that is not marked by the expert by using the fault diagnosis learner to obtain a pseudo labeled sample; if the pseudo labeled sample is consistent with the state correlation rule template corresponding to the pseudo label, the pseudo labeled sample is included in a training set, and the training set further comprises the labeled samples and the labeled unlabeled samples; Training the integrated model by using the training set to obtain the trained feature extractor and the fault diagnosis learner.

[0009] In another aspect, the embodiment of the present application further provides a system for diagnosing faults of a rotor bearing of a UAV based on multi-source data, comprising: A to-be-diagnosed data acquisition module, configured to obtain multi-source data of the rotor bearing of the UAV during operation, preprocess the multi-source data, and obtain corresponding to-be-diagnosed samples; The fault diagnosis module is configured to extract a multi-source fusion hidden code feature vector of a sample to be diagnosed by using the feature extractor, input the multi-source fusion hidden code feature vector into the fault diagnosis learner, and output a fault diagnosis result. The training module for training the feature extractor and the fault diagnosis learner comprises: The training data acquisition submodule is configured to collect multi-dimensional data of the unmanned aerial vehicle rotor bearing during operation, and divide the multi-dimensional data into labeled samples and unlabeled samples according to whether the fault condition of the multi-dimensional data is known. The model construction submodule is configured to construct an integrated model, and the integrated model comprises the feature extractor, the ranker, and the fault diagnosis learner. The correlation enhancement submodule comprises calculating a correlation feature vector of the unlabeled sample, supplementing the correlation feature vector to the original features of the unlabeled sample, and forming an enhanced feature vector. The sample selection index construction submodule is configured to input the enhanced feature vector into the ranker to obtain a corresponding ranking value, input the unlabeled sample into the feature extractor and the fault diagnosis learner in sequence to obtain a corresponding prediction entropy, cluster the enhanced feature vector to obtain a clustering representative, calculate the similarity between the unlabeled sample and the correlation mode template to determine the multi-source correlation significance. The sample labeling submodule is configured to weight and sum the ranking value, the prediction entropy, the clustering representative, and the multi-source correlation significance to obtain a comprehensive score, label the K unlabeled samples with the highest comprehensive scores by experts, and supplement the labeled unlabeled samples to the set of labeled samples. The training set construction submodule is configured to predict pseudo-labels for the unlabeled samples that are not labeled by experts by using the fault diagnosis learner to obtain pseudo-labeled samples, and if the pseudo-labeled samples are consistent with the state correlation rule templates corresponding to the pseudo-labels, the pseudo-labeled samples are included in the training set, and the training set further comprises the labeled samples and the labeled unlabeled samples. The training submodule is configured to train the integrated model by using the training set to obtain the trained feature extractor and the fault diagnosis learner.

[0010] The unmanned aerial vehicle rotor bearing fault diagnosis method and system based on multi-source data in the present application have the following advantages: (1) Solving the problem of insufficient utilization of multi-source features. The present application calculates the correlation features such as mutual information and Pearson coefficient between multi-source data such as vibration, temperature, and working condition, constructs an enhanced feature system of single-source features + correlation features, and then strengthens the key correlation information related to faults through an attention mechanism. This design enables the model not only to identify the explicit fault features of single-source data, but also to capture the implicit rules of multi-source data linkage, greatly improving the recognition ability of early faults such as mild faults.

[0011] (2) Solve the problem of serious waste of unlabeled samples. The application makes full use of unlabeled samples through pseudo-label strategy: high-confidence pseudo-label samples are included in the training set after correlation consistency verification; low-confidence samples are not used for fault diagnosis training, but are used to optimize the reconstruction ability of the VAE (Variational Autoencoder) network. Let all unlabeled samples play a value, not only expand the training data size, but also avoid data resource waste, and improve the model generalization without increasing the labeling cost.

[0012] (3) Make up for the lack of representative of unlabeled samples. The application constructs a four-dimensional sampling index, which integrates multi-source correlation significance, model prediction entropy, ranking value and clustering representativeness, which not only ensures that the selected samples contain key fault information, but also covers different data distribution areas through clustering analysis to avoid sample redundancy. The selected samples are more representative, so that each expert annotation can maximize the supplement of model knowledge gap and significantly reduce the labeling cost. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0014] Figure 1 The flowchart of the unmanned aerial vehicle rotor bearing fault diagnosis method based on multi-source data provided by the embodiment of the present application. DETAILED DESCRIPTION

[0015] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0016] Figure 1 The flowchart of the unmanned aerial vehicle rotor bearing fault diagnosis method based on multi-source data provided by the embodiment of the present application. The embodiment of the present application provides an unmanned aerial vehicle rotor bearing fault diagnosis method based on multi-source data, which comprises: S100, acquiring multi-source data of unmanned aerial vehicle rotor bearing operation, pre-processing the multi-source data to obtain corresponding samples to be diagnosed.

[0017] Exemplarily, the multi-source data includes vibration signal and temperature data of the rotor bearing, and flight parameter and sensor state data of the unmanned aerial vehicle.

[0018] Specifically, the vibration signal can be collected by an acceleration sensor installed near the rotor bearing seat, which covers X, Y, Z three directions, and the sampling frequency is set to 20 kHz to ensure the capture of high-frequency fault characteristics, and the bearing surface temperature data is collected by a temperature sensor, the sampling frequency is 1 Hz, and the temperature change trend with running time is recorded.

[0019] The real-time flight parameters can be obtained from the UAV flight control system, including flight speed, load size, environmental temperature and altitude. The sensor state data includes sensor power supply voltage and signal transmission stability index, which is used to exclude abnormal data interference caused by sensor failure.

[0020] The preprocessing of multi-source data includes the following contents: Noise filtering: wavelet filtering algorithm is used to denoise the vibration signal, by decomposing the vibration signal into different scales, the effective signal components related to bearing failure (such as fault impact signal) are retained, and the high-frequency noise caused by electromagnetic interference and airflow disturbance is filtered out; moving average filtering is used for temperature data to smooth short-term fluctuations and highlight temperature trends.

[0021] Time axis alignment: taking the timestamp of the UAV flight control system as the reference, the vibration signal, temperature data and flight data are time-synchronized. For example, the temperature data collected at 1 Hz is matched with the vibration signal collected at 20 kHz according to the time node, to ensure that the multi-source data at the same time are corresponding and associated, and to avoid the deviation caused by time misalignment.

[0022] Data normalization: all data are normalized by Min-Max normalization to map different dimensional data to the interval [0, 1]. For example, the vibration signal amplitude and flight speed are normalized to eliminate the influence of dimensional difference on model training, so that each dimension of data has equal weight contribution.

[0023] S110, using a feature extractor to extract a multi-source fusion hidden code feature vector of the sample to be diagnosed, inputting the multi-source fusion hidden code feature vector into a fault diagnosis learner, and outputting a fault diagnosis result.

[0024] Exemplarily, after obtaining the fault diagnosis result, the fault diagnosis result is compared with the UAV operation log and the artificial inspection result, if the fault diagnosis result does not match the actual state, the current sample is marked as a diagnosis error sample and supplemented to the set of unlabeled samples, and the next round of iteration optimization is started.

[0025] The training method of the feature extractor and the fault diagnosis learner is as follows: S200, collect multi-dimensional data of the unmanned aerial vehicle rotor bearing during operation, and divide the multi-dimensional data into labeled samples and unlabeled samples according to whether the fault condition is known.

[0026] Exemplarily, the multi-dimensional data and the multi-source data have the same data content, the multi-dimensional data is preprocessed by adopting the same preprocessing procedure as the multi-source data, and after preprocessing, the preprocessed multi-dimensional data is segmented into independent samples according to a fixed time window, and each sample contains a vibration signal segment, a temperature average, a flight parameter instantaneous value and sensor state data in the time window.

[0027] According to whether the fault condition is known, the samples in the sample set are divided into labeled samples and unlabeled samples, wherein the labeled samples are clearly labeled with fault types such as wear, crack and normal, and fault degrees such as mild, moderate and severe, and the unlabeled samples are not labeled with fault-related information.

[0028] S210, constructing an integrated model, the integrated model comprising a feature extractor, a ranker and a fault diagnosis learner.

[0029] Exemplarily, the integrated model further comprises a VAE network and a discriminator, and after the integrated model is constructed, the feature extractor, the ranker, the fault diagnosis learner, the VAE network and the discriminator in the integrated model are parameterized and pre-trained.

[0030] Specifically, the feature extractor adopts a hybrid structure of CNN (Convolutional Neural Network) + attention mechanism. The CNN part is composed of 3 convolutional layers and 2 pooling layers, which is used to extract local features in multi-dimensional data, such as fault impact features of vibration signals and mutation features of temperature data. The attention mechanism part is used to strengthen the weight of the correlation features of multi-dimensional data, so that the model pays more attention to the key correlation information for fault diagnosis. The core function of the feature extractor is to compress high-dimensional multi-dimensional data into low-dimensional and compact hidden code feature vectors, while retaining the core information and correlation rules of the data.

[0031] The ranker adopts a 3-layer fully connected network structure, the input is the original features of the unlabeled samples, and the output is a single numerical value, i.e. the ranking value. Its core function is to sort the unlabeled samples according to the fault information value contained in the samples, and the higher the ranking value, the more likely the sample contains key fault features or information that the model does not grasp.

[0032] The fault diagnosis learner is composed of 4 layers of fully connected network and 1 layer of time sequence fusion layer. The time sequence fusion layer is used to capture the time sequence evolution features of the bearing fault, and the fully connected network is used to map the hidden code feature vector to the fault diagnosis result. Its core function is to predict the fault type and fault degree of the bearing according to the hidden code features output by the feature extractor.

[0033] The VAE network comprises an encoder and a decoder, both of which are 3-layer fully connected network structures. The function of the encoder is to map the sample features to the parameters of the latent space distribution, i.e., the mean and variance, and the function of the decoder is to reconstruct the original sample data according to the latent space parameters. The core function of the VAE network is to optimize the feature extraction quality through the reconstruction loss, and to align the data distribution through the adversarial training.

[0034] The discriminator adopts a 2-layer fully connected network structure, and the input is the fusion feature of the latent code feature vector and the ranking information, and the output is a binary classification result, i.e., whether the feature comes from a labeled sample or an unlabeled sample. The core function is to promote the feature extractor to learn more discriminative general features through adversarial training.

[0035] Randomly initializing the parameters of each module in the integrated model can ensure that the parameter values are within a reasonable range. In the pre-training process, the initial labeled samples are used to pre-train the model, and the training rounds are set to 50 rounds and the batch size is set to 32. The core goal of pre-training is to enable the model to have preliminary fault recognition capability. The specific process is as follows: input the labeled samples into the feature extractor to obtain the latent code feature vector, input the latent code feature vector into the fault diagnosis learner to obtain the fault prediction result, and update the parameters of the feature extractor and the fault diagnosis learner by calculating the difference between the prediction result and the true label, i.e., the fault diagnosis loss. After pre-training, the recognition accuracy of the model for known fault types needs to reach more than 60%, which lays a foundation for subsequent iterative optimization.

[0036] In S220, the correlation feature vector of the unlabeled sample is calculated, and the correlation feature vector is supplemented to the original feature of the unlabeled sample to form an enhanced feature vector.

[0037] For example, based on mutual information and Pearson coefficient, the correlation indicators between different dimensions of data in the unlabeled sample are calculated, and all the correlation indicators are spliced to form a correlation feature vector.

[0038] The running state of the rotor bearing of the unmanned aerial vehicle is strongly related to multi-dimensional data. Under normal conditions, the correlation between the data is stable, such as the weak positive correlation between the vibration amplitude and the temperature, and the slight increase in the vibration amplitude when the flight speed increases. Under fault conditions, the correlation will change significantly, such as the strong positive correlation between the vibration amplitude and the temperature in the wear fault, and the negative correlation between the vibration signal and the load in the crack fault, and the greater the load, the more intense the vibration amplitude fluctuation. By calculating the correlation between multi-dimensional data, the potential state information behind the sample can be mined; Two complementary correlation analysis methods are used to calculate the correlation, which can fully capture the correlation patterns of multi-dimensional data: Mutual information: used to measure the degree of dependence between two data variables, suitable for data pairs with strong nonlinear correlation, such as vibration signal and temperature, vibration signal and load. For example, calculate the mutual information of the X-direction vibration signal and the bearing surface temperature. The higher the mutual information value, the stronger the dependence between the two, and the more likely it is to reflect a specific operating state. Pearson coefficient: used to measure the degree of linear correlation between two data variables, suitable for data pairs with obvious linear correlation, such as flight speed and vibration amplitude, and environmental temperature and bearing temperature. The value of the Pearson coefficient ranges from -1 to 1. A value close to 1 indicates strong positive correlation, a value close to -1 indicates strong negative correlation, and a value close to 0 indicates no linear correlation.

[0039] For each unlabeled sample, calculate the correlation indicators of all key data pairs, including vibration-temperature, vibration-flight speed, vibration-load, temperature-environmental temperature, and other 8 core data pairs. Concatenate these correlation indicators into a correlation feature vector (dimension 8), and supplement this vector to the original features of the sample to form an enhanced feature vector (dimension: original feature dimension + 8) of original features + correlation features, providing a more comprehensive basis for subsequent sample selection.

[0040] S230, input the enhanced feature vector into the ranker to obtain the corresponding ranking value; input the unlabeled sample into the feature extractor and the fault diagnosis learner in turn to obtain the corresponding prediction entropy; cluster the enhanced feature vector to obtain the clustering representation; calculate the similarity between the unlabeled sample and the correlation mode template to determine the multi-source correlation significance.

[0041] Exemplarily, input the enhanced feature vector of the unlabeled sample into the ranker, and the ranker outputs the ranking value of each sample according to the pre-trained parameters θ_R. The calculation process of the ranking value is: the ranker learns the feature mode of the labeled sample to score the probability of containing fault information in the unlabeled sample. The higher the score, the more likely the sample contains fault features that the model has not fully learned.

[0042] Entropy is an index for measuring uncertainty, and model prediction entropy is used to quantify the uncertainty of the model in judging the fault state of the unlabeled sample. The calculation method is: input the unlabeled sample into the pre-trained feature extractor + fault diagnosis learner to obtain the prediction probability distribution of each fault type, and the prediction entropy is calculated through the probability distribution. The higher the entropy value, the more uncertain the model is in judging the fault state of the sample, and the greater the value of the sample in optimizing the model after labeling.

[0043] The clustering representation is used to ensure that the screened samples cover different data distribution areas, avoiding repeated labeling of similar samples. The K-Means clustering algorithm is used to cluster the enhanced feature vectors of all unlabeled samples, and the number of clusters is set to 1 / 10 of the total number of unlabeled samples, and each cluster corresponds to a data distribution area. The distance of each sample to the cluster center is calculated, and the closer the distance, the more representative the sample is of the core features of the cluster, and the stronger the clustering representation.

[0044] The multi-source association significance is used to evaluate the matching degree of the sample association features and the known fault state association mode, and the core is to judge whether the association features of the sample belong to the key association mode that the model does not master. First, based on the initial labeled samples, the association mode templates of the normal state and each type of fault state are constructed; then, the similarity of the association features of the unlabeled samples and each template is calculated, if the similarity of the association features of the sample to all templates is lower than 0.5, it means that its association mode is not mastered by the model, then the association significance is high; if the similarity to a certain template is higher than 0.8, it means that its association mode has been familiar to the model, then the association significance is low.

[0045] S240, the ranking value, the prediction entropy, the clustering representation and the multi-source association significance are weighted and summed to obtain a comprehensive score, and the K unlabeled samples with the highest comprehensive score are labeled by experts, and the labeled unlabeled samples are supplemented to the set of labeled samples.

[0046] Exemplarily, according to the influence degree of each index on the sampling value, the weight coefficients are set. The multi-source association significance weight is 0.35, the model prediction entropy weight is 0.25, the ranker output value weight is 0.2, and the clustering representation weight is 0.2. The sum of each weight coefficient is 1, and the weight setting is adjusted based on the test results of the verification set to ensure the accuracy of the sampling.

[0047] The four-dimensional indexes of each unlabeled sample are normalized, and then weighted and summed according to the weights to obtain the comprehensive score of the sample. For example, the normalized association significance of a sample is 0.9, the prediction entropy is 0.8, the ranking value is 0.7, and the clustering representation is 0.6, then the comprehensive score is 0.9x0.35+0.8x0.25+0.7x0.2+0.6x0.2=0.795.

[0048] The unlabeled samples are ranked from high to low according to the comprehensive scores, and the top K samples are screened out, and the value of K is set according to the labeling ability of experts, and is usually 5%-10% of the total number of unlabeled samples. These high-value samples are handed over to domain experts for labeling, and the labeling content includes fault type (normal, wear, crack, deformation, etc.), fault degree (mild, moderate, severe), and associated feature verification (experts confirm whether the associated features of the sample meet the associated rules of the corresponding fault). After labeling, these samples are converted into labeled samples and supplemented to the set of labeled samples, and the size of the labeled samples is updated, that is, from the original size M of the labeled samples to M+K.

[0049] In S250, pseudo labels are predicted for the unlabeled samples that are not labeled by experts through the fault diagnosis learner, and pseudo labeled samples are obtained; if the pseudo labeled samples are consistent with the state association rule template corresponding to the pseudo labels, the pseudo labeled samples are included in the training set, and the training set also contains labeled samples and labeled unlabeled samples.

[0050] For example, the pseudo labels include fault types and prediction confidence. After the pseudo labels are predicted, the pseudo labeled samples are screened according to the prediction confidence, and the pseudo labeled samples with a prediction confidence greater than or equal to 0.9 are retained, and the pseudo labeled samples with a prediction confidence less than 0.9 are removed.

[0051] Specifically, the remaining unlabeled samples after step S240 are input into the model in the current training state, that is, the model after pre-training and supplement of expert labeled samples. The model extracts an enhanced feature vector through a feature extractor, and then outputs a fault prediction result of each sample through a fault diagnosis learner.

[0052] The prediction result output by the fault diagnosis learner is used as the pseudo label of the sample, and the pseudo label includes two parts of information: one is the fault type, and the other is the prediction confidence. For example, the prediction result of a sample is “wear fault, confidence 0.92”, and the pseudo label of the sample is “wear fault” with a confidence of 0.92.

[0053] All pseudo labeled samples are screened, and samples with a confidence greater than or equal to 0.9 are retained as high-confidence pseudo labeled samples, and samples with a confidence less than 0.9 are removed as low-confidence pseudo labeled samples. The pseudo labels of high-confidence samples are highly reliable and have the value of supplementing the training set; the pseudo labels of low-confidence samples have a large error risk and are not included in the fault diagnosis training process.

[0054] Further, the method for determining whether the pseudo labeled sample is consistent with the state association rule template corresponding to the pseudo label comprises: If all 8 core data pair indexes in the associated feature vector of the pseudo label sample fall within the associated index range of the state corresponding to the pseudo label, and the change trend of the 8 core data pair indexes is consistent with the associated rule template, it is determined that the pseudo label sample is consistent with the state associated rule template. If 3 or more core data pair indexes in the associated feature vector of the pseudo label sample exceed the associated index range, or the change trend of 3 or more core data pair indexes is opposite to the associated rule template, it is determined that the pseudo label sample is inconsistent with the state associated rule template. If 1-2 core data pair indexes in the associated feature vector of the pseudo label sample exceed the associated index range, but the similarity of the change trend of all 8 core data pair indexes in the associated feature vector to the associated rule template is higher than 0.8, it is determined that the pseudo label sample is consistent with the state associated rule template.

[0055] Specifically, based on the set of labeled samples supplemented in step S240, the associated rule template of multi-dimensional data under each state is refined, and the template includes two parts: one is the associated index range (such as the Pearson coefficient of vibration-flying speed under normal state is 0.1-0.3, the mild wear fault is 0.3-0.5, the moderate wear fault is 0.5-0.7, and the severe wear fault is 0.7-0.9); the second is the associated change trend (such as the mutual information between vibration signal and temperature gradually increases over time when there is a crack fault).

[0056] For high-confidence pseudo label samples, verify whether the multi-dimensional data associated features are consistent with the state associated rule template corresponding to the pseudo label, and the judgment criteria are as follows: If 8 core data pair indexes of the sample associated features fall within the associated index range of the state corresponding to the pseudo label, and the associated change trend is consistent with the template, it is determined that the association is consistent; If 3 or more core data pair indexes exceed the corresponding range, or the associated change trend is opposite to the template, it is determined that the association is inconsistent; If 1-2 core data pair indexes exceed the range, but the overall associated mode has a similarity higher than 0.8 with the template, it is determined that the association is basically consistent, which can be equivalent to consistent association.

[0057] The high-confidence pseudo-labels + associated consistent samples are included in the training set as pseudo-label samples. To distinguish the reliability difference between the expert-labeled samples and the pseudo-label samples, the pseudo-label samples are given a lower training weight, and the weight coefficient is 0.3, and the weight coefficient of the expert-labeled samples is 1.0, so as to avoid the possible error of the pseudo-label from having too great an impact on the model training. The high-confidence pseudo-labels + associated inconsistent samples are excluded from the fault diagnosis training link, because the associated features and the pseudo-label state do not match, and there may be a label error. The low-confidence pseudo-label samples do not participate in the fault diagnosis loss calculation, and are only input into the VAE network for reconstruction loss calculation, so as to optimize the feature extraction capability of the VAE network by using the data distribution information of the low-confidence pseudo-label samples, and at the same time, avoid introducing an error fault label.

[0058] In S260, the integrated model is trained by using the training set, to obtain a trained feature extractor and a fault diagnosis learner.

[0059] Exemplarily, the method for training the integrated model includes: Multi-source fusion hidden code feature extraction: the single-source features of each sample in the training set are extracted by the feature extractor, the core data on the index in the associated feature vector is weighted by the attention mechanism, the strengthened associated features are obtained, and the single-source features and the strengthened associated features are spliced, and the multi-source fusion hidden code features of the spliced features are extracted by the feature extractor; Loss calculation: the matching degree between the prediction result of the fault diagnosis learner and the label labeled by the expert or the pseudo-label is used to determine the fault diagnosis loss; the ranking loss is determined based on the relationship between the ranking value of the unlabeled sample labeled by the expert and the actual information value; the multi-source fusion hidden code features are input into the VAE decoder to obtain the reconstructed data of the multi-dimensional data, and the VAE reconstruction loss is determined according to the difference between the reconstructed data and the multi-dimensional data; the hidden code feature vector of the data in the training set is input into the discriminator, and the VAE adversarial loss is determined according to the difference between the discrimination result of the discriminator on the hidden code feature vector and the true source; the discriminator loss is determined based on the difference between the discrimination result of the discriminator on the labeled sample and the unlabeled sample and the true category; Parameter updating: the gradient descent method is used to update the parameters of each module in the integrated model according to the fault diagnosis loss, the ranking loss, the VAE reconstruction loss, the VAE adversarial loss and the discriminator loss.

[0060] Specifically, the fused training set is input into the feature extractor, and the feature extractor extracts the multi-source fusion hidden code features according to the following process: Single-source feature extraction: Extract features from the vibration signal, temperature data, flight parameters, and sensor status data of each sample. For the vibration signal, extract time-domain features (including mean, peak, and root mean square) and frequency-domain features (including characteristic frequency and harmonic amplitude). For the temperature data, extract trend features (including rate of change and mean). For the flight parameters, extract stability features (including fluctuation amplitude). For the sensor status data, extract reliability features (including fault-free duration ratio).

[0061] Through the attention mechanism, assign weights to the associated features of the 8 core data pairs. Assign higher weights to the associated features that are strongly related to the fault state, and strengthen the key associated information.

[0062] Concatenate the single-source features and the strengthened associated features. Compress and reduce the dimensionality through the CNN layer and the fully connected layer of the feature extractor. Finally, output a 128-dimensional multi-source fusion hidden code feature. This feature contains the core information of each dimension of data and the associated rules of multi-dimensional data.

[0063] The fault diagnosis loss L_F measures the difference between the predicted fault result and the true label (or pseudo label). For expert-labeled samples, compare the matching degree of the predicted result output by the fault diagnosis learner and the expert-labeled label. For pseudo-label samples, compare the matching degree of the predicted result and the pseudo label (multiplied by a weight of 0.3). The lower the matching degree, the larger the loss value, and vice versa. The core role of this loss is to promote the fault diagnosis learner to learn the mapping relationship between fault features and fault types.

[0064] The ranking loss L_R measures the accuracy of the ranker's ranking of unlabeled samples. Based on the newly added samples (K) labeled by experts, compare the ranking values of these samples in step S230 with the actual information value. If high-value samples are ranked high, the loss value is small, and vice versa. The core role of this loss is to optimize the ranking ability of the ranker and improve the accuracy of subsequent sampling.

[0065] The VAE reconstruction loss L_VAE_recon measures the reconstruction accuracy of the VAE network for the original sample. The calculation process is as follows: input the multi-source fusion hidden code feature into the VAE decoder, the decoder reconstructs the original sample data, and compares the difference between the reconstructed data and the original data. The smaller the difference, the smaller the loss value. The core role of this loss is to ensure that the multi-source fusion hidden code feature can fully represent the distribution information of the original data and improve the quality of feature extraction.

[0066] The VAE confrontation loss L VAE_con measures the distribution alignment degree of the latent code feature vector between the labeled samples and the unlabeled samples. The calculation process is as follows: the latent code feature vectors of the expert labeled samples, the pseudo-labeled samples and the low-confidence unlabeled samples are input into the discriminator, the discriminator tries to distinguish the features from which kind of samples, and the VAE network makes the latent code feature distribution as consistent as possible through confrontation training, so that the discriminator cannot accurately distinguish, so as to improve the generality of the features. The smaller the confrontation loss value is, the better the feature distribution alignment effect is.

[0067] The discriminator loss L D measures the distinguishing ability of the discriminator to the latent code feature vector. The calculation process is as follows: the difference between the distinguishing results of the discriminator to the features of the labeled samples and the unlabeled samples and the true categories is compared, the smaller the difference is, the smaller the loss value is. The core role of this loss is to optimize the distinguishing ability of the discriminator, and then promote the VAE network to learn better latent code features.

[0068] The gradient descent algorithm is adopted to update the parameters of each module of the model according to the weighted sum of each loss. The parameter update logic is as follows: Fault diagnosis learner and ranker parameter update: the total loss is the fault diagnosis loss + the ranking loss x the scale adjustment hyperparameter η1, η1 is 0.2, which is used to balance the influence degree of the two kinds of losses, and then the gradient of the total loss to the fault diagnosis learner parameter θ F and the ranker parameter θ R is calculated, and then the parameters are updated according to the following process: θ F = θ F - α1 x gradient, α1 is the first learning coefficient, which is 0.001, and is used to control the parameter update step; θ R = θ R - α2 x gradient, α2 is the second learning coefficient, which is 0.0008, and the step is slightly smaller because the complexity of the ranker is lower than that of the fault diagnosis learner.

[0069] VAE network parameter update: the total loss is the VAE reconstruction loss + the VAE confrontation loss x the scale adjustment hyperparameter η2, η2 is 0.3, which balances the reconstruction accuracy and the distribution alignment, and then the gradient of the total loss to the VAE encoder parameter θ P and the decoder parameter θ Q is calculated, and then the updating process is as follows: θ P = θ P - α3 x gradient, α3 is the third learning coefficient, which is 0.0005; θ Q = θ Q - α4 x gradient, α4 is the fourth learning coefficient, which is 0.0006.

[0070] Discriminator parameter update: the total loss is the discriminator loss L D, and the gradient of the loss to the discriminator parameter θ D is calculated, and the updating logic is as follows: θ D = θ D - α5 x gradient, α5 is the fifth learning coefficient, which is 0.0012.

[0071] Parameter update constraint: after all parameter updates, ensure that the parameter values are within a reasonable range to avoid parameter explosion and cause the model to fail. At the same time, calculate the fault diagnosis accuracy of the validation set every 10 rounds of parameter update. If the accuracy rate has decreased for 3 consecutive rounds, reduce the learning coefficient to stabilize the model training.

[0072] Repeat steps S220-S260 to form an iterative optimization closed loop, and the iteration termination condition is as follows: Model convergence: the fault diagnosis accuracy of the validation set has improved by less than 0.5% for 5 consecutive rounds, and the accuracy is more than 90%, indicating that the model has fully learned the fault features and correlation rules, and the performance tends to be stable; Upper limit of iteration rounds: if the model does not converge, but the iteration rounds reach 200 rounds, the iteration is also terminated to avoid overfitting caused by overtraining; Sample depletion: if all unlabeled samples are selected as expert-labeled samples or pseudo-labeled samples, and the model does not converge, then continue iteration after supplementing new unlabeled samples to ensure that the model has sufficient data support.

[0073] During the iteration process, the correlation rule template needs to be updated, the sample selection index weight needs to be adjusted, and the loss function scale adjustment hyperparameters need to be optimized after each iteration to ensure continuous optimization of the model performance during the iteration process.

[0074] When the model iteration converges, enter the actual fault diagnosis phase, and the specific process is as follows: Diagnosed sample processing: obtain the real-time multi-source data of the unmanned aerial vehicle rotor bearing, and perform denoising, time axis alignment, and normalization processing according to the preprocessing process of step S100 to construct the diagnosed sample; Multi-source fusion hidden code feature extraction: input the diagnosed sample into the trained feature extractor to extract a multi-source fusion hidden code feature vector z_target, which contains single-source features and multi-source correlation features of the diagnosed sample; Model inference: input the multi-source fusion hidden code feature vector z_target into the fault diagnosis learner, and the fault diagnosis learner outputs the fault diagnosis result according to the optimal parameter θ_F*, i.e., the fault diagnosis learner parameter obtained after iteration convergence; Diagnosis result output: the diagnosis result contains two parts of core information: Fault type: clearly indicates the current state of the bearing, such as normal, wear, crack, deformation, etc. Fault degree: quantifies the severity of the fault, i.e., mild, moderate, and severe.

[0075] The embodiment of the application also provides an unmanned aerial vehicle rotor bearing fault diagnosis system based on multi-source data, which comprises: The data to be diagnosed is obtained by a data acquisition module, which is configured to acquire multi-source data of the rotor bearing of the unmanned aerial vehicle in operation, and to preprocess the multi-source data to obtain corresponding samples to be diagnosed. The fault diagnosis module is configured to extract a multi-source fusion hidden code feature vector of the sample to be diagnosed by using a feature extractor, and to input the multi-source fusion hidden code feature vector into a fault diagnosis learner to output a fault diagnosis result. The training module for training the feature extractor and the fault diagnosis learner includes: The training data acquisition submodule is configured to collect multi-dimensional data of the rotor bearing of the unmanned aerial vehicle in operation, and to divide the multi-dimensional data into labeled samples and unlabeled samples according to whether the multi-dimensional data is known to have a fault condition. The model construction submodule is configured to construct an integrated model, and the integrated model includes the feature extractor, the ranker, and the fault diagnosis learner. The correlation enhancement submodule includes calculating a correlation feature vector of the unlabeled sample, supplementing the correlation feature vector to the original features of the unlabeled sample to form an enhanced feature vector. The sample selection index construction submodule is configured to input the enhanced feature vector into the ranker to obtain a corresponding ranking value, and to input the unlabeled sample into the feature extractor and the fault diagnosis learner in turn to obtain a corresponding prediction entropy. The enhanced feature vector is clustered to obtain a clustering representative, and the similarity between the unlabeled sample and the correlation mode template is calculated to determine the multi-source correlation significance. The sample labeling submodule is configured to weight and sum the ranking value, the prediction entropy, the clustering representative, and the multi-source correlation significance to obtain a comprehensive score, and to label the K unlabeled samples with the highest comprehensive score by experts, and to supplement the labeled unlabeled samples to the set of labeled samples. The training set construction submodule is configured to predict pseudo-labels for the unlabeled samples that have not been labeled by experts by using the fault diagnosis learner to obtain pseudo-labeled samples, and to include the pseudo-labeled samples in the training set if the pseudo-labeled samples are consistent with the state correlation rule template corresponding to the pseudo-labels. The training set also includes the labeled samples and the labeled unlabeled samples. The training submodule is configured to train the integrated model using the training set to obtain the trained feature extractor and the fault diagnosis learner.

[0076] Experimental description (I) Experimental purpose The performance advantage of the method of fusing multi-source data correlation, pseudo-labels, and active learning proposed in the present application (hereinafter referred to as the present method) compared with the method of the prior patent application CN120234570A (hereinafter referred to as the original method) is verified. The differences between the two methods in terms of fault diagnosis accuracy, labeling cost, training efficiency, and robustness are focused on, and the actual effect of the improvement strategy is determined.

[0077] (II) Data source The experimental data is measured from industrial-grade unmanned aerial vehicle rotor bearing, covering the whole life cycle of the bearing operation state, as follows: Data type: vibration signal (X / Y / Z three directions, sampling frequency 20 kHz), bearing surface temperature (sampling frequency 1 Hz), flight parameters (flight speed 0-80 m / s, load 0-5 kg, environment temperature -10~45℃); Fault type: including normal state, ball wear, inner ring crack, outer ring deformation, among which the proportion of mild fault samples is 30%, moderate 30%, severe 20%, and normal 20%; Sample size: a total of 10000 samples (each sample is 1 second of time window data), divided into training set (7000), validation set (2000), and test set (1000) according to the ratio of 7:2:1. There are M=1000 labeled samples and N=6000 unlabeled samples in the training set, and both the validation set and the test set are labeled samples.

[0078] (Three) Evaluation index Diagnosis performance index: fault diagnosis accuracy (the proportion of correctly predicted samples), fault type recall rate (the proportion of correctly identified faults), F1 value (the harmonic mean of precision and recall, which comprehensively reflects the classification effect); Labeling cost index: the proportion of expert labeled samples when achieving the same diagnosis accuracy (labeled sample number / total training sample number); Training efficiency index: the number of iterations required for model convergence (the convergence standard is that the validation set accuracy improves by less than or equal to 0.5% for 5 consecutive rounds); Robustness index: the decline in diagnosis accuracy after adding 5% Gaussian noise to the test set.

[0079] II. Experimental parameter setting To ensure fairness, the basic experimental conditions (hardware environment: Intel i9-13900K CPU, NVIDIA RTX4090 GPU; software environment: PyTorch 2.0) and core model structure of the two methods are consistent, only the difference is that the improved method adds multi-source correlation utilization, pseudo-label strategy and optimized active learning sampling logic. The specific parameters are as follows: (One) Data processing parameters Time window length 1 second, vibration signal uses wavelet filter denoising (decomposition scale 5 layers), data normalization range [0, 1], and multi-source data time axis alignment benchmark is flight control system timestamp.

[0080] (Two) Model basic parameters Feature extractor: CNN + attention mechanism (3 convolutional layers, 2 pooling layers, 1 attention layer), output hidden code feature vector dimension 128; Ranker: 3-layer fully connected network (input dimension = single-source feature dimension, output dimension 1); Fault diagnosis learner: 4-layer fully connected network + 1-layer time series fusion layer (hidden layer neuron numbers are 256, 128, 64, and 32, respectively); VAE network: both the encoder and the decoder are 3-layer fully connected networks (hidden layer neuron numbers are 128, 64, and 32); Discriminator: 2-layer fully connected network (hidden layer neuron numbers are 64 and 32).

[0081] (Three) Training related parameters Learning coefficients: a1 = 0.001 (fault diagnosis learner), a2 = 0.0008 (ranker), a3 = 0.0005 (VAE encoder), a4 = 0.0006 (VAE decoder), a5 = 0.0012 (discriminator); Scale adjustment hyperparameters: η1 = 0.2 (ranking loss), η2 = 0.3 (VAE adversarial loss); Batch size: 32, maximum iteration rounds 200.

[0082] (Four) Method-specific parameters Active learning: four-dimensional index weight (correlation significance 0.35, prediction entropy 0.25, ranking value 0.2, clustering representativeness 0.2), K = 300 samples (5% of unlabeled samples) are selected each round; Pseudo-label: confidence threshold 0.9, pseudo-label sample training weight 0.3; Correlation calculation: 8 groups of core data pairs (vibration-X and temperature, vibration-Y and load, vibration-Z and flight speed, etc.), mutual information calculation window 0.5 seconds, Pearson coefficient sliding window 10 seconds.

[0083] Three, experimental process (I) Data preprocessing Both methods use exactly the same preprocessing procedure: denoising, time axis alignment, and normalization of the original multi-source data, dividing the training set, validation set, and test set in proportion to ensure uniform data input format.

[0084] (II) Model training Original method training process: Initialize model parameters, pre-train with initial labeled samples (1000) for 50 rounds; In each round of iteration, the unlabeled samples are ranked by the ranker, and the top K=300 samples are selected for expert annotation and supplemented to the labeled sample set. The fault diagnosis loss, ranking loss, VAE reconstruction / adversarial loss, and discriminator loss are calculated, and the parameters of each module are updated. Repeat the iteration until convergence or reach the maximum number of rounds.

[0085] The training process of the method is as follows: Initialize the model parameters and pre-train for 50 rounds with the initial labeled samples (consistent with the original method); In each round of iteration, first calculate the multi-source correlation features of the unlabeled samples, construct a four-dimensional sampling index, and select the top K=300 high-value samples for annotation. Generate pseudo-labels for the remaining unlabeled samples, and select samples with "confidence ≥ 0.9 + correlation consistency" to be included in the training set (weight 0.3); Extract multi-source fusion hidden code features (including correlation features), calculate various losses, and update parameters; Repeat the iteration, update the correlation rule template in each round, and dynamically adjust the sampling weight until convergence.

[0086] (Three) Performance test Basic performance test: input the test set into the trained two models, output the fault diagnosis results, and calculate the accuracy, recall rate, and F1 value; Labeling cost test: record the total number of expert-labeled samples when the two methods reach an accuracy of 90% on the validation set, and calculate the labeling proportion; Robustness test: add 5% Gaussian noise to the test set vibration signals, retest the diagnosis accuracy, and calculate the decrease; Stability verification: repeat the experiment 5 times for each method, take the average of each index, and avoid the influence of random errors.

[0087] Four, comparison results and analysis (I) Statistical results of the experiment

[0088] (II) Results analysis Improved diagnostic performance: The accuracy of the method is improved by 9.5% compared to the original method, and the recall rate of mild faults is significantly improved (15.8%). The core reason is that the method utilizes the correlation of multi-source data - when the fault is mild, the single-source feature does not change significantly, but the correlation pattern between vibration and temperature, load has already been abnormal. By strengthening the correlation features, the model can capture the fault signal earlier; while the original method only relies on single-source features, it is difficult to identify early weak faults.

[0089] The annotation cost is reduced: the proportion of annotated samples is reduced from 28.3% to 17.5%, and the annotation cost is reduced by about 38%. On the one hand, the four-dimensional sampling index selects samples with higher information value, avoiding redundant annotation caused by the single ranker of the original method. On the other hand, the high-confidence pseudo-label sample supplements the training data, reducing the dependence on expert annotation, while ensuring accuracy and reducing labor costs.

[0090] The training efficiency is optimized: the convergence round is reduced by 34 rounds, and the efficiency is improved by 23.9%. This is because the pseudo-label sample expands the training data size, and the multi-source correlation feature provides more rich constraint information for the model, so that the model can learn the mapping relationship between faults and features faster, and can converge without too many iterations.

[0091] Robustness is enhanced: the accuracy of the original method decreases by 11.8% after adding noise, while the accuracy of the improved method decreases by only 4.8%. The multi-source data correlation has stability - even if the single-source data is disturbed by noise, the correlation rules under normal or fault conditions remain the same. This constraint significantly improves the anti-interference ability of the model, while the original method only relies on single-source features and is easily affected by noise.

[0092] V. CONCLUSION The experimental results show that the improved method of fusing multi-source data correlation, pseudo-label and active learning is superior to the original method in various core indicators: the fault diagnosis accuracy is improved by 9.5%, the light fault recognition ability is greatly enhanced, the annotation cost is reduced by 38%, the training efficiency is improved by 23.9%, and the anti-noise robustness is stronger. The improved strategy fully excavates the potential correlation information of multi-source data, optimizes sample selection and data utilization efficiency, effectively solves the problems of "single-source feature limitation, high annotation cost and weak robustness" of the original method, and is more suitable for the rotor bearing fault diagnosis scene of unmanned aerial vehicles under complex working conditions.

[0093] Although the preferred embodiments of the application have been described, those skilled in the art can make further changes and modifications to the embodiments once they know the basic inventive concept. Therefore, the appended claims are intended to be interpreted as including all changes and modifications falling within the scope of the application.

[0094] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application also intends to include these modifications and variations.

Claims

1. A method for diagnosing faults of a rotor bearing of a UAV based on multi-source data, characterized in that, The method comprises the following steps: acquiring multi-source data of a rotor bearing of a UAV in operation, preprocessing the multi-source data to obtain corresponding samples to be diagnosed; using a feature extractor to extract a multi-source fusion hidden code feature vector of the samples to be diagnosed, inputting the multi-source fusion hidden code feature vector into a fault diagnosis learner, and outputting a fault diagnosis result; wherein the training method of the feature extractor and the fault diagnosis learner is as follows: collecting multi-dimensional data of the rotor bearing of the UAV in operation, and dividing the multi-dimensional data into labeled samples and unlabeled samples according to whether the fault condition of the multi-dimensional data is known; constructing an integrated model, wherein the integrated model comprises the feature extractor, a ranker, and the fault diagnosis learner; calculating a correlation feature vector of the unlabeled samples, supplementing the correlation feature vector to the original features of the unlabeled samples to form an enhanced feature vector; inputting the enhanced feature vector into the ranker to obtain a corresponding ranking value; inputting the unlabeled samples into the feature extractor and the fault diagnosis learner in turn to obtain a corresponding prediction entropy; clustering the enhanced feature vector to obtain a clustering representative; calculating the similarity between the unlabeled samples and a correlation mode template to determine multi-source correlation significance; weighting and summing the ranking value, the prediction entropy, the clustering representative, and the multi-source correlation significance to obtain a comprehensive score, labeling the K unlabeled samples with the highest comprehensive score, and supplementing the labeled unlabeled samples to the set of labeled samples; predicting pseudo labels for the unlabeled samples that have not been labeled by experts through the fault diagnosis learner to obtain pseudo labeled samples; if the pseudo labeled samples are consistent with the state correlation rule template corresponding to the pseudo labels, the pseudo labeled samples are included in a training set, and the training set also contains the labeled samples and the labeled unlabeled samples; training the integrated model using the training set to obtain the trained feature extractor and the fault diagnosis learner. 2.The method of claim 1, wherein, The multi-source data and the multi-dimensional data both contain vibration signals and temperature data of the rotor bearing, and flight parameters and sensor state data of the UAV. 3.The method of claim 1, wherein, The integrated model further comprises a VAE network and a discriminator, and after the integrated model is constructed, the feature extractor, the ranker, the fault diagnosis learner, the VAE network, and the discriminator in the integrated model are parameterized and pre-trained. 4.The method of claim 1, wherein, Based on mutual information and Pearson correlation coefficient, the correlation between different dimensional data in the unlabeled samples is calculated to form the correlation feature vector.

5. The method for diagnosing faults of a rotor bearing of a UAV based on multi-source data according to claim 1, characterized in that, The pseudo labels contain fault types and prediction confidence, and after the pseudo labels are predicted, the pseudo labeled samples are screened according to the prediction confidence, and the pseudo labeled samples with a prediction confidence greater than or equal to 0.9 are retained, and the pseudo labeled samples with a prediction confidence less than 0.9 are removed. 6.The method of claim 1, wherein, The method for determining whether the pseudo labeled samples are consistent with the state correlation rule template corresponding to the pseudo labels comprises: If all 8 core data pair indexes in the association feature vector of the pseudo label sample fall within the association index range of the pseudo label corresponding state, and the change trend of the 8 core data pair indexes is consistent with the association rule template, it is determined that the pseudo label sample is consistent with the state association rule template; If 3 or more core data pair indexes in the association feature vector of the pseudo label sample exceed the association index range, or the change trend of 3 or more core data pair indexes is opposite to the association rule template, it is determined that the pseudo label sample is inconsistent with the state association rule template; If 1-2 core data pair indexes in the association feature vector of the pseudo label sample exceed the association index range, but the similarity of the change trend of all 8 core data pair indexes in the association feature vector to the association rule template is higher than 0.8, it is determined that the pseudo label sample is consistent with the state association rule template.

7. The method for diagnosing faults of a rotor bearing of a UAV based on multi-source data according to claim 1, characterized in that, The method for training the integrated model comprises: Multi-source fusion hidden code feature extraction: the feature extractor extracts single-source features for each sample in the training set, assigns weights to the features of the core data pair indexes in the association feature vector through an attention mechanism to obtain reinforced association features, concatenates the single-source features and the reinforced association features, and extracts multi-source fusion hidden code features of the concatenated features by the feature extractor; Loss calculation: determine a fault diagnosis loss based on the matching degree of the prediction result of the fault diagnosis learner and the label marked by an expert or the pseudo label; determine a ranking loss based on the relationship between the ranking value of the unlabeled sample marked by an expert and the actual information value; input the multi-source fusion hidden code features into a VAE decoder to obtain reconstructed data of the multi-dimensional data, and determine a VAE reconstruction loss based on the difference between the reconstructed data and the multi-dimensional data; input the hidden code feature vectors of the data in the training set into a discriminator, and determine a VAE adversarial loss based on the difference between the discrimination result of the discriminator on the source of the hidden code feature vector and the true source; determine a discriminator loss based on the difference between the discrimination result of the discriminator on the labeled sample and the unlabeled sample and the true category; Parameter update: use the gradient descent method to update the parameters of each module in the integrated model according to the fault diagnosis loss, the ranking loss, the VAE reconstruction loss, the VAE adversarial loss, and the discriminator loss. 8.The method of claim 1, wherein, After obtaining the fault diagnosis result, compare the fault diagnosis result with the unmanned aerial vehicle operation log and the artificial inspection result, if the fault diagnosis result does not match the actual state, mark the current sample as a diagnosis error sample and supplement it to the set of unlabeled samples, and start the next round of iterative optimization.

9. The system for diagnosing faults in the rotor bearing of the unmanned aerial vehicle based on multi-source data according to any one of claims 1-8, characterized in that, The system comprises: A to-be-diagnosed data acquisition module for acquiring multi-source data of unmanned aerial vehicle rotor bearing operation, pre-processing the multi-source data, and obtaining corresponding to-be-diagnosed samples; The fault diagnosis module is configured to extract a multi-source fusion hidden code feature vector of the sample to be diagnosed by using the feature extractor, input the multi-source fusion hidden code feature vector into a fault diagnosis learner, and output a fault diagnosis result. The training module for training the feature extractor and the fault diagnosis learner includes: A training data acquisition submodule is configured to collect multi-dimensional data of a rotor bearing of a UAV in operation, and divide the multi-dimensional data into labeled samples and unlabeled samples according to whether the multi-dimensional data is known to have a fault condition. A model construction submodule is configured to construct an integrated model, the integrated model including the feature extractor, a ranker, and the fault diagnosis learner. An association enhancement submodule includes calculating an association feature vector of the unlabeled samples, supplementing the association feature vector to original features of the unlabeled samples to form an enhanced feature vector. A sample selection index construction submodule is configured to input the enhanced feature vector into the ranker to obtain a corresponding ranking value, input the unlabeled samples into the feature extractor and the fault diagnosis learner in sequence to obtain a corresponding prediction entropy, cluster the enhanced feature vector to obtain a clustering representation, calculate a similarity of the unlabeled samples to an association mode template to determine multi-source association significance. A sample labeling submodule is configured to weight and sum the ranking value, the prediction entropy, the clustering representation, and the multi-source association significance to obtain a comprehensive score, label K unlabeled samples with the highest comprehensive score by experts, supplement the labeled unlabeled samples to a set of labeled samples, and obtain a training set. A training set construction submodule is configured to predict pseudo labels of unlabeled samples that are not labeled by experts by using the fault diagnosis learner to obtain pseudo labeled samples, and if the pseudo labeled samples are consistent with a state association rule template corresponding to the pseudo labels, include the pseudo labeled samples in the training set, the training set also including the labeled samples and the labeled unlabeled samples. A training submodule is configured to train the integrated model by using the training set to obtain the feature extractor and the fault diagnosis learner after training.

Images