Robust joint sub-domain network gearbox fault diagnosis method based on noise label

Abstract

The present application relates to the field of fault diagnosis, in particular to a kind of joint sub-domain network gearbox fault diagnosis method based on noise label robustness, the present application will combine the deep sub-domain adaptive network in unsupervised field with label noise representation learning, propose joint sub-domain network gearbox fault diagnosis method based on noise label robustness;Through deep sub-domain adaptive network, fine-grained local alignment instead of global alignment between different fault types under cross-domain condition is realized;At the same time, by introducing label noise representation learning into joint sub-domain network, the noise label problem caused by poor recognizable or wrong labeling of marked fault samples in the process of fault diagnosis is effectively alleviated, and the fault diagnosis precision is further improved.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis, and in particular to a fault diagnosis method for a joint subdomain network gearbox based on noise label robustness. Background Technology

[0002] With the continuous upgrading and transformation of global manufacturing, mechanical equipment is also becoming increasingly complex through constant iteration and updates. As a key component of rotating machinery, gearboxes frequently need to undergo repeated changes in operating conditions based on actual needs. Therefore, under various alternating loads, internal parts of the gearbox can experience various types of damage, leading to gearbox failure and even incalculable economic losses. Therefore, conducting research on gearbox fault diagnosis to ensure the stable and safe operation of mechanical equipment is crucial for industrial production and other fields.

[0003] Generally, there are three methods for gearbox fault diagnosis: 1. Model-based methods; 2. Signal processing-based methods; and 3. Data-driven methods. Model-based methods simulate fault types by building a dynamic simulation model of a complex gearbox and incorporating mechanistic knowledge. Signal processing-based methods require expertise in signal processing to uncover the potential correlation between signals and fault types. Data-driven methods primarily use big data to build intelligent identification models, automatically revealing the feature mapping relationship between data and fault types, without requiring complex simulation models or specialized knowledge.

[0004] Recently, data-driven methods based on deep learning have attracted researchers' attention and been widely applied in gearbox fault diagnosis. However, the success of existing deep learning-based fault diagnosis models depends on two conditions: (1) a large amount of fault data is available for model training; and (2) the training data and test data share the same data distribution. However, in actual industrial applications, gearboxes need to operate under different working conditions according to actual needs. Therefore, the distribution of fault data collected under different working conditions varies. In addition, collecting and labeling all fault data under each working condition is quite resource-intensive and manpower-intensive. Therefore, it is difficult to establish an effective fault diagnosis model under working conditions without fault label information.

[0005] Transfer learning provides researchers with a viable path to transfer fault information from one domain (operating condition) to another through paradigms such as distance metrics or adversarial training, thus constructing a transfer fault diagnosis model. Currently, transfer learning has seen many successful applications in gearbox fault diagnosis. However, these successful models still have some shortcomings. Specifically, when labeling data, the subjectivity of professionals and the poor identifiability of fault samples inevitably produce noisy labels (incorrectly classifying faults). When there are many noisy labels, the network will continuously fit noisy labeled fault samples from the source domain during training, thereby transmitting incorrect fault information to the target domain, establishing a low-performance transfer fault diagnosis model, and resulting in low accuracy in gearbox fault diagnosis and identification.

[0006] Therefore, how to improve the low accuracy of fault diagnosis caused by the presence of noise labels in the source domain in cross-domain gearbox fault diagnosis has become an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0007] To address the problem of noise-labeled faults in the source domain, existing cross-domain gearbox fault diagnosis methods cannot handle this issue. This invention proposes a noise-label-robust joint subdomain network-based gearbox fault diagnosis method. In the proposed noise-label-robust joint subdomain network, three deep subdomain adaptive networks are constructed to filter noise-label-free samples based on the prediction consistency principle. This reduces the impact of noise-labeled fault samples on network training, thus minimizing the impact on fault diagnosis accuracy. Furthermore, the filtered noise-label-free fault sample set is processed using a maximum mean local difference metric loss to facilitate fault information transfer between different operating conditions, ultimately achieving effective and reliable gearbox fault diagnosis in the target domain.

[0008] To achieve the above objectives, this invention provides a noise-label-based robust joint subdomain network gearbox fault diagnosis method, the method comprising:

[0009] Vibration data of the gearbox under multiple operating conditions are collected using an accelerometer, and the vibration data is preprocessed. Specifically, based on whether the vibration data is labeled, labeled data under the same operating condition can be used as the source domain training set sample, and unlabeled data under another operating condition can be used as the target domain training set sample and the target domain test set sample.

[0010] Construct three deep subdomain adaptive networks with identical network structures, and initialize the parameters of each network with different parameters;

[0011] High-dimensional feature representations of source domain training set samples and target domain training set samples are extracted using feature extractors of adaptive networks in each deep subdomain.

[0012] By utilizing the sub-classifiers of adaptive networks in various deep subdomains, noise-free labeled source domain training set samples and reliable target domain training set samples are selected based on the prediction consistency principle.

[0013] The distance between the high-dimensional features of the noise-free labeled source domain training set samples and the trustworthy target domain training set samples is measured by the local maximum mean difference distance metric, and a joint subdomain network loss is constructed.

[0014] By optimizing the joint subdomain network loss, a noise-label-robust joint subdomain network is trained, which is composed of three deep subdomain adaptive networks.

[0015] The target domain test set samples are input into a pre-trained noise-labeled robust joint subdomain network, which outputs the fault identification results of the vibration data of the gearbox under test.

[0016] The beneficial effects of this invention are:

[0017] This invention addresses the problem of low accuracy in gearbox fault diagnosis due to noise labels in the source domain during cross-domain fault diagnosis. It proposes a noise-label-robust joint subdomain network-based gearbox fault diagnosis method. By constructing three deep subdomain adaptive networks, noise-label-free sample selection is achieved, improving the accuracy of cross-domain gearbox fault diagnosis under the influence of noise labels. Experimental results from embodiments of this invention show that:

[0018] (1) This method uses three deep subdomain adaptive networks to filter source domain training set samples and credible target domain training set samples with no noise labels according to the prediction consistency principle, thereby improving the reliability of the training set samples.

[0019] (2) To address the problem of poor gearbox fault diagnosis performance when there are noise-labeled fault samples in the source domain and fault samples may be incorrectly predicted in the target domain, a noise-labeled robust joint subdomain network gearbox fault diagnosis method is proposed, which effectively improves the accuracy of cross-domain gearbox fault diagnosis. Attached Figure Description

[0020] To more clearly illustrate the objectives, technical solutions, and advantages of this invention, the invention will be further described in detail below with reference to the accompanying drawings, wherein:

[0021] Figure 1 This is a flowchart of the fault diagnosis method for a joint subdomain network gearbox based on noise tag robustness according to the present invention;

[0022] Figure 2 A schematic diagram of the structure of a noise-label-based robust joint subdomain network (LRJSN).

[0023] Figure 3Simulation plot showing the trend of accuracy and variance under a 20% noise sample ratio;

[0024] Figure 4 The simulation graph shows the trend of accuracy variation and variance under a 40% noise sample ratio. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0026] In gearbox cross-domain fault diagnosis, the presence of noisy labeled samples in the source domain causes the network to transmit incorrect fault information to the target domain, resulting in low accuracy in gearbox cross-domain fault diagnosis. Existing transfer models are therefore inapplicable.

[0027] In an embodiment of the present invention, Figure 1 This is a flowchart illustrating the fault diagnosis method for gearboxes in joint subdomain networks based on noise label robustness according to the present invention; as shown. Figure 1 As shown, the method mainly includes two processes: training a noise-label-robust joint subdomain network and gearbox fault diagnosis.

[0028] In the training of the noise-label-robust joint subdomain network, the gearbox vibration data is preprocessed and divided into a labeled source domain training set, an unlabeled target domain training set, and an unlabeled target domain test set. The training set is then input into the network for training. The classifiers of the three deep subdomain adaptive networks select noise-label-free samples based on the output consistency principle. The high-dimensional features corresponding to the selected noise-label-free samples are then used to narrow the high-dimensional distance between the source and target domains through the maximum mean difference distance, thereby enabling the transmission of fault information and ultimately allowing the network to identify the fault type of the fault samples in the target domain.

[0029] In cross-domain fault diagnosis of gearboxes, it is necessary to preprocess the vibration data of the gearbox to be tested; input the preprocessed data into the noise-labeled robust joint subdomain network after training; and output the fault diagnosis result of the vibration signal of the gearbox to be tested.

[0030] Specifically, in this embodiment of the invention, a method for fault diagnosis of a joint subdomain network gearbox based on noise label robustness includes the following steps:

[0031] Vibration data of the gearbox under multiple operating conditions are collected using an accelerometer, and the vibration data is preprocessed.

[0032] In this embodiment of the invention, depending on whether the vibration data is labeled, labeled data under the same working condition is used as the source domain training set sample, and unlabeled data under another working condition is used as the target domain training set sample and the target domain test set sample. The source domain training set sample and the target domain training set sample are mainly used to jointly train the deep subdomain adaptive network, so that the trained network model can accurately predict the fault identification result of the target domain test set sample.

[0033] In this embodiment of the invention, the process of preprocessing the vibration data may include performing a fast Fourier transform on the acquired gearbox time-domain vibration data to convert it into a one-dimensional frequency-domain vibration signal, and then creating corresponding source domain training set samples, target domain training set samples, and target domain test set samples under the one-dimensional frequency-domain vibration signal.

[0034] In this embodiment of the invention, the gearbox vibration data can be fault data from existing publicly available datasets or actually collected fault data. The source domain training set samples can be gearbox fault data with known fault category labels, and the target domain training set samples and target domain test set samples can be gearbox fault data with unknown fault category labels. For example, gearbox fault data for operating conditions A and B can be used as source domain training set samples, gearbox fault data for operating condition C can be used as target domain training set samples, and gearbox fault data for operating condition D can be used as target domain test set samples. Alternatively, planetary gearbox fault data of type a and type b can be used as source domain training set samples, and gearbox fault data of type c can be used as target domain training set samples and target domain test set samples. Here, "type" can refer to different gearboxes produced by different manufacturers.

[0035] Construct three deep subdomain adaptive networks with identical network structures, and initialize the parameters of each network with different parameters;

[0036] In this embodiment of the invention, in order to improve the recognition effect of the network model trained by the training set samples, the present invention uses three deep subdomain adaptive networks to filter the training set samples, so that the filtered training set samples can better train the network model and improve the recognition accuracy of the network model. Among them, the network structure of each deep subdomain adaptive network is the same, but the initialization of the three network parameters is different, so different prediction results can be output for each training set sample.

[0037] High-dimensional feature representations of source domain training set samples and target domain training set samples are extracted using feature extractors of adaptive networks in each deep subdomain.

[0038] By utilizing the sub-classifiers of adaptive networks in various deep subdomains, noise-free labeled source domain training set samples and reliable target domain training set samples are selected based on the prediction consistency principle.

[0039] In embodiments of the present invention, such as Figure 2 As shown, the deep subdomain adaptive network includes a feature extractor and a subclassifier. The feature extractor can adopt the common VGGNet network architecture, and the subclassifier can also be a common classifier structure. After inputting the source domain training set samples and the target domain training set samples into the feature extractor of the deep subdomain adaptive network, high-dimensional feature representations of the source domain training set samples and the target domain training set samples can be extracted. These high-dimensional feature representations contain rich features about gearbox faults in the training set samples. Inputting these high-dimensional feature representations into the subclassifier allows for label prediction of the source domain training set samples and the target domain training set samples. Since the source domain training set samples are labeled, while the target domain training set samples are unlabeled, the source domain training set samples are hard-labeled, and the label prediction result of the target domain training set samples is a soft pseudo-label. The present invention further performs consistency screening on the training set samples based on the label prediction results of the subclassification of each deep subdomain adaptive network. The so-called prediction consistency principle specifically selects samples with consistent classification results from any two networks as noise-free labeled source domain fault samples and reliable target domain fault samples, which participate in the network parameter optimization of the third network. Since the prediction results of source domain fault samples and / or trusted target domain fault samples without noise labels are generally stable and beneficial to network convergence, while the prediction results of training set samples with noise labels are unstable, an excessive number of noisy labeled samples will not only affect network convergence but also lead to incorrect predictions of target domain fault samples; therefore, in this embodiment of the invention, identical source domain training set samples and target domain training set samples are respectively input into any two deep subdomain adaptive networks; if the classification results are consistent, the source domain training set samples and target domain training set samples with consistent classification results are used as noise-free source domain training set samples and trusted target domain training set samples, respectively, and the selected noise-free source domain training set samples and trusted target domain training samples are input into a third deep subdomain adaptive network for training; if the classification results are inconsistent, the source domain training set samples and target domain training set samples with inconsistent classification results are discarded as noisy labeled source domain training set samples and untrusted target domain training set samples.

[0040] In this embodiment of the invention, the formula used to select noise-free labeled source domain training set samples and reliable target domain training set samples is expressed as follows:

[0041]

[0042] In the formula, These represent the noise-free labeled source domain training set samples selected by the three deep subdomain adaptive networks, in that order. These represent, in order, the training set samples of the trustworthy target domain selected by the three deep subdomain adaptive networks. These represent the hard labels of the source domain training set samples selected by the sub-classifiers in each deep subdomain adaptive network, respectively. These represent the soft pseudo-labels of the target domain training set samples selected by the sub-classifiers in each deep subdomain adaptive network, respectively. These represent the number of source domain training set samples selected by the sub-classifiers in each deep subdomain adaptive network, respectively. These represent the number of target domain training set samples selected by the sub-classifiers in each deep subdomain adaptive network.

[0043] The formula used to predict training set samples using three deep subdomain adaptive networks is as follows:

[0044]

[0045] This represents the prediction output of each classifier for the source domain samples and the prediction output for the target domain samples.

[0046] The distance between the high-dimensional features of the noise-free labeled source domain training set samples and the trustworthy target domain training set samples is measured by the local maximum mean difference distance metric, and a joint subdomain network loss is constructed.

[0047] In this embodiment of the invention, the Local Maximum Mean Difference Distance (LMMD) is used to measure the distance between the high-dimensional features of the noise-free labeled source domain fault samples and the target domain fault samples. The LMMD formula can be expressed as:

[0048]

[0049] Among them, D H (p,q) represents the high-dimensional feature between a noise-free labeled source domain training sample p and a trustworthy target domain training sample q. The weights of the training set samples represent the training set samples when they belong to the Class category. The weights of the training set samples that represent the target domain training set samples when they belong to the Class category. Φ(·) represents the weight of the training set sample when it belongs to the Class category, Φ(·) represents the kernel function, which can be a Gaussian kernel or other kernel functions, and p and q represent the class distributions in the source domain training set samples and the target domain training set samples, respectively; x s,i D represents the training set samples in the source domain; s Represents the set of samples in the source domain training set; Indicates x s,iFeature extraction results of deep subdomain adaptive networks; x t,i This represents the training set samples of the target domain; Indicates x t,i Feature extraction results of deep subdomain adaptive networks; H This represents the H-norm.

[0050] In a preferred embodiment of the present invention, considering the inconsistent distribution of the source domain training set samples and the target domain training set samples, that is, the inconsistent feature positions and distributions of the high-dimensional features of the source domain training set samples and the high-dimensional features of the target domain training set samples in the high-dimensional feature space, in order to enable the network to learn highly correlated and diverse high-dimensional feature knowledge matching relationships through the training set samples, the present invention adaptively adjusts the Gaussian kernel radius, so that the Gaussian kernel radius changes adaptively with the sample distance between the target domain training set samples and the source domain training set samples; this solves the problem of inconsistent feature distribution between the source domain training set samples and the target domain training set samples, improves the accuracy of the feature extractor and the diversity of feature knowledge, and enhances the diagnostic accuracy of the network model.

[0051] The adaptive Gaussian kernel function is expressed as follows:

[0052]

[0053]

[0054] in, The variable radius Gaussian kernel function is represented. The variance represents the feature distribution of source domain fault samples in the training set. It represents the variance of the feature distribution of fault samples in the target domain in the training set.

[0055] Based on the above distance, the joint subdomain network loss function of the present invention can be expressed as:

[0056]

[0057] in, The following are the joint subdomain network losses of the adaptive networks for each depth subdomain, respectively. These represent the number of source domain training set samples selected by the sub-classifiers in each deep subdomain adaptive network, respectively. These represent samples x from the same source domain training set in each of the deep subdomain adaptive networks, respectively. s,i The corresponding tag; These represent the same source domain training set samples input into each deep subdomain adaptive network, in turn. These represent the training set samples of the same target domain input to each deep subdomain adaptive network, respectively; F(·) represents the feature extractor of the deep subdomain adaptive network; C(·) represents the subclassifier of the deep subdomain adaptive network; λ represents the weight factor; Φ * (·) represents the adaptive Gaussian kernel function proposed in this invention. Representative source domain training set sample x s,i The training set sample weights when the sample belongs to the Class category. x represents the training set sample of the target domain t,i Training set sample weights when belonging to the Class category; D H (p,q) represents the distance between high-dimensional features of a noise-free labeled source domain training set sample p and a trusted target domain training set sample q. Here, p and q are just general terms. The specific parameters are calculated according to the above calculation formula. E[·] represents the expected value of the local maximum mean difference distance under all fault categories.

[0058] By optimizing the joint subdomain network loss, a noise-label-robust joint subdomain network is trained, which is composed of three deep subdomain adaptive networks.

[0059] The target domain test set samples are input into a pre-trained noise-labeled robust joint subdomain network, which outputs the fault identification results of the vibration data of the gearbox under test.

[0060] In this embodiment of the invention, the target domain test set samples can be input into any one of the deep subdomain adaptive networks in the joint subdomain network. After the above training, the structures of the three deep subdomain adaptive networks are consistent, so consistent diagnostic results can be obtained by inputting them into any one of the networks.

[0061] To verify the effectiveness of the open-set fault diagnosis method disclosed in this invention, gearbox fault vibration data collected on a transmission system diagnostic simulation test bench was used. The test bench includes a motor, a planetary gearbox, a parallel shaft gearbox, and a magnetic powder brake. Several gearbox faults under different operating conditions were simulated by changing the motor speed and the load on the magnetic powder brake. Eight different gearbox health states of the bearings and gears within the parallel shaft gearbox were simulated on the transmission system diagnostic simulation test bench under four different operating conditions (as summarized in Tables 1 and 2).

[0062] Table 1. Description of Planetary Gearbox Status

[0063]

[0064]

[0065] Table 2. Description of operating conditions for planetary gearbox fault datasets.

[0066]

[0067] Based on the available gearbox fault data, the noise-label-based robust joint subdomain network proposed in this invention is compared with the DSAN model in IEEE Trans. Neur.Net.Lear.Syst., May. 2020, “Deep subdomain adaptation network for image classification”, the ButterflyNet model in Proc. ArXivabs / 1905.07720., May. 2019, “Butterfly: A panacea for all difficulties in wildly unsupervised domain adaptation.”, and the GearNet model in Proc., ArXivabs / 2201.06001., “GearNet: Stepwise dual learning for weakly supervised domain adaptation.”. To ensure experimental fairness, all the above methods use the same feature extractor and classifier structure.

[0068] In this experiment, 12 cross-domain fault diagnosis tasks for gearboxes were set up, and diagnostic experiments were carried out simultaneously under two noise label sample ratios (20% and 40% source domain noise sample ratios). We also simulated three different operating conditions: 1. Same speed, but different load variations; 2. Same load, but different speed variations; 3. Different speeds, and different load variations. To obtain stable experimental results and minimize the impact of randomness, each method was tested five times, with 30 model training iterations per test. Specific fault diagnosis accuracy results are shown in Tables 3 and 4, respectively, for noise sample ratios of 20% and 40%.

[0069] Table 3. Fault diagnosis accuracy at a 20% noise sample ratio

[0070]

[0071] Table 4. Fault diagnosis accuracy at a 40% noise sample ratio

[0072]

[0073] It can be seen that when the noise sample ratio is 20% and 40%, DSAN, due to the presence of noise samples in the source domain, transmits erroneous fault information to the target domain, resulting in the lowest fault diagnosis accuracy among all methods, with accuracy rates of only 86.97% and 70.75%. Figure 3 and Figure 4 In the initial analysis, DSAN's diagnostic accuracy peaked, but gradually decreased due to the misinterpretation of fault information from noisy samples. ButterflyNet outperformed DSAN, improving accuracy by 3.33% and 5.15% at LNRs of 20% and 40%, respectively. GearNet also outperformed DSAN, improving accuracy by 5.77% and 6.37%, respectively. Both ButterflyNet and GearNet effectively mitigated the impact of noisy labels by employing a selection strategy based on low-loss samples. Compared to the aforementioned methods, the noise-label-robust joint subdomain network uses prediction consistency across multiple networks as a sample selection strategy, replacing the low-loss-based sample selection strategy in collaborative teaching, thus avoiding the proportional threshold setting required in low-loss sample selection strategies. At LNRs of 20% and 40%, TJSAN achieved fault diagnosis accuracies of 94.77% and 79.84%, respectively, outperforming basic DSAN by 7.80% and 9.09%, and outperforming the current best method, GearNet, by 1.02% and 2.72%, respectively.

[0074] This invention addresses the low fault diagnosis accuracy in gearbox cross-domain fault diagnosis caused by noise-labeled fault samples in the source domain. It proposes a noise-label-robust joint subdomain network-based gearbox fault diagnosis method. Experimental results show that:

[0075] (1) The three deep subdomain adaptive networks constructed can effectively screen out noise-free labeled samples according to the prediction consistency principle, which effectively improves the problem of low fault diagnosis accuracy caused by noise-labeled samples.

[0076] (2) The proposed noise-robust joint subdomain network can effectively prevent the network from transmitting incorrect fault sample information to the target domain, while narrowing the maximum mean difference distance, enabling fine-grained alignment of fault types in the source and target domains, and effectively improving the gearbox fault diagnosis accuracy.

[0077] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include ROM, RAM, disk, or optical disk, etc.

[0078] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A noise-label-based robust joint subdomain network gearbox fault diagnosis method, characterized in that, The method includes: Vibration data of the gearbox under multiple operating conditions are collected using an accelerometer, and the vibration data is preprocessed. Depending on whether the vibration data is labeled, labeled data under the same operating condition is used as the source domain training set sample, and unlabeled data under another operating condition is used as the target domain training set sample and the target domain test set sample. Construct three deep subdomain adaptive networks with identical network structures, and initialize the parameters of each network with different parameters; High-dimensional feature representations of source domain training set samples and target domain training set samples are extracted using feature extractors of adaptive networks in each deep subdomain. Using the sub-classifiers of each deep subdomain adaptive network, noise-free labeled source domain training set samples and reliable target domain training set samples are selected based on the prediction consistency principle. Specifically, this involves inputting identical source domain training set samples and target domain training set samples into any two deep subdomain adaptive networks. If the classification results are consistent, the source domain training set samples and target domain training set samples with consistent classification results are used as noise-free labeled source domain training set samples and reliable target domain training set samples, respectively. Then, the selected noise-free labeled source domain training set samples and reliable target domain training set samples are input into a third deep subdomain adaptive network for training. If the classification results are inconsistent, the source domain training set samples and target domain training set samples with inconsistent classification results are discarded as noisy labeled source domain training set samples and unreliable target domain training set samples. The distance between the high-dimensional features of the noise-free labeled source domain training set samples and the trustworthy target domain training set samples is measured by the local maximum mean difference distance metric, and a joint subdomain network loss is constructed. By optimizing the joint subdomain network loss, a noise-label-robust joint subdomain network is trained, which is composed of three deep subdomain adaptive networks. The target domain test set samples are input into a pre-trained noise-labeled robust joint subdomain network, which outputs the fault identification results of the vibration data of the gearbox under test.

2. The method for fault diagnosis of a joint subdomain network gearbox based on noise label robustness according to claim 1, characterized in that, Data preprocessing of the gearbox vibration data includes converting the acquired time-domain vibration data of the gearbox into a frequency-domain signal after fast Fourier transform.

3. The method for fault diagnosis of a joint subdomain network gearbox based on noise label robustness according to claim 1, characterized in that, The formula used to select noise-free labeled source domain training set samples and reliable target domain training set samples is expressed as follows: ， In the formula, , , These represent the noise-free labeled source domain training set samples selected by the three deep subdomain adaptive networks, in that order. , , These represent, in order, the training set samples of the trustworthy target domain selected by the three deep subdomain adaptive networks. , , These represent the hard labels of the source domain training set samples selected by the sub-classifiers in each deep subdomain adaptive network, respectively. , , These represent the soft pseudo-labels of the target domain training set samples selected by the sub-classifiers in each deep subdomain adaptive network, respectively. , , These represent the number of source domain training set samples selected by the sub-classifiers in each deep subdomain adaptive network, respectively. , , These represent the number of target domain training set samples selected by the sub-classifiers in each deep subdomain adaptive network.

4. The method for fault diagnosis of a joint subdomain network gearbox based on noise label robustness according to claim 1, characterized in that, The formula used to calculate the distance between the high-dimensional features of noise-free labeled source domain training set samples and trustworthy target domain training set samples selected using the local maximum mean difference distance metric is as follows: ， in, Represents noise-free labeled source domain training set samples training set samples with trusted target domain The distance between high-dimensional features The source domain training set samples belong to Training set sample weights for class classification The training set samples representing the target domain belong to Training set sample weights for class classification Represents the kernel function. These represent the class distributions in the source domain training set samples and the class distributions in the target domain training set samples, respectively. Represents samples from the source domain training set; Represents the set of samples in the source domain training set; Indicates to Feature extraction results of deep subdomain adaptive networks; This represents the training set samples of the target domain; Indicates to Feature extraction results of deep subdomain adaptive networks; This represents the H-norm.

5. The method for fault diagnosis of a joint subdomain network gearbox based on noise label robustness according to claim 1, characterized in that, The joint subdomain network loss is expressed as: ， in, , , The following are the joint subdomain network losses of the adaptive networks for each depth subdomain, respectively. , , These represent the number of source domain training set samples selected by the sub-classifiers in each deep subdomain adaptive network, respectively. , , These represent the training set samples from the same source domain in each deep subdomain adaptive network, respectively. The corresponding tag; , , These represent the same source domain training set samples input into each deep subdomain adaptive network, in turn. , , These represent the training set samples of the same target domain input into each deep subdomain adaptive network, respectively. A feature extractor for a deep subdomain adaptive network; This represents a subclassifier in a deep subdomain adaptive network. Indicates the weighting factor; Represents the kernel function. Representative source domain training set samples belong Training set sample weights for class classification Representative target domain training set samples belong Training set sample weights when classifying; Represents noise-free labeled source domain training set samples training set samples with trusted target domain The distance between high-dimensional features This represents the expected value of the local maximum mean difference distance across all fault categories.

6. The method for fault diagnosis of a joint subdomain network gearbox based on noise label robustness according to claim 5, characterized in that, The kernel function It is an adaptive Gaussian kernel function, whose Gaussian kernel radius changes adaptively with the sample distance between the training set samples in the target domain and the training set samples in the source domain.

Images