Training method and training device for bearing lubrication fault data generation model
By combining conditional cyclic generative adversarial networks and noise-reducing autoencoders, high-fidelity bearing lubrication fault samples are generated, solving the problem of inaccurate samples in existing technologies and improving the accuracy and robustness of lubrication fault diagnosis. This method is applicable to a variety of rotating machinery.
Patent Information
- Application Number
- CN202511608776.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-05
- Publication Date
- 2026-02-03
AI Technical Summary
In existing technologies, the data generation methods for bearing lubrication failures generate inaccurate samples, resulting in poor performance of deep learning models in bearing lubrication condition monitoring and evaluation. This is mainly due to the lack of sufficient labeled data and the difficulty in simulating complex lubrication conditions.
A bearing lubrication fault data generation model is adopted. By collecting data from industrial sites and laboratories, a conditional recurrent generative adversarial network and a noise-reducing autoencoder are used to construct various loss functions for training, generating high-fidelity pseudo-fault samples, including feature consistency loss and style invariant classification loss, to ensure the fidelity of key fault features during the transfer process.
The generated pseudo-samples are of high quality, which can effectively improve the accuracy and robustness of lubrication fault diagnosis, lower the technical threshold, and are suitable for intelligent diagnosis of various rotating machinery.
Smart Images
Figure CN121456473A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent fault diagnosis technology for rolling shafts, and in particular to a training method and training device for a bearing lubrication fault data generation model. Background Technology
[0002] Rotating machinery has wide applications in modern industrial systems, and rolling bearings are one of the most critical components, directly affecting the overall performance of the equipment. Statistics show that 30% of failures in rotating machinery originate from bearing failures, and 80% of these failures are related to bearing lubrication. In fact, most physical damage to bearings, such as wear, fatigue, abrasion, and spalling, stems from direct contact between metal surfaces and the resulting friction and high stress. Proper lubrication can significantly prevent most types of physical damage, thereby substantially extending bearing life. Therefore, ensuring effective bearing lubrication is crucial for the safe and healthy operation of rotating equipment.
[0003] Currently, in the field of bearing health management technology, intelligent fault diagnosis methods for bearing physical damage, especially those based on deep learning, have become a research hotspot and have achieved significant progress. In stark contrast, the development of technologies for monitoring and evaluating bearing lubrication conditions has lagged behind. Analysis of the underlying technology reveals that the effectiveness of deep learning models heavily relies on two fundamental assumptions: the sufficiency of the training dataset (containing a large number of labeled samples) and the consistency of data distribution (the training and test sets are independent and identically distributed). In bearing physical damage diagnosis, these assumptions can be satisfied under certain conditions. However, for dynamic and complex lubrication conditions, obtaining sample data that meets these conditions is extremely difficult, especially in real industrial settings. This directly restricts the in-depth application and development of deep learning technology in this area.
[0004] The root cause of the above problems lies in the significant discrepancy between the manifestations of physical damage and lubrication failure in bearings. Taking bearing vibration signals as an example, when a bearing experiences physical damage, at a given rotational speed, each time an internal component passes the damaged area, a pulse is generated, ultimately manifesting as a fault characteristic frequency in the frequency domain. The specific frequency value depends on the location of the damage. It can be seen that when a bearing experiences physical damage, both the manifestation of the damage and the vibration signal are deterministic. This is highly advantageous for simulation experiments (using lasers or electrical discharge machining to process damage on specific bearing components) or for labeling signal samples in industrial settings. However, when a bearing experiences lubrication failure, the internal manifestation is random metal-to-metal contact between the rolling elements and raceways. The result is merely an increase in the energy of the high-frequency components of the vibration signal, rather than a specific indicator. This necessitates highly precise design for lubrication failure simulation experiments, and the lack of specific bearing lubrication condition indicators makes labeling industrial data extremely challenging.
[0005] Currently, common solutions to the problem of missing target domain faults in intelligent diagnosis of bearing physical damage mainly involve augmenting the target domain fault data. Therefore, a similar approach is considered for addressing the missing target domain faults in intelligent diagnosis of bearing lubrication faults. For example, patent CN117932413A proposes using recurrent adversarial networks for data augmentation; however, the unconditional network used cannot guarantee that the key fault features of the generated samples are not distorted, leading to inaccurate generated samples. The article "Anovel bearing fault data generation strategy combining physical modeling and CycleGAN variant for fault diagnosis without real samples" proposes using a physical simulation model to generate idealized fault simulation signals, followed by style transfer based on a CycleGAN variant to obtain lubrication fault samples. However, unlike bearing physical damage simulation, the physical mechanisms behind lubrication faults are far more complex, therefore the generated fault simulation signals are inherently idealized and inaccurate, resulting in inaccurate lubrication fault samples. Summary of the Invention
[0006] Based on the shortcomings of the existing technology, the present invention provides a training method and training device for a bearing lubrication fault data generation model, which solves the problem of inaccurate fault samples obtained by existing fault data augmentation methods.
[0007] The present invention adopts the following technical solution: In a first aspect, the present invention provides a training method for a bearing lubrication fault data generation model. The generation model includes a first generator, a first discriminator, a second generator, and a second discriminator. The training method includes the following steps: Collect real samples of the target domain from bearings operating under normal lubrication conditions in industrial settings, and real samples of multiple source domains from the same bearings operating under different lubrication conditions in the laboratory. The source domain real samples are used to generate pseudo target domain samples through the first generator, and the pseudo target domain samples are used to generate source domain reconstructed samples through the second generator; the target domain real samples are used to generate pseudo source domain samples through the second generator, and the pseudo source domain samples are used to generate target domain reconstructed samples through the first generator. A cycle consistency loss is constructed based on source domain real samples, source domain reconstructed samples, target domain real samples, and target domain reconstructed samples. An adversarial loss and an identity loss are constructed based on source domain real samples, pseudo target domain samples, target domain real samples, and pseudo source domain samples. A feature consistency loss is constructed based on the features of source domain real samples, source domain reconstructed samples, target domain real samples, and target domain reconstructed samples extracted by the first discriminator and the second discriminator. A style invariant classification loss is constructed based on pseudo target domain samples and their corresponding real labels. The total loss is the sum of adversarial loss, cycle consistency loss, identity loss, feature consistency loss, and style invariant classification loss. The generative model is trained based on the total loss to obtain the trained generative model.
[0008] Preferably, the feature consistency loss is as follows: ; In the formula, L Feat For feature consistency loss, G For the first generator, F For the second generator, As the first discriminator, For the second discriminator, As the expected value, For real samples from the source domain, For real samples in the target domain, P data（A） Let represent the distribution of all real samples in the source domain. P data（B） The distribution of all real samples in the target domain. yes Extracted source domain samples The deep features, yes Extracted reconstructed samples The deep features, yes Extracted target domain samples The deep features, yes Extracted reconstructed samples The deep features, A This is a source domain experimental dataset containing multiple lubrication state categories. B This is a target domain field dataset that includes unique background noise and the influence of transmission paths.
[0009] Preferably, the style invariant classification loss is as follows: ; In the formula, L style For style invariance classification loss, for The true category label, for right and via the first generator G Generated pseudo-target domain samples The predicted class probability after classification.
[0010] Preferably, the total loss is as follows: ; In the formula, To combat the losses, For cycle consistency loss, For the loss of identity, , , and These are the weight parameters.
[0011] Preferred options also include: Multiple real samples from the source domain are input into the trained generative model to obtain a pseudo dataset of the target domain. After data augmentation of the target domain pseudo dataset, the data is input into the autoencoder for training, resulting in the trained autoencoder. The samples in the target domain pseudo dataset are input into the trained autoencoder to obtain the hidden layer representation. The hidden layer representation is then input into the classifier for training to obtain the trained classifier. The signal to be tested in the industrial field is input into the trained autoencoder and classifier to obtain the corresponding lubrication status.
[0012] Preferably, the step of inputting the data augmentation of the target domain pseudo-dataset into the autoencoder for training includes the following steps: Gaussian noise is applied to the target domain pseudo dataset; The target domain pseudo dataset after applying Gaussian noise is input into the autoencoder, where the encoder outputs the hidden layer representation, and the decoder outputs the reconstructed data based on the hidden layer representation; The mean squared error is constructed using a pseudo dataset of the target domain and reconstructed data, and the autoencoder is trained by minimizing the mean squared error.
[0013] Secondly, the present invention provides a training device for a bearing lubrication fault data generation model. The generation model includes a first generator, a first discriminator, a second generator, and a second discriminator. The training device includes: The acquisition module is used to acquire real samples of the target domain of bearings in the industrial field under normal lubrication conditions, as well as real samples of multiple source domains of the same bearings in the laboratory under different lubrication conditions. The generation module is used to generate pseudo-target domain samples from real source domain samples using a first generator, and to generate source domain reconstructed samples from pseudo-target domain samples using a second generator; and to generate pseudo-source domain samples from real target domain samples using a second generator, and to generate target domain reconstructed samples from pseudo-source domain samples using a first generator. The module is used to construct cycle consistency loss based on source domain real samples, source domain reconstructed samples, target domain real samples, and target domain reconstructed samples; to construct adversarial loss and identity loss based on source domain real samples, pseudo target domain samples, target domain real samples, and pseudo source domain samples; to construct feature consistency loss based on the features of source domain real samples, source domain reconstructed samples, target domain real samples, and target domain reconstructed samples extracted by the first discriminator and the second discriminator; and to construct style invariant classification loss based on pseudo target domain samples and their corresponding real labels. The training module uses the sum of adversarial loss, cycle consistency loss, identity loss, feature consistency loss, and style-invariant classification loss as the total loss, and trains the generative model based on this total loss to obtain the trained generative model. Compared with the prior art, the above-mentioned at least one technical solution adopted by the present invention can achieve the following beneficial effects: This invention first collects real samples of the target domain from bearings operating under normal lubrication conditions in industrial settings, as well as multiple real samples of the same bearings in the laboratory under different lubrication conditions. Adopting an "experiment-to-field" paradigm, it uses real experimental data as the source, avoiding the source distortion problem caused by oversimplification in physical simulation models. Then, based on the features extracted by the first and second discriminators—the real samples of the source domain, the reconstructed samples of the source domain, the real samples of the target domain, and the reconstructed samples of the target domain—a feature consistency loss is constructed. This requires the reconstructed signal to remain consistent with the original signal within the deep feature space extracted by the discriminator, achieving deep content preservation. Based on pseudo-target domain samples and their corresponding real labels, a style-invariant classification loss is constructed, forcing the discriminator to have the ability to reverse-engineer the original source domain fault category from a pseudo-sample that has been converted to the target domain style. By introducing feature consistency loss and style-invariant classification loss constraints, this invention ensures high fidelity of key fault features during style transfer, resulting in higher quality pseudo-samples. Attached Figure Description
[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0015] Figure 1 This is a flowchart of a training method for a bearing lubrication fault data generation model according to the present invention; Figure 2 This is a diagram of the conditional cyclic adversarial generative network structure of the present invention; Figure 3 This is a schematic diagram illustrating the source domain identity loss of the present invention; Figure 4 This is a schematic diagram illustrating the target domain identity loss of the present invention; Figure 5 This is a schematic diagram of the source domain signal in the embodiment; Figure 6 Diagnose the confusion matrix for real target domain data; Figure 7 Visualize the features of the real target domain. Detailed Implementation
[0016] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0017] Example 1 To address the aforementioned problems, this invention proposes a training method for a bearing lubrication fault data generation model. The method first collects a real source domain experimental dataset covering all lubrication states under controlled laboratory conditions. Then, a conditional recurrent generative adversarial network (GAN) incorporating multiple advanced constraints is used to fuse the "fault content" of the source domain data with the "environmental style" of normal data from the target domain (industrial field), generating high-fidelity pseudo-fault samples with the same distribution as the target domain. Finally, a decoupled two-stage network with a denoising autoencoder learns essential features insensitive to signal morphological changes and performs accurate classification. (Refer to...) Figure 1 Specifically, it includes the following steps:
[0018] S1: Data Acquisition and Domain Definition.
[0019] S11: Source Domain Data Acquisition: To facilitate subsequent maintenance work, industrial sites typically purchase rotating machinery in batches with bearings of the same or similar models. Therefore, in the laboratory testing platform, identical or similar rolling bearing entities are introduced, and precision control experiments are conducted under different lubrication conditions (e.g., normal, insufficient oil, or deteriorated oil) while controlling the rotational speed, forming a source domain experimental dataset containing multiple lubrication condition categories (labels).
[0020] S12: Target Domain Data Acquisition: In industrial sites requiring lubrication fault monitoring, only relevant signals from the target equipment operating under normal lubrication conditions are collected to form a target domain field dataset. This dataset includes unique background noise and transmission path effects, representing the "environmental style" of the target domain.
[0021] S2: High-fidelity fake sample generation.
[0022] A bearing lubrication failure data generation model is constructed to achieve style transfer from the source domain to the target domain. The specific structure of the model is as follows: Figure 2 As shown, where A This represents a source domain experimental dataset containing multiple lubrication state categories. B This represents a target domain field dataset that includes unique background noise and the influence of transmission paths. As the first discriminator, G is the second discriminator, G is the first generator, F is the second generator, and G(A F( represents the pseudo-target domain sample generated by the first generator G from the real source domain sample, and F( represents the pseudo-target domain sample generated by the first generator G). B () is the target domain real sample passed through the second generator F The generated pseudo-source domain sample.
[0023] Unlike traditional CycleGAN, to ensure that the core fault features are not distorted during the transfer process, the proposed Conditional Recurrent Adversarial Generative Network (CRGAN) employs a joint constraint system including standard adversarial loss, cycle consistency loss, and identity loss, as well as innovatively introduced feature consistency loss and style invariant loss. Feature consistency loss mandates that the reconstructed signal and the original signal remain consistent within the deep feature space extracted by the discriminator, achieving deep content preservation. Style invariant classification loss forces the discriminator to have the ability to reverse-engineer the original source domain fault category from a pseudo-sample that has been converted to the target domain style.
[0024] exist Figure 2 In China, the use of countermeasures against losses Refers to the use of circular consistency loss. Refers to, loss of identity Refers to the additional feature consistency loss used. Referencing, style invariance classification loss usage Referential. The specific calculation formula is as follows:
[0025] Combat losses: ; ; in, For real samples from the source domain, For real samples in the target domain, As the expected value, P data（A） Let represent the distribution of all real samples in the source domain. P data（B） The distribution of all real samples in the target domain. This represents the probability that the second discriminator will judge a real sample in the target domain. Let be the probability that the first discriminator judges a real sample in the source domain. For the second discriminator to test real samples in the source domain Through the first generator G Generated pseudo-target domain samples The probability of judgment. For the first discriminator to target real samples Through the second generator F Generated pseudo-source domain samples The probability of judgment.
[0026] Cyclic consistency loss: ; in, For real samples in the source domain Passing through the first generator in sequence G Second generator F The obtained source domain reconstruction samples, For the target domain real samples Passing through the second generator in sequence F and the first generator G The reconstructed sample obtained.
[0027] Reference Figure 3 and Figure 4 Loss of identity: ; in, For real samples in the source domain Through the first generator G The generated pseudo-target domain samples, For the target domain real samples Through the second generator F The generated pseudo-source domain sample.
[0028] Feature consistency loss: ; in, It is a discriminator Extracted source domain samples The deep features; It is a discriminator Extracted reconstructed samples The deep features; It is a discriminator Extracted target domain samples The deep features; It is a discriminator Extracted reconstructed samples The deep features of.
[0029] Style invariant classification loss: ; in, represent The true category label. Based on the above, the calculated losses are weighted and summed to obtain the final total loss:
[0030] ; in, , , , These are weight parameters.
[0031] Based on the total loss, the bearing lubrication fault data generation model is trained, and the internal weights of the generator and discriminator are updated through backpropagation. During training, all lubrication fault signals from the source domain experimental dataset and all normal samples from the target domain experimental dataset are selected as inputs. After training, the trained network is used to perform style transfer on the lubrication fault signals from the source domain experimental dataset A to generate a pseudo dataset covering all lubrication states and with the same distribution as the target domain B. .
[0032] S3: Robust feature extraction.
[0033] To further extract the essential features of lubrication faults, a denoising autoencoder (DAE) is constructed and trained using the aforementioned target domain pseudo dataset. During training, structural data augmentation (such as random time shifts and amplitude variations) is applied to the input data, forcing the DAE to learn more essential structural features that are insensitive to minute changes in signal morphology.
[0034] Specifically, it includes the following steps: S31: Input the target domain pseudo dataset generated in step S2 The dataset is divided into training and testing sets. Gaussian noise and other structural data augmentation methods (such as random time shifts and amplitude variations) are applied to all data in the training set to generate... .
[0035] S32: Encoder computes hidden layer representation The decoder outputs reconstructed data. ,in and These are the encoder and decoder parameters, respectively.
[0036] S33: Using the mean squared error (MSE) or another similar function as the loss function, minimize the reconstruction error, forcing the model to ignore noise and restore the original data distribution. The mean squared error can be expressed as... .
[0037] S34: Optimize parameters using backpropagation algorithm Complete model training and output optimal parameters. .
[0038] S4: Classification diagnosis.
[0039] After the DAE is trained, all parameters of its encoder are frozen, treating it as a fixed, robust feature extractor. Subsequently, a lightweight classifier (such as a convolutional neural network CNN) is constructed and trained using the hidden layer representations extracted from the frozen DAE encoder. The input data for training is the target domain pseudo-dataset generated in step two. After the overall training is completed, the diagnostic model consisting of "frozen DAE encoder + trained CNN classifier" will be deployed and applied. The current input signal is the new test signal in the industrial field, and the output result is the current lubrication status of the equipment.
[0040] Example 2 This project utilizes a self-designed lubrication fault monitoring test bench to collect vibration data under three operating conditions: dry friction, lubrication deterioration, and normal lubrication. The vibration data sampling frequency is 204.8kHz, and 1000 sets of experimental data were collected under each condition. These experimental data are referred to as source domain data in this invention. A schematic diagram of the signal is shown below. Figure 5 As shown.
[0041] Vibration data of the field test motor under dry friction, lubrication degradation, and normal lubrication conditions are collected. This vibration data is referred to as target domain data in this invention. Specifically, to verify the feasibility of the current method, the field test motor was modified to perform bearing vibration tests under different lubrication conditions. However, most industrial field motors only have the capability to collect vibration signals under normal lubrication conditions.
[0042] A conditional recurrent adversarial generative network (CRG) is trained using all source domain data and normal target domain data. A high-fidelity pseudo dataset of the target domain is then generated using the trained CRG and all source domain data.
[0043] A diagnostic model was trained using a high-fidelity target domain pseudo-dataset, and then used to diagnose real target domain data. The specific diagnostic classification results are as follows: The accuracy rate for identifying normal samples was 99.95%, the accuracy rate for identifying deteriorated samples was 97.97%, and the accuracy rate for identifying dry friction samples was 100%.
[0044] The relevant confusion matrix and feature distribution are shown below. Figure 6 and Figure 7 .
[0045] To evaluate the model performance, Figure 6 This demonstrates the detailed confusion matrix of the current diagnostic model on real target domain data. Furthermore, Figure 7 Using t-SNE visualization technology, the two-dimensional spatial distribution of features extracted from the current diagnostic model of the real target domain data is presented intuitively.
[0046] This invention fundamentally improves the fidelity of generated data: It adopts an "experiment-to-field" paradigm, using real experimental data as the source, thus avoiding the source distortion problem caused by oversimplification of physical simulation models. Compared to the unconditional network used in patent CN117932413A, this invention introduces multiple constraints, such as style-invariant classification loss, to ensure high fidelity of key fault features during style transfer, resulting in higher quality generated pseudo-samples.
[0047] This invention enhances the robustness and accuracy of diagnostic models: through a unique two-stage feature learning architecture, utilizing DAE and structural data augmentation, the model is able to learn essential features that are insensitive to morphological changes, effectively resisting minute signal fluctuations and noise interference in real industrial environments, and significantly improving diagnostic accuracy.
[0048] The technical solution of this invention is more versatile and practical: This invention eliminates the dependence on complex and time-consuming physical modeling of specific equipment, and can be implemented by collecting easily accessible laboratory data and normal field data, which greatly reduces the technical threshold and application cost, and makes it easier to promote in a variety of rotating machinery.
[0049] This invention effectively solves the pain points of real industrial scenarios: addressing the fundamental problem of difficulty in obtaining and labeling lubrication failure data in industrial sites, this invention provides a complete and reliable technical solution that can build a high-performance intelligent diagnostic model using only normal field data.
[0050] Example 3 Based on the same concept, the present invention also provides a training device for generating a bearing lubrication fault data model, including an acquisition module, a generation module, a construction module and a training module.
[0051] The acquisition module is used to acquire real samples of the target domain of bearings in industrial sites under normal lubrication conditions, as well as multiple real samples of the source domain of the same bearings in the laboratory under different lubrication conditions.
[0052] The generation module is used to generate pseudo-target domain samples from real source domain samples using a first generator, and then generate source domain reconstructed samples from pseudo-target domain samples using a second generator; real target domain samples are used to generate pseudo-source domain samples from the second generator, and then generate target domain reconstructed samples from pseudo-source domain samples using the first generator.
[0053] The construction module is used to construct cycle consistency loss based on source domain real samples, source domain reconstructed samples, target domain real samples, and target domain reconstructed samples; to construct adversarial loss and identity loss based on source domain real samples, pseudo target domain samples, target domain real samples, and pseudo source domain samples; to construct feature consistency loss based on the features of source domain real samples, source domain reconstructed samples, target domain real samples, and target domain reconstructed samples extracted by the first discriminator and the second discriminator; and to construct style invariant classification loss based on pseudo target domain samples and their corresponding real labels.
[0054] The training module uses the sum of adversarial loss, cycle consistency loss, identity loss, feature consistency loss, and style invariant classification loss as the total loss, and trains the generative model based on the total loss to obtain the trained generative model.
[0055] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.
[0056] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
Claims
1. A method of training a bearing lubrication failure data generation model, characterized by, The generation model comprises a first generator, a first discriminator, a second generator and a second discriminator, and the training method comprises the following steps: Collect target domain real samples of bearings in an industrial field under normal lubrication state and a plurality of source domain real samples of the same bearings under different lubrication states in a laboratory; The source domain real samples generate pseudo-target domain samples through the first generator, and the pseudo-target domain samples generate source domain reconstructed samples through the second generator; the target domain real samples generate pseudo-source domain samples through the second generator, and the pseudo-source domain samples generate target domain reconstructed samples through the first generator; Cycle consistency loss is constructed based on the source domain real samples, the source domain reconstructed samples, the target domain real samples and the target domain reconstructed samples, and adversarial loss and identity loss are constructed based on the source domain real samples, the pseudo-target domain samples, the target domain real samples and the pseudo-source domain samples; feature consistency loss is constructed based on features of the source domain real samples, the source domain reconstructed samples, the target domain real samples and the target domain reconstructed samples extracted by the first discriminator and the second discriminator, and style invariance classification loss is constructed based on the pseudo-target domain samples and corresponding real labels; The sum of the adversarial loss, the cycle consistency loss, the identity loss, the feature consistency loss and the style invariance classification loss is used as total loss, and the generation model is trained based on the total loss to obtain a trained generation model.
2. The method of claim 1, wherein, The feature consistency loss is specifically as follows: ； wherein, L Feat is a feature consistency loss, G is a first generator, F is a second generator, is a first discriminator, is a second discriminator, is an expected value, is a source domain real sample, is a target domain real sample, P data（A） is a distribution of all real samples in the source domain, P data（B） is a distribution of all real samples in the target domain, is a deep feature of the extracted source domain sample is a deep feature of the extracted reconstructed sample is a deep feature of the extracted target domain sample is a deep feature of the extracted reconstructed sample is a deep feature of the extracted reconstructed sample is a deep feature of the extracted reconstructed sample A is a source domain experimental dataset containing multiple lubrication condition categories, B is a target domain field dataset containing unique background noise and transmission path effects in the field.
3. The training method for a bearing lubrication fault data generation model as described in claim 2, characterized in that, The style invariance classification loss is specifically as follows: ； In the formula, L style is a style-invariant classification loss, is a true class label, is to and generated by the first generator G pseudo target domain samples predicted class probability after classification.
4. The training method for a bearing lubrication fault data generation model as described in claim 3, characterized in that, The total loss is specifically as follows: ； wherein, is a loss against, is a cycle consistency loss, is an identity loss, , , and are weight parameters.
5. The method of claim 1, wherein, Further comprising: The plurality of source domain real samples are input into the trained generation model to obtain a target domain pseudo data set; The target domain pseudo data set is input into the autoencoder after data enhancement to train the autoencoder to obtain a trained autoencoder; The samples in the target domain pseudo data set are input into the trained autoencoder to obtain a hidden layer representation, and the hidden layer representation is input into a classifier to train the classifier to obtain a trained classifier; The to-be-measured signal of the industrial field is input into the trained autoencoder and the classifier to obtain a corresponding lubrication state.
6. The method of training a bearing lubrication failure data generation model of claim 5, wherein, The target domain pseudo data set is input into the autoencoder after data enhancement to train the autoencoder, comprising the following steps: Gaussian noise is applied to the target domain pseudo data set; The target domain pseudo data set to which the Gaussian noise is applied is input into the autoencoder, wherein the encoder outputs a hidden layer representation, and the decoder outputs reconstructed data according to the hidden layer representation; Mean square error is constructed through the target domain pseudo data set and the reconstructed data, and the autoencoder is trained by minimizing the mean square error.
7. A training apparatus of a bearing lubrication failure data generation model, characterized by, The generation model comprises a first generator, a first discriminator, a second generator and a second discriminator, and the training device comprises: A collection module for collecting target domain real samples of bearings in an industrial field under normal lubrication state and a plurality of source domain real samples of the same bearings under different lubrication states in a laboratory; A generation module for generating pseudo-target domain samples from the source domain real samples through the first generator, generating source domain reconstructed samples from the pseudo-target domain samples through the second generator; generating pseudo-source domain samples from the target domain real samples through the second generator, and generating target domain reconstructed samples from the pseudo-source domain samples through the first generator; The construction module is configured to construct a cycle consistency loss based on the source domain real sample, the source domain reconstructed sample, the target domain real sample and the target domain reconstructed sample, construct an adversarial loss and an identity loss based on the source domain real sample, the pseudo target domain sample, the target domain real sample and the pseudo source domain sample, construct a feature consistency loss based on features of the source domain real sample, the source domain reconstructed sample, the target domain real sample and the target domain reconstructed sample extracted by the first discriminator and the second discriminator, and construct a style invariance classification loss based on the pseudo target domain sample and a corresponding real label; The training module is configured to use a sum of the adversarial loss, the cycle consistency loss, the identity loss, the feature consistency loss and the style invariance classification loss as a total loss, train the generative model based on the total loss, and obtain the trained generative model.