Intelligent mechanical fault diagnosis method based on relationship transfer domain generalization network (rtdgn)
The RTDGN method addresses the limitations of existing mechanical fault diagnosis techniques by employing a two-stage training approach within the RTDGN framework, leveraging CNNs and adversarial networks to enhance diagnostic accuracy and robustness across diverse mechanical fault conditions.
Patent Information
- Application Number
- GB2023003406
- Authority / Receiving Office
- GB · GB
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2022-09-01
- Filing Date
- 2023-03-08
- Publication Date
- 2024-11-27
- Estimated Expiration
- 2043-03-08
AI Technical Summary
Current mechanical fault diagnosis methods based on domain adaptation techniques fail when prior assumptions about data availability in target domains are not met, leading to inconsistent data distribution and reduced robustness and generalization ability in practical engineering applications.
An intelligent mechanical fault diagnosis method using a Relationship Transfer Domain Generalization Network (RTDGN) is developed, which includes two stages of training: task-irrelevant domain adaptation (TIDA) and task-relevant domain generalization (TRDG), utilizing a Convolutional Neural Network (CNN) with adversarial networks and a relationship transfer framework to address data distribution discrepancies and enhance diagnostic accuracy.
The RTDGN achieves high generalization ability and robustness in diagnosing mechanical faults across varying conditions, even with incomplete data, by effectively transferring distributional differences from labeled auxiliary domains to unseen target domains, maximizing data utilization and reducing domain discrepancies.
Smart Images

Figure 00000001_0000 
Figure 00000001_0001 
Figure 00000002_0000
Abstract
Description
TECHNICAL FIELD
[0001] The present disclosure belongs to the technical field of mechanical fault diagnosis, and relates to an intelligent mechanical fault diagnosis method based on a relationship transfer domain generalization network (RTDGN). BACKGROUND
[0002] Gears, bearings and other key mechanical parts are widely used in aerospace, rail transit, petrochemical engineering and other important engineering fields. Once such mechanical part has a fault, it will cause serious economic losses and irreversible casualties. At present, in view of the inconsistent distribution of training data and test data, a large number of transfer diagnosis methods based on domain adaption techniques have been widely used in the field of mechanical diagnosis. However, the success of these methods depends entirely on the priori assumption that data of target domains can be obtained during model training. However, this priori assumption often fails in most practical engineering applications.
[0003] Therefore, there is an urgent need for a novel intelligent mechanical diagnosis method to address the above problems. SUMMARY
[0004] In view of this, an objective of the present disclosure is to provide an intelligent mechanical fault diagnosis method based on an RTDGN, so that a rotating mechanical part can obtain a diagnostic result representing high generalization ability and strong robustness in case of inconsistent data distribution and unseen data of target domains.
[0005] To achieve the above objective, the present disclosure provides the following technical solutions:
[0006] The present disclosure provides an intelligent mechanical fault diagnosis method based on an RTDGN, the method including the following steps:
[0007] step SI: acquiring original vibration signals from mechanical equipment under different working conditions, and then dividing a sample set based on the acquired original vibration signals so as to build a subsequent relationship transfer diagnosis task;
[0008] step S2: building a relationship transfer framework according to a relationship transfer diagnosis task to be implemented;
[0009] step S3: building an RTDGN based on the relationship transfer framework built in step S2 and a Convolutional Neural Network (CNN);
[0010] step S4: training and optimizing the RTDGN in two stages: 1, in a stage of task-irrelevant domain adaptation (TIDA), inputting all task-irrelevant auxiliary data into the RTDGN, and then conducting iterative training on weight parameters of a built adversarial network, where the weight parameters of the adversarial network include weight parameters ®FE, ®FC and ®Dt of a feature extractor, a fault classifier and a domain discriminator; and 2, in a stage of task-relevant domain generalization (TRDG), first inheriting the weight parameters of the feature extractor in the stage of TIDA, and then inputting data of source domains from task-relevant data into the RTDGN for fine-tuning training; and
[0011] step S5: after multiple times of iterative training, observing whether a loss curve tends to converge smoothly, finishing training the RTDGN if the curve converges, and conducting, by the trained RTDGN, cross-working-condition transfer fault diagnosis in a relationship transfer task of a rotating mechanical part.
[0012] Said building a relationship transfer framework in step S2 specifically includes: firstly, defining that task-irrelevant auxiliary data include a plurality of auxiliary domains, where D = ID 1® =R(X‘ Y'1® o 1C Y' 1 ' 1=1 , 2 denotes a quantity of the auxiliary domains, and denote an zth sample and a corresponding label vector of a z / th auxiliary domain, respectively, and n D q denotes a sample size of the z / th auxiliary domain; and defining that the task-relevant data R include source domains and target domains, where anc[ t tv rdenote a labeled source domain and an unseen unlabeled target domain, X' y! respectively, s denote an zth sample and a corresponding label vector of a source domain, respectively, 5 denotes a sample size of a source domain, T denotes an zth sample and a corresponding label vector of a target domain, and T denotes a sample size of the target domain; and
[0013] defining a relationship between any two task-irrelevant auxiliary domains as where [1,2,---,2] , and '* / ; denotes a relationship mapping AP between the task-irrelevant auxiliary data, ij denotes a distributional difference between an zth •r j j -i- j • j , AP = 0 D / D7 . auxiliary domain and a / th auxiliary domain, and when ,J , 1 • \ and similarly, defining a relationship between a source domain and a target domain in the task-relevant data as ^(ds’^st), where denotes a relationship mapping between the task-relevant data, AP and 57 denotes a distributional difference between a source domain and an unlabeled target domain; according to similarity ~ in the distributional difference of the task-irrelevant auxiliary data and the distributional difference of the task-relevant data, transferring a distribution discrepancy relationship of the task-irrelevant auxiliary data into the task-relevant data via relationship transfer so as to indirectly measure and eliminate a distributional difference between an unseen target domain and a source domain.
[0014] Further, said building the RTDGN in step S3 is performed in the two stages of TIDA and TRDG, and the CNN is used as a backbone network to extract features in both stages; in the stage of TIDA, a general feature extractor characterized by domain confusion and classification capability is built based on the task-irrelevant auxiliary data; and then distributional relationship knowledge learned by the feature extractor in the stage of TIDAis transferred to the stage of TRDG through the relationship transfer framework to realize domain confusion; and
[0015] in the backbone network, a batch normalization (BN) layer is used to accelerate convergence of a neural network and avoid variance drift; in addition, a Dropout layer and global average pooling (GAP) are used to prevent network overfitting; a Rectified Linear Unit (ReLU) is used as an activation function for a hidden layer, and Softmax and Sigmoid are used as activation functions for an output layer.
[0016] Further, step S3 further includes building the adversarial network with a plurality of domain discriminators in the stage of TIDA, where the adversarial network includes a feature extractor , a fault classifier gfc(^fc) anc[ binary domain discriminators k e [I’2’’ ’,, K ~CQ~ Q(Q -1) / 2 , Q jenotes a combinatorial number obtained by taking two of 2 elements at a time, and , ®FC and n,J denote weight parameters of the feature extractor, the fault classifier and the domain discriminator, respectively.
[0017] Further, step S3 further includes building a domain generalization diagnosis network in the stage of TRDG, where the domain generalization diagnosis network includes a feature extractor gfe^fe) and a residual fusion classifier ^(^c); the feature extractor in the stage of TRDG have a structure and weight parameters the same as a structure and weight parameters of the feature extractor in the stage of TIDA; the residual fusion classifier is a fault classifier that fuses multiple features through residual connections.
[0018] Further, said conducting iterative training on weight parameters of a built adversarial network in the stage of TIDA in step S4 specifically includes iteratively updating the weight parameters , ®FC and ®Dt of the adversarial network in the stage of TIDA by an Adam optimizer, a domain loss L°, a classification loss Lc and an inverse entropy loss Lie as follows: FC FC S0
[0019]
[0020]
[0021] ^FE EE £0 d0m d0„„ d0m \ rii rc rii
[0022] where s° denotes a learning rate in the stage of TIDA, and denote weighing parameters, denotes a weight parameter of each domain classifier, and D ' °1’ °2 ' !,: h and in order to directly update gradient parameters, a Gradient reversal layer (GRL) is introduced into the adversarial network.
[0023] Further, the domain loss L° is defined as follows:
[0024] = E LDt(GDJGPE(X),V) i=1 nt +rij XeDj ud,
[0025] where d denotes a one-hot encoding vector of a domain label, and Dt denotes a Ath binary cross-entropy loss,
[0026] the classification loss Lc is defined as follows: Lc^FEfic')^- E 4,(^c(^(X),Y))
[0027] q=\ nq
[0028] where Y denotes a label vector corresponding to an auxiliary domain sample, and denotes a multi-classification cross-entropy loss of a c / th auxiliary domain;
[0029] the inverse entropy loss Lie is defined as follows:
[0030] wk = expQJ pk =^- E LDt(GDt(GPE(X)^
[0031] where , and ni +nj XeD^D, (iFf) n n
[0032] Further, in step S4, the weight parameters FE and RC in the stage of TRDG are updated as follows:
[0035] where S1 and Sz denote a learning rate, and si<sz ; anc[ ! denotes a multiclassification cross-entropy loss of a source domain.
[0036] Further, the multi-classification cross-entropy loss ^Cs of a source domain is defined as 4C(^AC) = — y lcgrc(gfe(xs\ys) r C \ rt “ r\G / / i r\G X rt j ~ o _ _ follows: 5 XseDs , where s denotes a label vector corresponding to a source domain sample.
[0037] The present disclosure has the following beneficial effects:
[0038] (1) According to the present disclosure, a relationship transfer framework is built; through the similarity of distributional differences, the framework can transfer the distributional difference information obtained on one rotating mechanical part with a complete label data set (task-irrelevant auxiliary data) to another rotating mechanical part with incomplete data (task-relevant data). It can not only maximize the use of data resources, but also help rotating mechanical parts with incomplete data to measure and reduce the distributional difference between a source domain and a target domain indirectly even when data of target domains are not available. Finally, the built transfer diagnosis model can obtain a diagnostic result representing high generalization ability and strong robustness in case of inconsistent data distribution and unseen data of target domains.
[0039] (2) An intelligent machinery diagnosis method based on an RTDGN is designed based on the built relationship transfer framework. The RTDGN is trained and optimized in two stages: 1, a stage of TIDA for task-irrelevant auxiliary data; and 2, a stage of TRDG for task-relevant data. In the stage of TIDA, an adversarial network with a plurality of domain discriminators is built to enhance the domain confusion ability. In addition, an inverse entropy loss is designed to enable the network to obtain a more general domain confusion capability. In the stage of TRDG, a residual fusion classifier is designed to effectively restrain fault classifier from falling into overfitting, and at the same time improve the generalization ability of the fault classifier.
[0040] Other advantages, objectives and features of the present invention will be illustrated in the subsequent description in some degree, and will be apparent to those skilled in the art in some degree based on study on the following description, or those skilled in the art may obtain teachings by practicing the present invention. The objectives and other advantages of the present disclosure can be implemented and obtained by the description below. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] In order to make the objectives, technical solutions, and advantages of the present disclosure clearer, the present disclosure will be preferably described in detail below with reference to the accompanying drawings.
[0042] FIG. lisa schematic diagram of a relationship transfer task according to an embodiment of the present disclosure;
[0043] FIG. 2 is a structural block diagram of an RTDGN;
[0044] FIG. 3 is a schematic diagram of a Drivetrain Diagnostics Simulator (DDS); and
[0045] FIG. 4 is a schematic diagram of a Rotor Dynamics Simulator (DDS). DETAILED DESCRIPTION OF THE EMBODIMENTS
[0046] The implementations of the present disclosure are described below through specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. The present disclosure can also be implemented or applied through other different specific implementations. Based on different viewpoints and applications, various modifications or amendments can be made to various details of this specification without departing from the spirit of the present disclosure. It should be noted that the diagrams provided in the following embodiments merely illustrate the basic conception of the present disclosure only schematically, and the following embodiments or features in the embodiments may be combined in a non-conflicting manner.
[0047] The accompanying drawings are schematic diagrams rather than physical diagrams, which are only for illustrative description and should not be construed as a limitation to the present disclosure. In order to better describe the embodiments of the present disclosure, some components may be omitted, enlarged or reduced in the accompanying drawings, and thus do not represent true sizes of physical products. Those skilled in the art should understand that some well-known structures and descriptions thereof may be omitted in the accompanying drawings.
[0048] The same or similar labels in the accompanying drawings of the embodiment of the present disclosure correspond to the same or similar components. In the description of the present disclosure, it should be understood that orientation or position relationships indicated by terms such as "upper", "lower", "left", "right", "front" and "rear" are orientation or position relationships as shown in the drawings. These terms are merely intended to facilitate and simplify the description of the present disclosure, rather than to indicate or imply that the mentioned device or components must have a specific orientation or must be constructed and operated in a specific orientation. Therefore, the terms describing the positional relationship in the accompanying drawings are only used for illustrative description, and should not be understood as a limitation to the present disclosure. A person of ordinary skill in the art will understand specific meanings of such terms based on a specific situation.
[0049] Referring to FIG. 1 to FIG. 4, the embodiment of the present disclosure provides an intelligent mechanical fault diagnosis method based on an RTDGN of a bearing and a gear, the method including the following steps:
[0050] Step 1: Install an acceleration sensor to acquire original vibration signals from mechanical equipment under different working conditions. Divide a sample set based on the acquired original vibration signals so as to build a subsequent relationship transfer diagnosis task.
[0051] Step 2: Build a relationship transfer framework (see FIG. 1) according to a relationship transfer diagnosis task to be implemented.
[0052] Build a relationship transfer framework: firstly, define that task-irrelevant auxiliary data D D =[D ]e = n(X' Y)}”5F o 1 include a plurality of auxiliary domains, where 1 ’ 1 !=1 ’ 1, denotes Xz Yz a quantity of the auxiliary domains, and denote an zth sample and a corresponding label 77 vector of a z / th auxiliary domain, respectively, and ? denotes a sample size of the z / th auxiliary domain. Define that the task-relevant data include source domains and target domains, where , DS-{(XS,YS)L1 anj Dr-{(Xr)}j=1 denote a labeled source domain and an Y* unseen unlabeled target domain, respectively, s ’ s denote an zth sample and a corresponding n label vector of a source domain, respectively, 5 denotes a sample size of a source domain, T n denotes an zth sample and a corresponding label vector of an unlabeled target domain, and T denotes a sample size of the unlabeled target domain.
[0053] Define, in a relationship transfer framework, a relationship between any two task- , .,. , D, ,;AP) , i / ell ? ... z)| (h 6) irrelevant auxiliary domains as 1 • 'J , where , and 7 ; AP denotes a relationship mapping, iJ denotes a distributional difference between an zth auxiliary domain and a / th auxiliary domain, and when “ 0, b . Similarly, define a relationship between a source domain and a target domain in the task-relevant data as -^r^s^st) However, as the target domain sample is unknown, the relationship between the source domain and the target domain cannot be measured and eliminated. However, it is possible to indirectly measure and eliminate the distributional difference between a target domain and a source domain by transferring a distribution discrepancy relationship of the task-irrelevant auxiliary data into the task-relevant data according to similarity (^-^^ ) in the distributional difference of the task-irrelevant auxiliary data and the distributional difference of the task-relevant data via relationship transfer.
[0054] For example, FIG. 1 shows a gear-to-bearing relationship transfer task. Assuming that the prior knowledge of the vibration signal and its label under one of the three bearing working conditions (source domain) of a bearing is known, it is necessary to diagnose the fault types under other working conditions (target domain). In the domain generalization task, the data of target domains are not available in training, so the distributional difference caused by different loads will greatly reduce the robustness and generalization ability of the diagnostic model trained using the source domain. Through the relationship transfer framework proposed in the present disclosure, the distributional difference caused by the load change can be indirectly obtained from the complete and labeled task-irrelevant auxiliary data, so as to realize domain confusion of taskrelevant data when the target domain is unseen.
[0055] S3: Build an RTDGN based on the relationship transfer framework built in S2 and a Convolutional Neural Network (CNN), as shown in FIG. 2.
[0056] RTDGN includes two stages of TIDA and TRDG, as shown in FIG. 2. In both stages, a CNN is used as a backbone network to extract features. In the stage of TIDA, a general feature extractor characterized by domain confusion and classification capability is built based on task-irrelevant auxiliary data. The feature extractor is then be transferred to the stage of TRDG through the relationship transfer framework proposed according to the present disclosure. In the backbone network, a batch normalization (BN) layer is used to accelerate convergence of a neural network and avoid variance drift; in addition, a Dropout layer and global average pooling (GAP) are used to prevent network overfitting. A Rectified Linear Unit (ReLU) is used as an activation function for a hidden layer, and Softmax and Sigmoid are used as activation functions for an output layer. The two stages of TIDA and TRDG will be described in detail below.
[0057] l)Stage of TIDA
[0058] In the stage of TIDA, an adversarial network with a plurality of domain discriminators is built to reduce the distributional difference between Q auxiliary domains and guarantee the robustness in relationship transfer. As shown in FIG. 2, the adversarial network includes a feature extractor , a fault classifier gfc(^fc) and K binary domain discriminators where , and K , and ®FF, ®FC and <>F denote weight parameters of the feature extractor, the fault classifier and the domain discriminator, respectively. Through adversarial training between K domain discriminators and feature extractors, domain confusion between any two auxiliary domains can be achieved. By the following three loss functions, domain-invariant but discriminative high-dimensional general features can be obtained.
[0059] (1) Domain loss
[0060] A domain loss can be used to remove the a distributional difference between two arbitrary auxiliary domains. The feature extractor can use the domain loss to extract common domaininvariant features, as defined below: = Z LDt(GDt{GPE(X),dy)
[0061] Q)
[0062] Where d denotes a one-hot encoding vector of a domain label, denotes a Ath binary cross-entropy loss, denotes a weight parameter of each domain classifier, and
[0063] (2) Classification loss Lc
[0064] The feature extractor can only achieve domain confusion through domain loss, but loses the classification capability. It is desirable that the feature extractor can obtain domain-invariant but discriminative features. Therefore, through supervisory training, classification loss is used to minimize the classification risk. Lc(ePE,ec)=Y—Y LCq(GPC(GPE(X)y})
[0065] n“Xl> (2)
[0066] Where Lc- denotes a multi-classification cross-entropy loss of a c / th auxiliary domain.
[0067] (3) Inverse entropy loss Lie
[0068] Assuming that all domain discriminators have gained excellent discriminant ability, the loss output of each domain discriminator can be used to measure the distribution similarity between the two auxiliary domains. For example, a greater Ath binary cross-entropy loss indicates that the feature extractor pays more attention to the distribution alignment between the two auxiliary domains corresponding to the Ath domain discriminator. Generally, a feature extractor is expected to uniformly realize the domain confusion between all auxiliary domains, because paying too much attention to the distribution alignment between two domains will cause domain-invariant features learned by the feature extractor to be over-fitted. Therefore, in order to effectively implement the relationship transfer framework proposed in the present disclosure, an inverse entropy loss is designed to help the feature extractor obtain more universal high-dimensional general features. In order to facilitate the derivation of the inverse entropy loss, the domain loss of the Ath domain discriminator is first defined as follows: Pk =
[0069] Z LDt(GDt(GPE(X),d)) (3)
[0070] A weight corresponding to Pk may be calculated as follows: Wk=
[0071] S,=iexPte) 0) ix=i
[0072] Where k=l .Finally, the inverse entropy loss is expressed as follows:
[0073] (5)
[0074] As shown in FIG. 2, the inverse entropy loss is only used for updating the weight parameters of the feature extractor.
[0075] Based on the Adam optimizer and the above three losses, the weight parameters ®FF, ®FC and Dt of the adversarial network in the stage of TIDA can be iteratively updated as follows: n > a _F UFC ^UFC fc0
[0076] (6)
[0077] (7)
[0078] ^FE @FE £0 ( dLc dLrP dLn < d0PE d0PE (8)
[0079] Where s° denotes a learning rate in the stage of TIDA, and M and denote weighing parameters. In order to directly update gradient parameters, a Gradient reversal layer (GRL) is introduced into the adversarial network.
[0080] 2) Stage of TRDG
[0081] As shown in FIG. 2, in the stage of TRDG, a domain generalization diagnosis network including a feature extractor gfb@fb) and a residual fusion classifier grc^rc) is built. The feature extractor in the stage of TRDG have a structure and weight parameters the same as a structure and weight parameters of the feature extractor in the stage of TIDA. The features obtained by the feature extractor are more general and transferable than those obtained by the residual fusion classifier. In the meanwhile, it also reflects that the residual fusion classifier is liable to overfit the source domain. Therefore, in order to prevent the classifier from overfitting and obtain a more generalized diagnostic result, a fault classifier that fuses multiple features through residual connections is built.
[0082] In the stage of TRDG, because the target domain is unseen, the distributional difference io between the source domain and the target domain is immeasurable, which reflects that the domain generalization network trained only by the source domain samples is not a network with a feature of transfer diagnosis. Through the relationship transfer framework proposed in the present disclosure, the distributional relationship knowledge learned by the transfer feature extractor in the stage of TIDA can be transferred to the stage of TRDG, thereby realizing domain confusion.
[0083] A fusion classification loss of the residual fusion classifier is defined as follows: LPc^fe^rc>- X LcGrc(Gfe(Xs),Xs)
[0084] ns (9)
[0085] Where ^Cs denotes a multi-classification cross-entropy loss of a source domain.
[0086] In the gradient back propagation of the fusion classification loss, two optimizers with two different learning rates are used to implement fine-tuning operations. The optimizer with a lower learning rate is used to update learning for the feature extractor, and the optimizer with a higher n n learning rate is used by the fusion classifier. The weight parameters ( FE , RC ) are updated as follows: n _F ^FC UFE ^UFE fcl
[0087] (10) n _p ^FC URC URC - ~n
[0088] (11)
[0089] Where a learning rate S1 and a learning rate Sz should meet the following condition: Sj < s2
[0090] S4: Conduct training and optimizing in two stages: 1) in a stage of TIDA, input all task-irrelevant auxiliary data into the RTDGN, and then conduct iterative training on weight parameters of a built network according to formulas (6) to (8). 2) In a stage of TRDG, first inherit the weight parameters of the feature extractor in the stage of TIDA, and then input, according to formulas (10) to (11), data of source domains from task-relevant data into the RTDGN for fine-tuning training.
[0091] Step 5: After multiple times of iterative training, observe whether a loss curve tends to converge smoothly, finish training the network model if the curve converges, and conduct, by the trained RTDGN, cross-working-condition transfer fault diagnosis in a relationship transfer task of a bearing and a gear.
[0092] The above introduce the process steps of using the proposed RTDGN for cross-workingcondition transfer fault diagnosis in a relationship transfer task of a bearing and a gear, and the effectiveness of the method is illustrated below by experiments.
[0093] Verification experiment:
[0094] Original gear vibration signals collected in this experiment are from a DDS, which is shown in FIG. 3. The DDS mainly includes five parts of sensors, a motor, a planetary gearbox, a parallel shaft gearbox and a magnetic powder brake. By loading simulation of the magnetic powder brake, signals under different working conditions can be obtained, which are respectively 0 N.m (Gl), 1.4 N.m (G2) and 2.8 N.m (G3). There are five kinds of health information in the planetary gearbox, including normal condition (NC), surface wear (SW), root crack (RC), chipped tooth (CT), and missing tooth (MT). An acceleration sensor is installed on the planetary gearbox to collect the original vibration signals.
[0095] Original bearing vibration signals collected in this experiment are from an RDS, which is shown in FIG. 4. The RDS includes a servo motor, a coupling, a rotor as well as bearings and bearing supports on both ends of the rotor. The signals under different working conditions can be obtained from rotors of different mass, which are ON (Bl), 28N (B2) and 44N (B3), respectively. There are four kinds of health information about the bearing, including normal condition (NC), inner-race fait (IF), outer-race fault (OF) and ball fault (BF). An acceleration sensor is installed on a planetary bearing support to collect the original vibration signals.
[0096] Through the bearing and gear data sets collected in the above manner, two relationship transfer tasks can be built, each transfer task including six cross-working-condition transfer diagnosis tasks, as shown in Table 1 and Table 2.
[0097] Comparative experiment:
[0098] In order to prove the superiority of the relationship transfer framework and the RTDGN built according to the present disclosure, the relationship transfer framework and the RTDGN are compared with a current typical fault diagnosis network model, and the experimental results are shown in Table 1 and Table 2. As can be seen from the data in Table 1 and Table 2, the relationship transfer framework and the RTDGN built according to the present disclosure have higher transfer diagnosis accuracy and stronger robustness and generalization ability.
[0099] Table 1 Experimental results of the gear-to-bearing relationship transfer task Method Cross-working-condition transfer diagnosis task for a bearing (%) Mea n B1^B2 B1^B3 B2^B1 B2^B1 B3^B1 B3^B2 SVM 33.37 + 1.41 23.53+ 2.21 32.29 + 1.37 25.50+ 1.21 24.65+ 1.19 24.61 + 0.79 27.33 RF 32.04 ± 0.83 23.57± 3.42 53.90 ± 1.32 25.17± 0.63 27.28± 1.45 25.57± 1.02 31.26 CNN 64.89+2.23 45.13 + 0.95 28.37+ 1.71 83.35+2.90 47.11 + 1.12 65.54 + 1.33 55.73 Whitening-net 67.31 + 1.85 47.28 ± 0.97 38.181 1.75 92.471 1.76 38.8H 2.12 69.151 1.21 58.87 DANN 85.721 1.35 65.23+ 0.81 76.861 1.50 81.401 1.54 59.751 1.25 78.991 1.60 74.66 DDC 87.61 + 0.81 72.14+ 2.05 88.0912.37 77.0512.01 63.101 1.50 81.081 1.38 78.18 RTDGN(wZo) 94.32 ± 1.25 72.20 ± 1.43 92.181 0.87 81.091 0.96 80.721 1.88 76.551 1.42 82.84 RTDGN 98.66 ± 0.48 89.00± 1.00 92.351 0.64 94.011 1.12 87.391 1.48 87.451 1.11 91.48
[0100] Table 2 Experimental results of the bearing-to-gear relationship transfer task Cross-working-condition transfer diagnosis task for a gear (%) Method G1^G2 G1^G3 G2^G1 G2^G1 G3^G1 G3^G2 Mean SVM 30.80 ± 0.40 26.57+ 0.30 32.961 0.79 47.481 0.35 27.901 0.30 50.821 0.75 36.09 RF 24.30+ 0.38 22.21+0.80 25.591 0.68 22.831 0.71 23.861 0.63 24.001 0.57 23.80 CNN 72.36+ 1.98 62.311 1.27 79.371 1.16 73.601 2.40 59.341 0.99 69.251 1.30 69.37 Whitening-net 73.93± 1.31 57.77± 1.40 82.421 1.00 79.961 0.71 67.221 1.06 74.561 1.13 72.64 DANN 90.89+ 1.31 79.8H 1.38 95.871 0.79 88.841 0.72 82.001 0.98 88.381 1.21 87.63 DDC 96.35 ± 0.90 80.11 ±0.91 91.401 1.36 89.061 1.51 79.601 0.73 89.521 1.39 87.67 RTDGN(wZo) 85.55± 1.41 74.301 0.50 92.341 1.06 86.201 1.34 74.011 1.23 84.061 1.02 82.74 RTDGN____ 94.271 0.69 82.581 0.56 94.301 i.ii 90.731 1.20 85.221 0.84 93.361 0.71 90.08
[0101] In Table 1 and Table 2, RTDGN (w / o) represents a diagnostic model obtained by a process of removing the inverse entropy loss from an RTDGN. This process falls under the category of an ablation experiment and is used to verify the effectiveness of the inverse entropy loss built according to the present disclosure.
[0102] Finally, it should be noted that the examples above are merely intended to explain, rather than to limit, the technical solutions of the present disclosure. Although the present disclosure is described in detail with reference to the preferred examples, those of ordinary skill in the art should understand that modifications or equivalent substitutions may be made to the technical solutions of the present disclosure without departing from the spirit and scope of the technical solutions of the present disclosure, and these modifications or equivalent substitutions shall fall within the scope of the claims of the present disclosure.
Claims
1. An intelligent mechanical fault diagnosis method based on a relationship transfer domain generalization network (RTDGN), the method comprising the following steps:step SI: acquiring original vibration signals from mechanical equipment under different working conditions, and then dividing a sample set based on the acquired original vibration signals;step S2: building a relationship transfer framework according to a relationship transfer diagnosis task to be implemented;step S3: building an RTDGN based on the relationship transfer framework built in step S2 and a Convolutional Neural Network (CNN);step S4: training and optimizing the RTDGN in two stages: 1, in a stage of task-irrelevant domain adaptation (TIDA), inputting all task-irrelevant auxiliary data into the RTDGN, and then conducting iterative training on weight parameters of a built adversarial network, wherein the weight parameters of the adversarial network comprise weight parameters , ®FC and ®Dt of a feature extractor, a fault classifier and a domain discriminator; and 2, in a stage of task-relevant domain generalization (TRDG), first inheriting the weight parameters of the feature extractor in the stage of TIDA, and then inputting data of source domains from task-relevant data into the RTDGN for fine-tuning training; andstep S5: after multiple times of iterative training, observing whether a loss curve tends to converge smoothly, finishing training the RTDGN if the curve converges, and conducting, by the trained RTDGN, cross-working-condition transfer fault diagnosis in a relationship transfer task of a rotating mechanical part.
2. The intelligent mechanical fault diagnosis method based on an RTDGN according to claim 1, wherein said building a relationship transfer framework in step S2 specifically comprises: firstly, defining that task-irrelevant auxiliary data ®1 comprise a plurality of auxiliary domains, wherein D =[D ]e_ = [{(X! Y )}”l ]e o Y Y1 ’ 1 , denotes a quantity of the auxiliary domains, anddenote an zth sample and a corresponding label vector of a z / th auxiliary domain, respectively, andO7 denotes a sample size of the z / th auxiliary domain; and defining that the task-relevant data R D =FD D 1 Dcomprise source domains and target domains, wherein « l s u s, an(jt T)h=\ denote a labeled source domain and an unseen unlabeled target domain, X‘ Y'respectively, s ’ s denote an zth sample and a corresponding label vector of a source domain,it X?respectively, 5 denotes a sample size of a source domain, T denotes an zth sample and a Tlcorresponding label vector of an unlabeled target domain, and T denotes a sample size of the target domain; anddefining a relationship between any two task-irrelevant auxiliary domains as wherein M e [1,2,---,2] , anc[ 0 / 0 denotes a relationship mapping APbetween the task-irrelevant auxiliary data, iJ denotes a distributional difference between an zth.,. , , .< , . , , AP =0 D / — ®7 j • -1 1auxiliary domain and a / th auxiliary domain, and when ,J , 1 ‘ \ and similarly,defining a relationship between a source domain and a target domain in the task-relevant data as D / -^(D^AP^), () denotes a relationship mapping between the task-relevant data,and 57 denotes a distributional difference between a source domain and an unlabeled target domain; according to similarity ~ in the distributional difference of the task-irrelevant auxiliary data and the distributional difference of the task-relevant data, transferring a distribution discrepancy relationship of the task-irrelevant auxiliary data into the task-relevant data via relationship transfer so as to indirectly measure and eliminate a distributional difference between an unseen target domain and a source domain.
3. The intelligent mechanical fault diagnosis method based on an RTDGN according to claim 2, wherein said building the RTDGN in step S3 is performed in the two stages of TIDA and TRDG, and the CNN is used as a backbone network to extract features in both stages; in the stage of TIDA, a general feature extractor characterized by domain confusion and classification capability is built based on the task-irrelevant auxiliary data; and then distributional relationship knowledge learned by the feature extractor in the stage of TIDA is transferred to the stage of TRDG through the relationship transfer framework to realize domain confusion; andin the backbone network, a batch normalization (BN) layer is used to accelerate convergence of a neural network and avoid variance drift; in addition, a Dropout layer and global average pooling (GAP) are used to prevent network overfitting; a Rectified Linear Unit (ReLU) is used as an activation function for a hidden layer, and Softmax and Sigmoid are used as activation functions for an output layer.
4. The intelligent mechanical fault diagnosis method based on an RTDGN according to claim 3, wherein step S3 further comprises building the adversarial network in the stage of TIDA, wherein the adversarial network comprises a feature extractor , a fault classifier 0000and K binary domain discriminators kK\, K Ce Q(Q 1) / 2, CQ jen0|es acombinatorial number obtained by taking two of elements at a time, and , ®FC and n,J denote weight parameters of the feature extractor, the fault classifier and the domain discriminator, respectively.
5. The intelligent mechanical fault diagnosis method based on an RTDGN according to claim 3, wherein step S3 further comprises building a domain generalization diagnosis network in the stage of TRDG, wherein the domain generalization diagnosis network comprises a feature extractor gfe^fe) and a residual fusion classifier the feature extractor in the stage ofTRDG have a structure and weight parameters the same as a structure and weight parameters of the feature extractor in the stage of TIDA; the residual fusion classifier is a fault classifier that fuses multiple features through residual connections.
6. The intelligent mechanical fault diagnosis method based on an RTDGN according to claim 4, wherein said conducting iterative training on weight parameters of a built adversarial network in the stage of TIDA in step S4 specifically comprises iteratively updating the weight parameters See , F- anc[ &Dt 0£ a(iversariai network in the stage of TIDA by an Adam optimizer, a domain loss , a classification loss Lc and an inverse entropy loss Lif as follows:a <—ft — FUFC ^UFC fc0^FC^FE @FE £0dLc\ riiwherein s° denotes a learning rate in the stage of TIDA, and denote weighing / 2 o = io 0 0 \parameters, '• denotes a weight parameter of each domain classifier, and ‘ " 5 J; andin order to directly update gradient parameters, a Gradient reversal layer (GRL) is introduced into the adversarial network.
7. The intelligent mechanical fault diagnosis method based on an RTDGN according to claim 6, wherein the domain loss L° is defined as follows:LD(GPE fin) = ^-^- £ LDt(GDt(Gpl(X),d))i=1 nt +rij XeDjuDjwherein d denotes a one-hot encoding vector of a domain label, and Dt denotes a Ath binarycross-entropy loss;the classification loss Lc is defined as follows:LC (GPC(GPE(X^q=\ nqwherein Y denotes a label vector corresponding to an auxiliary domain sample, and L' denotes a multi-classification cross-entropy loss of a c / th auxiliary domain;the inverse entropy lossis defined as follows:_ exp(ft)wk -------V exp(p)wherein 1Ew*=1, andPk =^- Z8. The intelligent mechanical fault diagnosis method based on an RTDGN according to claim5, wherein in step S4, the weight parameters FE and RC in the stage of TRDG are updated asfollows:a > a _p ^FC ^FF^^FE riin ,a ^FCURC URC fc2^RCwherein S1 and Sz denote a learning rate, and S1<Szand Cs denotes a multiclassification cross-entropy loss of a source domain.
9. The intelligent mechanical fault diagnosis method based on an RTDGN according to claim8, wherein the multi-classification cross-entropy loss of a source domain is defined asLFc(eFE^Rc)=— X LcGrc(Gfe(Xs\Xs)follows: 5 XseDs , wherein denotes a label vectorcorresponding to a source domain sample.
Citation Information
Patent Citations
Multi-source-domain adaptive intelligent mechanical fault diagnosis method and system
CN112926642A
Gearbox fault diagnosis model training method and gearbox fault diagnosis method
CN114357663A
Rolling bearing fault diagnosis method and system in strong noise environment
CN114459760A
Bearing cross-working-condition fault prediction method based on adversarial transfer learning
CN114722879A
Rolling bearing fault diagnosis method and system for sample imbalance under variable speed working condition
CN114838936A
