Cross-working-condition rotor unknown fault diagnosis method

Abstract

A cross-working-condition rotor unknown fault diagnosis method comprises the following steps that a, collected vibration signals of a rotating machine under different working conditions are cut and subjected to short-time Fourier transform, two-dimensional time-frequency domain signals are obtained, and a labeled source domain data set and an unlabeled target domain data set are constructed through the two-dimensional time-frequency domain signals; b, constructing a multi-classifier residual universal domain adaptive network model to obtain a fault diagnosis model; and c, inputting the label-free target domain data into the fault diagnosis model, and outputting a diagnosis result. The cross-working-condition rotor unknown fault is diagnosed by adopting the method based on the multi-classifier residual universal domain adaptation network model, common characteristics of the same fault category under different working conditions can be extracted, the cross-working-condition rotor unknown fault can be accurately recognized, accurate diagnosis of the rotor fault state is achieved, and the fault diagnosis efficiency is improved. And an effective tool is provided for overcoming diagnosis difficulty caused by source domain data shortage and unknown target domain fault mode.

Description

Technical Field

[0001] The present invention relates to a cross-condition rotor unknown fault diagnosis method based on a multi-classifier residual universal domain adaptation network model (MRUDA), belonging to the technical field of data processing. Background Art

[0002] The rotor is one of the key components of rotating machinery. Therefore, it is crucial to ensure that the rotor is in good operating condition to guarantee the safe operation of the equipment. In recent years, rotor intelligent fault diagnosis methods based on deep learning have emerged in an endless stream, providing a powerful tool for monitoring the safe operation of mechanical equipment.

[0003] The invention patent with the publication number CN118171065A in Chinese patents discloses a motor rotor fault diagnosis method based on CNN-BiLSTM-residual module-attention mechanism. Its technical solution includes the following steps: S1. Optimize the VMD parameters using the improved Chernobyl disaster algorithm ICDO; S2. After the VMD decomposes the initial data of the motor rotor fault, select the IMF components with fault feature signal components according to the correlation coefficient criterion and kurtosis, and then perform signal reconstruction; S3. Calculate the time-domain features and frequency-domain features of the reconstructed signal, and use the comprehensive analysis method to screen the data features; S4. Input the screened data features into the CNN-BiLSTM-residual module-attention mechanism network for fault diagnosis. The above-mentioned motor rotor fault diagnosis method based on CNN-BiLSTM-residual module-attention mechanism can accurately realize the motor rotor fault diagnosis. At the same time, through the diagnosis of motor rotor faults and different bearing faults when adding noise interference, it shows that the model has good anti-noise robustness and generalization. However, with the continuous optimization of rotor design, the progress of monitoring technology and the continuous improvement of fault prevention measures, the occurrence frequency of rotor faults has decreased, and faults can usually be detected and repaired in time at an early stage, resulting in insufficient rotor fault data being obtained. Existing cross-domain fault diagnosis methods have high requirements for data annotation, and require that the training and test data must follow similar distributions. The relative scarcity of fault samples and the difficulty of data collection lead to challenges in the learning ability and generalization ability of the model, especially when the fault type is unknown. At the same time, problems such as data privacy, data sharing restrictions or unavailable source data in reality may prevent us from accessing the data in the source domain. Therefore, how to reduce the dependence on data and improve the cross-condition rotor unknown fault diagnosis ability is very important. Summary of the Invention

[0004] The object of the present invention is to provide a method for diagnosing unknown faults of a rotor across different operating conditions, aiming at the drawbacks of the existing technology. This method can improve the diagnostic ability of unknown faults of the rotor across different operating conditions and ensure the safe operation of mechanical equipment.

[0005] The problems of the present invention are solved by the following technical solutions:

[0006] A method for diagnosing unknown faults of a rotor across different operating conditions, the method comprising the following steps:

[0007] a. Data sampling and preprocessing: Cut the vibration signals of the rotating machinery collected under different operating conditions and perform short-time Fourier transform to obtain two-dimensional time-frequency domain signals. Use the two-dimensional time-frequency domain signals to construct a labeled source domain dataset and an unlabeled target domain dataset. Among them, the target domain dataset includes the fault types that exist and do not exist in the source domain dataset;

[0008] b. Construct a multi-classifier residual common domain adaptation network model, including two stages: source model generation and model domain adaptation, to obtain a fault diagnosis model;

[0009] c. Input the unlabeled target domain data into the fault diagnosis model, and output the diagnosis result. Samples with abnormal values greater than the abnormal threshold are considered unknown faults, otherwise they are considered known faults.

[0010] For the above method for diagnosing unknown faults of a rotor across different operating conditions, the specific process of the source model generation is as follows:

[0011] First, input the labeled source domain dataset into the constructed multi-classifier residual network. The multi-classifier residual network includes a feature extractor based on a convolutional attention module, an open-set classifier, and a closed-set classifier. The closed-set classifier uses the softmax function for output. The open-set classifier consists of L c sub-classifiers. One classifier is set for each class of samples. For the i-th classifier, the i-th class of samples is recognized as positive, and any other class of samples is regarded as negative;

[0012] Then calculate the total loss L s of the source domain training, and backpropagate the model parameters until the maximum iteration parameter is reached. The cross-entropy loss function is used to train the closed-set classifier for classifying known classes. The cross-entropy loss function is expressed as:

[0013]

[0014] For the open-set classifier, select l ova (x s , y s ) as the open-set classification loss, which is expressed as:

[0015]

[0016] where C is the number of rotor fault conditions, N is the number of samples, and y i is the true label of the i-th sample, and x s is the input sample, and y s is the true label of the input sample. p represents the probability that the sample is judged as y s when the given input is x s . is the predicted output of the model. The predicted label is obtained by the softmax function in the fault classification, and its writing method is as follows:

[0017]

[0018] w i is the score of the model for the i-th category. The total loss of source domain training is:

[0019] L s = loss clc + l ova (x s , y s )

[0020] Finally, save the model parameters to obtain the trained source domain model.

[0021] For the above cross-condition rotor unknown fault diagnosis method, the specific process of model domain adaptation is as follows:

[0022] ① Input the unlabeled target domain dataset into the multi-classifier residual network with parameter initialization completed, and perform the source-free general domain adaptation process;

[0023] ② Use the pseudo-label refinement strategy based on cosine distance to refine the pseudo-labels obtained in the model adaptation process, reduce the influence of noise, and obtain more accurate pseudo-labels for model training;

[0024] ③ Calculate the L MRUDA loss function, minimize the loss function and backpropagate to obtain the fault diagnosis model.

[0025] For the above cross-condition rotor unknown fault diagnosis method, the specific process of the pseudo-label refinement strategy based on cosine distance is as follows:

[0026] Ⅰ. Input the unlabeled target domain dataset into the multi-classifier residual network with parameter initialization completed to obtain the initial pseudo-labels;

[0027] Ⅱ. Calculate the cosine distance between the target sample and other samples, and select the 10 samples with the closest distance as adjacent samples;

[0028] Ⅲ. Take the average of the predicted scores of adjacent samples to obtain the average score vector;

[0029] Ⅳ. Use the average score vector and perform an argmax operation to calculate the exact pseudo-labels, which are used for self-supervised target samples.

[0030] For the rotor unknown fault diagnosis method across working conditions, calculate L MRUDA The specific process of the loss function is as follows:

[0031] For the closed-set classifier, use a classification loss function with entropy reweighting for training. Assign different weights to pseudo-labels with different reliabilities to reduce the impact of mislabels. Calculate the predicted probability distribution i of the model for the target sample x The entropy is:

[0032]

[0033] where C is the number of rotor fault conditions, i is the index of the sample, is the predicted probability that the model assigns the sample x i to the c-th category. Normalize the entropy and obtain the sample weights:

[0034]

[0035] is the weight of the sample x i The reweighted classification loss is:

[0036]

[0037] Through the loss reweighting strategy, punish the pseudo-labels with high uncertainty and use reliable pseudo-labels to guide the model training;

[0038] For the open-set classifier, use the average entropy loss function for training. The average entropy of all binary classifiers is

[0039]

[0040] where x t is the input sample in the target domain, is the predicted label of the sample;

[0041] L MRUDA The loss function is

[0042] L MRUDA = loss ent (x t ) + loss wclc .

[0043] For the above cross-condition rotor unknown fault diagnosis method, the abnormal threshold is set to 0.5.

[0044] The present invention uses a multi-classifier residual common domain adaptation network model to diagnose cross-condition rotor unknown faults, which can not only extract the common features of the same fault category under different conditions, but also accurately identify the unknown faults of the rotor under cross-conditions, realizing the accurate diagnosis of the rotor fault state, and providing an effective tool to overcome the diagnostic difficulties caused by the lack of source domain data and the unknown target domain fault mode. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The present invention will be further described in detail below with reference to the drawings.

[0046] Figure 1 is the overall flowchart of the present invention;

[0047] Figure 2 is the analysis result diagram of the rotor fault signal of a certain instrument and meter company in Wuxi using the method of the present invention;

[0048] Figure 3 is the analysis result diagram of the rotor fault signal of the Bently test bench using the method of the present invention;

[0049] Figure 4 is the comparison diagram of the open set fault diagnosis results provided by the embodiment of the present invention.

[0050] Each symbol in the text represents: loss clc represents the cross-entropy loss function, l ova (x s , y s ) is the open set classification loss, C is the number of rotor fault conditions, N is the number of samples, p represents the probability that the sample is judged as y s when the given input is x s ; is the predicted output of the model, y i is the true label of the i-th sample, x s is the input sample, y s is the true label of the input sample, is the predicted label, w i is the score of the model for the i-th category, L s is the total loss of source domain training, is the predicted probability of the model for the target sample, represents entropy of is the predicted probability that the sample belongs to the c-th category, is the weight of the sample x i , x i is the input target sample, losswclc is the reweighted classification loss, loss ent (x t ) is the average entropy of all binary classifiers, x t is the input sample of the target domain, is the predicted label of the sample, L MRUDA is L MRUDA loss function. Specific implementation manner

[0051] The present invention provides a cross - working - condition rotor unknown fault diagnosis method based on a multi - classifier residual universal domain adaptation network model (multi - classifier residual universal domain adaptation network, MRUDA). This method can effectively extract the deep features of rotor faults, realize the accurate identification of rotor unknown faults under cross - working conditions, and ensure the safe operation of mechanical equipment.

[0052] Refer to Figure 1 , this method includes the following steps:

[0053] a. Data sampling and pre - processing. Collect the vibration signals of rotating machinery under different working conditions, classify the original vibration signals according to different working conditions, different fault degrees, and different fault types, cut the collected data set and perform short - time Fourier transform to obtain two - dimensional time - frequency domain signals as the input signals of the model, and construct a labeled source - domain data set and an unlabeled target - domain data set. Among them, the target - domain data set includes the fault types existing and not existing in the source - domain data set.

[0054] b. Construct a multi - classifier residual universal domain adaptation network model, including two stages: source model generation and model domain adaptation, to obtain a fault diagnosis model.

[0055] Input the source - domain data constructed in step a into the designed multi - classifier residual network, train to obtain a source model, and save the model parameters; initialize the parameters of the target - domain multi - classifier residual network with the source - model network parameters, construct an L MRUDA loss function to train the network, and at the same time introduce a pseudo - label refinement strategy based on cosine distance to obtain a fault diagnosis model.

[0056] The specific process of source model generation is as follows:

[0057] First, the labeled source domain dataset is input into the constructed multi-classifier residual network. The multi-classifier residual network includes a feature extractor based on a convolutional attention module. The feature extractor takes three 3×3 convolutional kernels as the input layer, connects three residual network modules, and incorporates a CBAM attention mechanism, an open-set classifier, and a closed-set classifier. The closed-set classifier uses the softmax function for output. The open-set classifier consists of L c sub-classifiers. One classifier is set for each class of samples. For the i-th classifier, the i-th class of samples is recognized as positive, and any other class of samples is regarded as negative.

[0058] Then, calculate the total loss L s of the source domain training, and backpropagate the model parameters until the maximum iteration parameter is reached. The cross-entropy loss function is used to train the closed-set classifier for classifying known classes. The cross-entropy loss function is expressed as:

[0059]

[0060] For the open-set classifier, select l ova (x s ,y s ) as the open-set classification loss, which is expressed as:

[0061]

[0062] where C is the number of rotor fault conditions, N is the number of samples, y i is the true label of the i-th sample, x s is the input sample, y s is the true label of the input sample, p represents the probability that the sample is judged as y s given the input x s , is the predicted output of the model, and the predicted label is obtained by the softmax function in the fault classification, and its writing is as follows:

[0063]

[0064] w i is the score of the model for the i-th class. The total loss of the source domain training is:

[0065] L s = loss clc + l ova (x s ,y s )

[0066] Finally, save the model parameters to obtain the trained source domain model.

[0067] The specific process of model domain adaptation is as follows:

[0068] ① Input the unlabeled target domain dataset into the multi-classifier residual network with initialized parameters to perform the source-free general domain adaptation process.

[0069] ② Use the pseudo-label refinement strategy based on cosine distance to refine the pseudo-labels obtained in the model adaptation process, reduce the influence of noise, and obtain more accurate pseudo-labels for model training.

[0070] The specific process of the pseudo-label refinement strategy based on cosine distance is as follows:

[0071] Ⅰ. Input the unlabeled target domain dataset into the multi-classifier residual network with initialized parameters to obtain initial pseudo-labels.

[0072] Ⅱ. Calculate the cosine distance between the target sample and other samples, and select the 10 samples with the closest distance as adjacent samples.

[0073] Ⅲ. Take the average of the prediction scores of these adjacent samples to generate an average score vector.

[0074] Ⅳ. Use the argmax operation to calculate the accurate pseudo-labels and use them for self-supervised target samples.

[0075] ③ Calculate the L MRUDA loss function, minimize the loss function and backpropagate to obtain the fault diagnosis model.

[0076] L MRUDA The specific process of calculating the loss function is as follows:

[0077] For the closed-set classifier, use the classification loss function with entropy reweighting for training, assign different weights to pseudo-labels with different reliabilities, reduce the influence of mislabels, and calculate the prediction probability distribution i of the model for the target sample x The entropy is:

[0078]

[0079] where C is the number of rotor fault conditions, i is the index of the sample, is the predicted probability that the model assigns the sample x i to the c-th category. Normalize the entropy and obtain the sample weight:

[0080]

[0081] is the weight of the sample x i The reweighted classification loss is:

[0082]

[0083] Through the loss reweighting strategy, the pseudo-labels with high uncertainty are penalized, and the model training is guided by reliable pseudo-labels.

[0084] The open-set classifier is trained using the average entropy loss function, and the average entropy of all binary classifiers is

[0085]

[0086] where x t is the input sample of the target domain, is the predicted label of the sample;

[0087] L MRUDA The loss function is calculated as

[0088] L MRUDA = loss ent (x t ) + loss wclc

[0089] c. Input the unlabeled target domain data into the fault diagnosis model to diagnose the unknown faults of the rotor and output the diagnosis results. The cross-domain fault diagnosis performance of the model is comprehensively evaluated. Samples with outliers greater than the outlier threshold are considered unknown faults, otherwise they are considered known faults. The outlier threshold is set to 0.5.

[0090] The effectiveness of the present invention is verified below by analyzing two groups of rotor fault signals.

[0091] Two groups of rotor fault data are used to verify the present invention. The first group is the data collected from the HZXT-DS-001 type test bench of Wuxi Houde Instrument Co., Ltd., designated as dataset A. The vibration signals are collected by the sensor at a frequency of 10 kHz. Five vibration state experiments of normal state, air flow disturbance, mass imbalance, bearing seat looseness, and rotor misalignment are simulated under three operating states of 2400 rpm, 2600 rpm, and 2800 rpm. The second group is the fault data collected from the Bently test bench, designated as dataset B. The vibration signals are collected by the sensor at a sampling frequency of 5120 Hz. Four vibration state experiments of normal state, 1 mm crack fault, 2 mm crack fault, and 3 mm crack fault are simulated under four different operating states of 1300 rpm, 1400 rpm, 1500 rpm, and 1600 rpm:

[0092] Table 1 Datasets

[0093]

[0094] The two groups of rotor signals are analyzed using the present invention, and the specific implementation process is as follows:

[0095] Two migration diagnosis tasks are set using two data sets. For each task, a closed-set DA task, a partial DA task, an open-set DA task, and a partial open-set DA task are designed respectively, and the overall performance of the proposed method is evaluated. The abnormal categories existing in the target domain are regarded as unknown classes, and the settings of tasks A and B are shown in Table 2.

[0096] Table 2 Migration tasks

[0097]

[0098] First, the source domain data is used to generate the source model. The source domain data is input into the constructed multi-classifier residual universal domain adaptation network, and the MRUDA is trained by minimizing the source domain loss function to obtain a well-trained source domain model and save the model parameters, so as to realize the classification of known rotor faults and the identification of unknown faults under the condition of no source domain.

[0099] Then, in the target domain environment, the model parameters saved in the first stage are used to initialize the network, classify the known faults and identify the unknown faults. By minimizing L MRUDA Train the MRUDA network; refine the pseudo-labels using the pseudo-label refinement strategy based on cosine distance to reduce the influence of pseudo-label noise.

[0100] Finally, the optimal network model is obtained by updating the network parameters through multiple iterations. The target domain data is input into the trained fault recognition network for testing, and the test results obtained by comparing with several different networks are shown in Tables 3 and 4 respectively. The average accuracy rate of Task A reaches 96.645%, and the average accuracy rate of Task B is 97.151%, which proves the strong diagnostic ability of the present invention in different task scenarios. The accuracy rate of OSM for all diagnostic tasks is very poor, indicating that the feature distribution difference between the source domain and the target domain is very large. The DDC method has a strong recognition ability for closed-set DA, but the recognition accuracy rate is not high in some cases of DA, open DA, and partially open DA. The average accuracy rates of the SHOT method reach 72.120% and 84.223%. In the case of closed-set DA, its recognition ability is lower than that of the DDC method, but in other tasks, its average performance is better than that of other groups, and it has the ability to recognize unknown faults. The present invention is competitive in both extreme task scenarios of closed-set DA and partial DA tasks. For open-set DA tasks and partially open DA tasks, the diagnostic accuracy rate of the method in this paper is higher than that of other experimental methods, and the recognition accuracy rate can reach 100% in 11 groups of experiments. For Task A, the proposed model has the ability to recognize different fault categories, and for Task B, the proposed model has the ability to recognize different fault sizes. These results not only verify the application value of the method in this paper in multiple task scenarios, but also provide new ideas and methods for solving complex fault diagnosis problems in the industrial field.

[0101] Table 3 Recognition accuracy rate of Task A (%)

[0102]

[0103] Table 4 Recognition accuracy rate of Task B (%)

[0104]

[0105] Take Task A 8 as a case for performing feature visualization to better understand the advantages of the proposed method. The diagnostic results on Task A 8 are presented using a confusion matrix. As Figure 4As shown in (a), the model using OSM for direct diagnosis can only accurately identify two fault categories and does not have the ability to diagnose unknown fault categories. The method using DDC for domain adaptation can effectively identify two faults. However, due to the class mismatch between the source domain and the target domain, some faults are identified as the classes existing in the source domain rather than the fault classes in the target domain. When using the DANCE method for domain adaptation of the target domain, there is a common situation that the target domain is identified as the private class of the source domain. Although the domain adaptation process of the target domain is carried out and the unknown classes are diagnosed by reducing the entropy, the results show that the effective diagnosis of unknown class faults has not been achieved. Although the SHOT method fails to separate the known classes more effectively, it can partially identify unknown faults. The method proposed in the present invention can well identify the known and unknown classes existing in the target domain and achieve the highest diagnostic performance.

[0106] Therefore, the present invention can effectively mine the common features between different working conditions of the same fault, and can judge the features of unknown faults, more accurately identify the unknown faults of the rotor under cross-working conditions, and achieve accurate diagnosis of the rotor fault state.

[0107] The present invention has the following advantages:

[0108] a. The present invention can perform fault diagnosis when the source domain data is unavailable and the target fault categories are unknown. Through domain adaptation of the target domain by a pre-trained model, known fault diagnosis and unknown fault identification are achieved.

[0109] b. In order to identify unknown faults and align known classes, the present invention proposes a multi-classifier residual general domain adaptation network model, which includes two stages: source model generation and model domain adaptation. The source domain data is input into the designed multi-classifier residual network, the source model is trained, and the model parameters are saved; the network parameters of the source model are used to initialize the network parameters of the multi-classifier residual network in the target domain, and the L MRUDA loss function is introduced to train the network, and a pseudo-label refinement strategy based on cosine distance is introduced to obtain a fault diagnosis model.

[0110] c. The present invention uses cosine distance to refine the pseudo-labels, corrects the pseudo-labels by using the labels of the ten samples closest to the distance samples, reduces the influence of a large number of incorrect pseudo-labels on the process of source-free general domain adaptation, and improves the diagnostic ability of the model.

Claims

1. A rotor unknown fault diagnosis method across operating conditions, characterized in that: The method comprises the following steps: a. Data sampling and preprocessing: The collected vibration signals of rotating machinery under different working conditions are cut and short-time Fourier transformed to obtain two-dimensional time-frequency domain signals. The two-dimensional time-frequency domain signals are used to construct a labeled source domain data set and an unlabeled target domain data set. The target domain data set includes the fault types that exist and do not exist in the source domain data set; b. Construct a multi-classifier residual universal domain adaptation network model, including two stages: source model generation and model domain adaptation, to obtain a fault diagnosis model; c. Input the unlabeled target domain data into the fault diagnosis model and output the diagnosis results. Samples with outliers greater than the outlier threshold are considered unknown faults, otherwise they are considered known faults.

2. A rotor unknown fault diagnosis method across operating conditions according to claim 1, characterized in that: The specific process of generating the source model is as follows: First, the labeled source domain dataset is input into the constructed multi-classifier residual network. The multi-classifier residual network includes a feature extractor based on a convolutional attention module, an open set classifier and a closed set classifier. The closed set classifier uses a softmax function for output, and the open set classifier is composed of L c Sub-classifiers are composed, and a classifier is set for each type of sample. For the i-th classifier, the i-th type of sample is identified as positive, and any other type of sample is considered negative; Then calculate the total loss L for source domain training s , back-propagating model parameters until the maximum iteration parameter is reached, and the closed set classifier is trained using the cross entropy loss function to classify known classes. The cross entropy loss function is expressed as: For the open set classifier, choose l ova (x s ,y s ) is used as the open set classification loss, expressed as: Where C is the number of rotor fault conditions, N is the number of samples, and y i is the true label of the i-th sample, x s is the input sample, y s is the true label of the input sample, p represents the given input x s When the sample is identified as y s The probability of is the predicted output of the model, the predicted label It is obtained by the softmax function in fault classification and is written as follows: w i is the score of the model for the i-th category, and the total loss of source domain training is: L s =loss clc +l ova (x s ,y s ) Finally, save the model parameters to obtain the trained source domain model.

3. A rotor unknown fault diagnosis method across operating conditions according to claim 2, characterized in that: The specific process of model domain adaptation is: ① Input the unlabeled target domain dataset into the multi-classifier residual network with initialized parameters to perform passive universal domain adaptation process; ② Use the pseudo-label refinement strategy based on cosine distance to refine the pseudo-labels obtained during the model adaptation process, reduce the impact of noise, and obtain more accurate pseudo-labels for model training; ③Calculate L MRUDA Loss function, minimize the loss function and back propagate to obtain the fault diagnosis model.

4. The method for diagnosing unknown rotor faults across operating conditions according to claim 3 is characterized in that: The specific process of the pseudo-label refinement strategy based on cosine distance is: Ⅰ. Input the unlabeled target domain dataset into the multi-classifier residual network with initialized parameters to obtain the initial pseudo-label; Ⅱ. Calculate the cosine distance between the target sample and other samples, and select the 10 samples with the closest distance as adjacent samples; III. Take the average of the prediction scores of adjacent samples to obtain the average score vector; IV. Using the average score vector, the accurate pseudo-label is calculated using the argmax operation and used for self-supervised target samples.

5. A rotor unknown fault diagnosis method across operating conditions according to claim 4, characterized in that: Calculate L MRUDA The specific process of the loss function is: For the closed set classifier, the classification loss function with entropy reweighting is used for training, and different weights are assigned to pseudo labels with different reliability to reduce the impact of wrong labels. The calculation model is used to calculate the target sample x i The predicted probability distribution of The entropy of is: Where C is the number of rotor fault conditions, i is the index of the sample, is the model for sample x i The predicted probability of belonging to the cth category, normalize the entropy, and get the sample weight: is the sample x i The weight of , the reweighted classification loss is: Through the loss reweighting strategy, pseudo labels with high uncertainty are penalized, and model training is guided by reliable pseudo labels; The open set classifier is trained using the average entropy loss function, and the average entropy of all binary classifiers is Where x t is the target domain input sample, is the predicted label of the sample; L MRUDA The loss function is L MRUDA =loss ent (x t )+loss wclc .

6. The method for diagnosing unknown rotor faults across operating conditions according to claim 1 is characterized in that: The abnormal threshold is set to 0.5.

Images