Cross-domain fault diagnosis method for rotating machinery based on DACODL framework

Abstract

The present invention discloses a rotating machinery cross-domain fault diagnosis method based on a DACODL framework, comprising: constructing a source domain data set and a target domain data set according to vibration data under different working conditions; performing data enhancement on the target domain data set to obtain an enhanced data set of the target domain data set; constructing a diagnosis model, and training the diagnosis model by generating a virtual data set through the source domain data set, the target domain data, the enhanced data set and cross-domain mixup; iteratively training the diagnosis model based on the above data set using a trusted order cross-domain mixup loss, a dynamic label distribution-aware regularized cross-entropy loss, a joint distribution multi-layer adaptation loss and a domain adversarial loss; obtaining a final diagnosis model, testing the final diagnosis model, and using the tested final diagnosis model for the vibration data of the rotating machinery under the target working condition to obtain a rotating machinery cross-domain fault diagnosis result.

Description

Technical Field

[0001] The present invention belongs to the technical field of rotating machinery fault diagnosis, and in particular relates to a rotating machinery cross-domain fault diagnosis method based on a DACODL framework. Background Art

[0002] Rotating machinery systems play a vital role in modern industry, and their normal operation has a key impact on production efficiency, product quality and safety. However, due to the complexity and variability of rotating equipment, the occurrence of faults is inevitable and often leads to production interruptions, increased costs and waste of resources. Therefore, the development of efficient and accurate mechanical fault diagnosis technology has become one of the research hotspots in the engineering field. Traditional rotating machinery fault diagnosis methods based on signal processing technology can complete the diagnosis task well, but they are too dependent on the rich experience and expertise of engineering and technical personnel, and do not conform to the trend of real-time monitoring and diagnosis of industrial automation. In recent years, with the rapid development of artificial intelligence methods, this technology has shown great potential in the field of mechanical fault diagnosis, which can not only make up for the shortcomings of traditional methods but also achieve relatively good performance. Deep neural networks have significantly improved the technical level of intelligent fault diagnosis and shown significant generalization ability. For data sets with similar or identical distributions of training and test sets, it has excellent performance. However, in real-world scenarios, the data collected by the machine under different working conditions may experience domain shift, that is, the data distribution collected under one working condition (source domain) is significantly different from that collected under another working condition (target domain). The distribution difference between source domain samples and target domain samples will cause the diagnostic performance of the model trained in the source domain to deteriorate for target domain samples. Therefore, it is urgent to develop new deep learning methods to solve the problem of domain shift in cross-working condition fault diagnosis. Summary of the invention

[0003] In order to solve the above technical problems, the present invention proposes a cross-domain fault diagnosis method for rotating machinery based on the DACODL framework to solve the problems existing in the above prior art.

[0004] To achieve the above object, the present invention provides a rotating machinery cross-domain fault diagnosis method based on the DACODL framework, comprising:

[0005] S1. Obtain vibration data of rotating machinery under different working conditions, and construct a labeled source domain data set and an unlabeled target domain data set according to the vibration data under the different working conditions, wherein the label is a fault diagnosis result;

[0006] S2. performing calculation order tracking transformation and other preprocessing on the source domain data set and the target domain data set, and performing data enhancement on the target domain data set to obtain an enhanced data set of the target domain data set;

[0007] S3. Build a diagnostic model and set hyperparameters;

[0008] S4. In each round of training, the preprocessed target domain data is processed by the diagnostic model to obtain high-confidence samples and pseudo-labels, and the high-confidence samples and pseudo-labels are stored in the sample storage of the current round;

[0009] S5. In the current round of training, adaptive sampling is performed from the sample storage of the previous round, and the sampling results are mixed up with the source domain samples of the current round to obtain a virtual data set, wherein the virtual data set includes virtual samples and corresponding labels;

[0010] S6. Training the diagnostic model using the preprocessed source domain dataset, target domain dataset, enhanced dataset and virtual dataset;

[0011] S7. Calculate the total loss of the diagnostic model trained in the current round, and update the model parameters of the diagnostic model according to the total loss of the current round;

[0012] S8. Repeat the above S4-S7 until the training rounds reach the maximum number of times to obtain the final diagnosis model;

[0013] S9. Test the final diagnostic model, and obtain the cross-domain fault diagnosis result of the rotating machinery by analyzing the vibration data of the rotating machinery under the current working condition through the tested final diagnostic model.

[0014] Optionally, in step S2, other preprocessing methods include normalization and data enhancement, and data enhancement methods include Gaussian noise, random cropping, time stretching and amplitude stretching.

[0015] Optionally, in step S3, the diagnostic model includes a feature extractor, a label classifier and a domain discriminator, wherein the feature extractor is connected to the label classifier and the domain discriminator respectively.

[0016] Optionally, in step S4, the process of acquiring the virtual data set includes:

[0017] The preprocessed target domain data and enhanced data set corresponding to the current round are processed by the diagnostic model, and the maximum label probability of each sample in the target domain data and the enhanced data set is obtained through the softmax function. Samples with maximum label probabilities higher than the set threshold are screened to obtain high-confidence samples, and the label corresponding to the maximum label probability is used as a pseudo-label.

[0018] Optionally, in step S5, the process of adaptive sampling includes:

[0019]

[0020] Among them, pc Indicates the sample storage of class c from the N-1th round The probability of being sampled, N c Indicates the sample storage of the N-1th round In , ν is a hyperparameter used to adjust the sampling frequency, and C represents the number of categories.

[0021] Optionally, in step S6, the total loss is:

[0022]

[0023] in, represents the confidence order cross-domain mixup module loss, represents the cross entropy loss with dynamic label distribution aware margin regularization, represents the joint distribution multi-layer adaptation loss, represents the domain adversarial loss, ε adv represents the weight of the domain adversarial loss, μ t represents the time-varying trade-off coefficient, which is t represents the current round number, t m Indicates the maximum number of rounds.

[0024] Optional, cross-entropy loss with dynamic label distribution aware margin regularization for:

[0025]

[0026] in

[0027] Among them, z c =G c (G f (x s )) c represents the c-th output of the label classifier, i.e., the c-class output, ν 2 It is a hyperparameter used to adjust the marginal distance of each category, N c Indicates the sample storage of the N-1th round The number of samples of category c, C represents the number of categories, and K is a preset constant used to control the maximum margin of the category.

[0028] Optional, domain adversarial loss for:

[0029]

[0030] Among them, D s and D t Represent the source domain samples x s and target domain sample xt The distribution it obeys.

[0031] Optionally, jointly distribute multiple layers of adaptation loss for:

[0032]

[0033] Among them, γ 1 and γ 2 The trade-off factor between the global distribution difference and subdomain distribution difference of the same layer features, γ 3 and γ 4 Represents the trade-off factor of the difference in joint distribution between different layers, G pc (·) represents the features obtained from the penultimate layer of the label classifier, and They represent the proportion of class C samples in the source domain and the target domain, respectively, and are calculated by the number of class C samples in the source domain true label and the target domain pseudo label, respectively. and are all equal to 1, and is the weighted sum of category c, Φ u (·) represents the features generated by the u-th feature kernel.

[0034] Optional, confident order cross-domain mixup module loss for:

[0035]

[0036] in, represents the cross-domain mixup loss, ε CDM represents the cross-domain mixup loss weight, L SENTRY represents the selective entropy optimization strategy loss, μ t represents the time-varying trade-off coefficient.

[0037] Among them, the cross-domain mixup loss L CDM for:

[0038]

[0039] Among them, n mix represents the total number of virtual samples, σ represents the softmax activation function, represents the i-th virtual sample. Similarly, represents the label of the i-th virtual sample.

[0040] Among them, the selective entropy optimization strategy loss for:

[0041]

[0042] in, represents information entropy, {a 1 (x t ),a 2 (x t ),···,a k (x t )} represents the target domain sample x t k data augmentation versions of , i and j represent the indices of the last consistent and inconsistent augmentation versions, respectively.

[0043] Compared with the prior art, the present invention has the following advantages and technical effects:

[0044] Beneficial effects of the present invention:

[0045] (1) The present invention extends the traditional intra-domain mixup to cross-domain mixup and proposes a confident order-by-order cross-domain mixup method. This method makes full use of high-confidence target domain samples and source domain samples to perform order-by-order mixup operations, which not only improves the model's learning of target domain samples, but also generates new samples that can become a bridge connecting the source domain and the target domain, thereby improving the generalization ability of the model.

[0046] (2) In order to reduce the error accumulation caused by cross-domain mixup unconstrained self-training and improve the performance of the diagnostic model, the present invention starts from two aspects. On the one hand, the quality of pseudo-label samples is improved by filtering low-confidence labels with a fixed threshold. On the other hand, a selective entropy optimization strategy is used to reduce the impact of error accumulation on the model.

[0047] (3) A dynamic label distribution-aware margin regularization strategy is proposed, which dynamically adjusts the margin between categories based on the difficulty of classifying various samples in the target domain in the previous round, helping the model generalize to the unlabeled target domain. This strategy will give larger feature space to sample categories that are difficult to diagnose in the target domain, and increase the attention and learning of difficult categories in the target domain. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0049] Figure 1 Diagnostic network framework diagram provided by the present invention;

[0050] Figure 2 The CBAM structure diagram provided by the present invention;

[0051] Figure 3 A schematic diagram of the dynamic label distribution-aware marginal regularization (DLM) provided by the present invention;

[0052] Figure 4 This is a comparison chart of the diagnosis results of all the comparison methods provided by the present invention on the PU dataset. DETAILED DESCRIPTION

[0053] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0054] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0055] The present invention discloses a method for cross-domain fault diagnosis of rotating machinery based on a DACODL framework, which is composed of a trusted order-level cross-domain mixup module, a dynamic label distribution-aware marginal regularization module, and a joint distribution multi-layer adaptation module. The trusted order-level cross-domain mixup module uses the high-confidence target domain samples obtained by filtering and the source domain samples to perform order-level mixup operations to generate virtual samples. These generated intermediate samples become a bridge connecting the source domain and the target domain, improving the generalization ability of the model domain, and it can reduce the problem of decreased model diagnostic ability caused by unreliable pseudo-labels in cross-domain mixup. The proposed dynamic label distribution-aware marginal regularization module is used to improve the learning of difficult category samples. It can dynamically adjust the margin between the classification boundaries of each category according to the diagnostic difficulty assessment of each category in each round, and increase the model's attention to difficult samples by increasing the feature space of difficult category samples, thereby improving the diagnostic accuracy of difficult category samples. The present invention effectively improves the accuracy of cross-domain fault diagnosis by increasing the robustness and generalization of the model and improving the model's attention to difficult categories.

[0056] The present invention discloses a rotating machinery cross-domain fault diagnosis method based on the DACODL framework. The method increases the robustness and generalization of the model by making full use of target domain samples with high-confidence pseudo-labels for cross-domain mixup, and increases the model's attention to difficult samples by increasing the feature space of difficult category samples. The model's attention to difficult categories effectively improves the accuracy of cross-domain fault diagnosis.

[0057] To achieve the above object, the present invention provides a rotating machinery cross-domain fault diagnosis method based on the DACODL framework, comprising the following steps:

[0058] S1. Construct a labeled source domain dataset and an unlabeled target domain dataset using vibration data collected under different working conditions, and randomly divide the data into training sets and test sets;

[0059] S2. Perform computational order tracking (COT) transform on the source and target domain samples and perform other preprocessing to generate three data-enhanced strongly enhanced versions for each target domain sample;

[0060] S3. Build a diagnostic model and set hyperparameters;

[0061] S4. Put the target domain samples into the diagnostic model. The target domain samples are passed through the model to obtain the output of the label classifier, and the label probability is obtained through the softmax function. The target domain samples with label probabilities greater than the set threshold are placed together with their pseudo labels in the sample storage R of this round (assuming the current round is N). N ;

[0062] S5. From the previous round of sample storage R N-1 Adaptively sample various target domain samples and perform a mixup operation with the source domain samples in this round to generate a virtual data set;

[0063] S6. Put the source domain dataset, the target domain dataset and its enhanced version, and the cross-domain mixup virtual dataset into the diagnosis model for model training;

[0064] S7. Calculate the confidence order cross-domain mixup loss, dynamic label distribution aware regularized cross entropy loss, joint distribution multi-layer adaptation loss, and domain adversarial loss respectively, calculate the total loss and update the model parameters;

[0065] S8. Repeat S4-S7 to the maximum number of iterations to obtain a diagnostic model with good performance.

[0066] S9. Put the test set samples into the trained model to obtain the diagnosis results and compare them with other advanced methods.

[0067] As some embodiments, in step S2, other preprocessing methods include using a normalization function to convert signal values ​​to [0, 1], and applying data enhancement methods of several one-dimensional signals to source domain samples, and the data enhancement methods used are Gaussian noise, random cropping, time stretching and amplitude stretching.

[0068] As some embodiments, in step S3, the diagnostic model is composed of a feature extractor G f , label classifier G c and domain discriminator G d composition.

[0069] As some embodiments, in step S4, the target domain samples released in each round are the same, all the target domain samples are passed through the model to obtain the label classifier output, and then the label probability is obtained through the softmax function, and the target domain samples with label probabilities greater than the set threshold are placed together with their pseudo labels into the sample storage of this round. The low confidence sample filtering process can be expressed by the following formula:

[0070]

[0071] Among them, x t is the COT transformation input of the target domain sample, τ is the set threshold size, σ(·) represents the softmax function, T is the temperature coefficient, G represents the high-confidence pseudo-label predicted by the model, -1 means discarding the low-confidence sample, f () indicates feature extractor processing, G c () indicates label classifier processing.

[0072] After unreliable pseudo labels are filtered out, high-confidence samples of each category in the target domain and their pseudo label pairs are obtained. Put them into the sample storage of this round one by one according to their categories inside.

[0073] As some embodiments, in step S5, from the previous round of sample storage In the adaptive sampling, various target domain samples are sampled and mixed up with the source domain samples in this round. The adaptive sampling strategy can be expressed as follows:

[0074]

[0075] Among them, p c Indicates that the c-type sample is from the sample storage The probability of being sampled, N c Indicates sample storage In , ν is a hyperparameter used to adjust the sampling frequency, and C represents the number of categories.

[0076] As some embodiments, in step S5, from the previous round of sample storage In the adaptive sampling, various target domain samples are sampled and mixed with the source domain samples in this round. The mixup operation can be expressed as follows:

[0077]

[0078] Among them, x mix ,y mix Represents the new samples generated and their labels, Indicates that the sample memory Adaptively sample a batch of high-confidence target domain samples and their corresponding pseudo labels. Indicates the number of source domain samples in a batch in this round, and the cross-domain mixup coefficient Each mini-batch is randomly sampled from the Beta distribution Beta(α,α), where α = 0.2, and

[0079] As some embodiments, in step S7, the total loss is calculated. It can be expressed as follows:

[0080]

[0081] in, represents the confidence order cross-domain mixup module loss, represents the cross entropy loss with dynamic label distribution aware margin regularization, represents the joint distribution multi-layer adaptation loss, represents the domain adversarial loss, ε adv represents the weight of the domain adversarial loss, μ t represents the time-varying trade-off coefficient, which is t represents the current round number, t m Indicates the maximum number of rounds.

[0082] As some embodiments, the It is expressed as follows:

[0083]

[0084] in

[0085] Among them, z c =G c (G f (x s )) c represents the c-th output of the label classifier, i.e., the c-class output, ν2 is a hyperparameter used to adjust the marginal distance of each category, N c Indicates sample storage The number of samples in category c, C represents the number of categories, and K is a preset constant used to control the maximum margin of the category, which is taken as 0.5 in this paper.

[0086] As some embodiments, the domain adversarial loss It can be expressed as follows:

[0087]

[0088] Among them, D s and D t Represent the source domain samples x s and target domain sample x t The distribution it obeys.

[0089] As some embodiments, the jointly distributed multi-layer adaptation loss It can be expressed as follows:

[0090]

[0091] Among them, γ 1 and γ 2 The trade-off factor between the global distribution difference and subdomain distribution difference of the same layer features, γ 3 and γ 4 Represents the trade-off factor of the difference in joint distribution between different layers, G pc (·) represents the features obtained from the penultimate layer of the label classifier, and Represents the proportion of class c samples in the source domain and target domain, respectively, which can be calculated by the number of class c samples in the source domain true label and the target domain pseudo label. and are all equal to 1, and is the weighted sum of category c, Φ u (·) represents the features generated by the u-th feature kernel.

[0092] As some embodiments, the It can be expressed by the following formula:

[0093]

[0094] in, represents the cross-domain mixup loss, ε CDM represents the cross-domain mixup loss weight, represents the selective entropy optimization strategy loss, μ t represents the time-varying trade-off coefficient.

[0095] As some embodiments, the It is expressed as follows:

[0096]

[0097] Among them, n mix represents the total number of virtual samples, σ represents the softmax activation function, represents the i-th virtual sample. Similarly, represents the label of the i-th virtual sample.

[0098] As some embodiments, the It is expressed as follows:

[0099]

[0100] in, represents information entropy, {a 1 (x t ),a 2 (x t ),···,a k (x t )} represents the target domain sample x t k data augmentation versions, in this paper k=3; when more than half of the augmentation version prediction results are the same as the original target domain samples, they are marked as consistent, otherwise they are marked as inconsistent. i and j represent the indexes of the last consistent and inconsistent augmentation versions, respectively.

[0101] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0102] like Figure 1 As shown, the present invention provides a rotating machinery cross-domain fault diagnosis method based on the DACODL framework, comprising the following steps:

[0103] S1. Construct a labeled source domain dataset and an unlabeled target domain dataset using vibration data collected under different working conditions, and randomly divide the data into training sets and test sets;

[0104] Specifically, the Paderborn University bearing dataset (PU dataset) is used as the experimental object. The data is collected by a piezoelectric accelerometer under four operating conditions to collect the vibration signal of the bearing seat, where the bearing model is 6203 deep groove ball bearing, and the sampling frequency is 64kHz. The detailed information of the four operating conditions of the PU dataset is shown in Table 1, which are respectively denoted as C0, C1, C2, and C3. The present invention divides all samples into 14 types of health states, and the PU data classification information is shown in Table 2. In the experiment of the present invention, 12 cross-domain tasks were carried out, such as Figure 3 As shown in the figure, C01 indicates that the data with working conditions C0 and C1 are selected as the source domain and target domain of the migration task respectively. The training set adopts overlapping sampling, the length of the sliding window is 2560, and the sliding step is 512. The test set is sampled in sequence without overlapping, and the length of the sliding window is the same as that of the training set, which is 2560. The training set contains 320 samples for each health state, while the test set contains 32 samples for each health state.

[0105] Table 1

[0106]

[0107] Table 2

[0108]

[0109] S2. Perform computational order tracking (COT) transform on the source and target domain samples and perform other preprocessing to generate three data-enhanced strongly enhanced versions for each target domain sample;

[0110] Specifically, the segmented data is resampled through COT and converted to the frequency domain through FFT, and the first 1024 points are intercepted to obtain the input samples. Then all data are mapped to [0, 1] using the normalization method, and the source domain data is processed using Gaussian noise, random cropping, time stretching, and amplitude stretching data enhancement methods. For each target domain data, three strongly enhanced versions are generated after three random cropping and amplitude stretching data enhancements.

[0111] S3. Build a diagnostic model and set hyperparameters;

[0112] Specifically, the model consists of a feature extractor G f , label classifier G c and domain discriminator G d The composition and specific structural parameters are shown in Table 3. The detailed structure of CABM in the table is shown in Figure 2 The learning rate is set to 0.001 at the beginning of training, and adjusted to 1e-4 when the epoch is 15. The batch size is 32, the maximum training epoch is set to 30, and other hyperparameter sizes are shown in Table 4.

[0113] Table 3

[0114]

[0115] Table 4

[0116]

[0117]

[0118] S4. Put the target domain samples into the diagnostic model. The target domain samples are passed through the model to obtain the output of the label classifier, and the label probability is obtained through the softmax function. The target domain samples with label probabilities greater than the set threshold are placed together with their pseudo labels into the current round of sample storage.

[0119] Specifically, the low confidence sample filtering step can be expressed by the following formula:

[0120] σ(G c (G f (x t )) / T)=p

[0121]

[0122] Among them, x t is the COT transformation input of the target domain sample, τ is the set threshold size, σ(·) represents the softmax function, T is the temperature coefficient, Represents the high-confidence pseudo label predicted by the model, and -1 means discarding the low-confidence sample.

[0123] Finally, the selected high-confidence target domain samples and their pseudo labels are placed in the current round sample storage R N .

[0124] S5. From the previous round of sample storage In the process, various target domain samples are adaptively sampled and mixed up with the source domain samples in this round to generate virtual samples.

[0125] Specifically, the adaptive sampling strategy can be expressed by the following formula:

[0126]

[0127] Among them, p c Indicates that class c samples in adaptive sampling are taken from the sample storage The sampling frequency, N c Indicates sample storage In , ν is a hyperparameter used to adjust the sampling frequency, and C represents the number of categories.

[0128] The cross-domain mixup operation can be expressed as follows:

[0129]

[0130] Among them, x mix ,y mix Represents the new samples generated and their labels, Indicates that the sample memory Adaptively sample a batch of high-confidence target domain samples and their corresponding pseudo labels. represents the number of source domain samples in a batch in this round, Each mini-batch is randomly sampled from the Beta distribution Beta(α,α), where α = 0.2, and

[0131] S6. Put the source domain dataset, the target domain dataset and its enhanced version, and the cross-domain mixup virtual samples into the diagnosis model for model training;

[0132] S7. Calculate the confidence order cross-domain mixup loss, dynamic label distribution aware regularized cross entropy loss, joint distribution multi-layer adaptation loss, and domain adversarial loss respectively, calculate the total loss and update the model parameters;

[0133] Specifically, the total loss It can be expressed as follows:

[0134]

[0135] in, represents the confidence order cross-domain mixup module loss, represents the cross entropy loss with dynamic label distribution aware margin regularization, represents the joint distribution multi-layer adaptation loss, represents the domain adversarial loss, ε adv represents the weight of the domain adversarial loss, μ t represents the time-varying trade-off coefficient, which is The t in the formula represents the current round number, t m Indicates the maximum number of rounds.

[0136] The calculation of total loss is divided into four parts:

[0137] In the first part, in order to increase the model's attention to difficult categories and improve its diagnostic ability for difficult categories, dynamic label distribution aware margin regularization is proposed. The cross entropy loss of dynamic label distribution aware margin regularization It is expressed as follows:

[0138]

[0139] in

[0140] Among them, z c =G c (G f (x s )) c represents the c-th output of the label classifier, ν 2 It is a hyperparameter used to adjust the marginal distance of each category, N c Indicates sample storage The number of samples in category c, C represents the number of categories, and K is a preset constant used to control the maximum margin of the category, which is taken as 0.5 in this paper.

[0141] In the second part, an adversarial training strategy is used to extract domain-invariant features from the source domain and the target domain to improve the generalization ability of the model in the source domain and the target domain. It can be expressed as follows:

[0142]

[0143] Among them, D s and D t Represent the source domain samples x s and target domain sample x t The distribution it obeys.

[0144] In the third part, in order to promote the consistency of low-level and high-level features between source and target domain samples, accelerate the convergence of the model and improve the adaptability of the model, multi-layer adaptation is used to align the three-layer features of the label classifier. Joint distribution multi-layer adaptation loss It can be expressed as follows:

[0145]

[0146] Among them, γ 1 and γ 2 The trade-off factor between the global distribution difference and subdomain distribution difference of the same layer features, γ 3 and γ 4 Represents the trade-off factor of the difference in joint distribution between different layers, G pc (·) represents the features obtained from the penultimate layer of the label classifier, and Represents the proportion of class c samples in the source domain and target domain, respectively, which can be calculated by the number of class c samples in the source domain true label and the target domain pseudo label. and are all equal to 1, and is the weighted sum of category c, Φ u (·) represents the features generated by the u-th feature kernel.

[0147] In the fourth part, we use the confidence order cross-domain mixup loss to narrow the gap between the two domains and improve the generalization ability of the model. It can be expressed by the following formula:

[0148]

[0149] in, represents the cross-domain mixup loss, ε CDM represents the cross-domain mixup loss weight, represents the selective entropy optimization strategy loss, μ t represents the time-varying trade-off coefficient.

[0150] Calculating the confidence-order cross-domain mixup loss can be divided into two steps.

[0151] The first step is to calculate the cross-domain mixup loss. It is expressed as follows:

[0152]

[0153] Among them, n mix represents the total number of virtual samples, σ represents the softmax activation function, represents the i-th virtual sample. Similarly, represents the label of the i-th virtual sample.

[0154] The second step is to use selective entropy to optimize the strategy loss to reduce the problem of error accumulation. It is expressed as follows:

[0155]

[0156] in, represents information entropy, {a 1 (x t ),a 2 (x t ),···,a k (x t )} represents the sample x t k data enhancement versions, k = 3 in this paper. When most of the data enhancement version predictions are the same as the original sample predictions, they are marked as consistent, otherwise they are marked as inconsistent. i and j represent the indexes of the last consistent and inconsistent enhancement versions, respectively.

[0157] S8. Repeat S4-S7 to the maximum number of iterations to obtain a diagnostic model with good performance.

[0158] S9. Put the test set samples into the trained model to obtain the diagnosis results and compare them with other advanced methods.

[0159] Specifically, the comparison results with other advanced methods are as follows Figure 4 As shown in the figure, the average accuracy of the proposed method on all tasks reached 92.58%. Compared with CORAL, the worst performing method among the comparison methods, the average accuracy of the proposed method on all tasks increased by 25.62%, and compared with DSAN, the best performing method among the comparison methods, the average accuracy of the proposed method on all tasks increased by 15.4%.

[0160] Therefore, the present invention adopts the above-mentioned cross-domain fault diagnosis method based on the DACODL framework to solve the fault diagnosis problem under complex working conditions. Among them, the proposed confidence order cross-domain mixup module uses the high-confidence target domain samples and source domain samples to perform order cross-domain mixup operations to generate virtual data. The generated intermediate samples become a bridge connecting the source domain and the target domain, improving the generalization ability of the model domain, and it can alleviate the problem of decreased model diagnostic ability caused by unreliable pseudo-labels in cross-domain mixup. In addition, the proposed dynamic label distribution-aware marginal regularization module is used to improve the learning of difficult category samples. It can increase the model's attention to difficult samples by increasing the feature space of difficult category samples according to the diagnostic difficulty of the sample category, thereby improving the diagnosis of difficult category samples. The experimental results show that the proposed method effectively improves the accuracy of cross-domain fault diagnosis.

[0161] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A cross-domain fault diagnosis method for rotating machinery based on DACODL framework, characterized in that: include: S1. Obtain vibration data of rotating machinery under different working conditions, and construct a labeled source domain data set and an unlabeled target domain data set according to the vibration data under the different working conditions, wherein the label is a fault diagnosis result; S2. performing calculation order tracking transformation and other preprocessing on the source domain data set and the target domain data set, and performing data enhancement on the target domain data set to obtain an enhanced data set of the target domain data set; S3. Build a diagnostic model and set hyperparameters; S4. In each round of training, the preprocessed target domain data is processed by the diagnostic model to obtain high-confidence samples and pseudo-labels, and the high-confidence samples and pseudo-labels are stored in the sample storage of the current round; S5. In the current round of training, adaptive sampling is performed from the sample storage of the previous round, and the sampling results are mixed up with the source domain samples of the current round to obtain a virtual data set, wherein the virtual data set includes virtual samples and corresponding labels; S6. Training the diagnostic model using the preprocessed source domain dataset, target domain dataset, enhanced dataset and virtual dataset; S7. Calculate the total loss of the diagnostic model trained in the current round, and update the model parameters of the diagnostic model according to the total loss of the current round; S8. Repeat the above S4-S7 until the training rounds reach the maximum number of times to obtain the final diagnosis model; S9. Test the final diagnostic model, and obtain the cross-domain fault diagnosis result of the rotating machinery by analyzing the vibration data of the rotating machinery under the current working condition through the tested final diagnostic model.

2. The method according to claim 1, characterized in that In step S2, other preprocessing methods include normalization and data enhancement, and the data enhancement methods include Gaussian noise, random cropping, time stretching and amplitude stretching.

3. The method according to claim 1, characterized in that In step S3, the diagnosis model includes a feature extractor, a label classifier and a domain discriminator, wherein the feature extractor is connected to the label classifier and the domain discriminator respectively.

4. The method according to claim 1, characterized in that: In step S4, the process of acquiring the virtual data set includes: The preprocessed target domain data and enhanced data set corresponding to the current round are processed by the diagnostic model, and the maximum label probability of each sample in the target domain data and the enhanced data set is obtained through the softmax function. Samples with maximum label probabilities higher than the set threshold are screened to obtain high-confidence samples, and the label corresponding to the maximum label probability is used as a pseudo-label.

5. The method according to claim 1, characterized in that In step S5, the process of adaptive sampling includes: Among them, p c Indicates the sample storage of class c from the N-1th round The probability of being sampled, N c Indicates the sample storage of the N-1th round In , ν is a hyperparameter used to adjust the sampling frequency, and C represents the number of categories.

6. The method according to claim 1, characterized in that In step S6, the total loss is: in, represents the confidence order cross-domain mixup module loss, represents the cross entropy loss with dynamic label distribution aware margin regularization, represents the joint distribution multi-layer adaptation loss, represents the domain adversarial loss, ε adv represents the weight of the domain adversarial loss, μ t represents the time-varying trade-off coefficient, which is t represents the current round number, t m Indicates the maximum number of rounds.

7. The method according to claim 6, characterized in that Cross-entropy loss with dynamic label distribution-aware margin regularization for: in Where zc = G c (G f (x s )) c represents the c-th output of the label classifier, i.e., the c-class output, ν2 is a hyperparameter used to adjust the marginal distance of each category, N c Indicates the sample storage of the N-1th round The number of samples of category c, C represents the number of categories, and K is a preset constant used to control the maximum margin of the category.

8. The method according to claim 6, characterized in that Domain Adversarial Loss for: Among them, D s and D t Represent the source domain samples x s and target domain sample x t The distribution it obeys.

9. The method according to claim 6, characterized in that Jointly distributed multi-layer adaptation loss for: Among them, γ1 and γ2 represent the trade-off factors between the global distribution difference and the subdomain distribution difference of the same layer features, γ3 and γ4 represent the trade-off factors of the joint distribution difference between different layers, and G pc (·) represents the features obtained from the penultimate layer of the label classifier, and They represent the proportion of class C samples in the source domain and the target domain, respectively, and are calculated by the number of class C samples in the source domain true label and the target domain pseudo label, respectively. and are all equal to 1, and is the weighted sum of category c, Φ u (·) represents the features generated by the u-th feature kernel.

10. The method according to claim 6, characterized in that Confident order cross-domain mixup module loss for: in, represents the cross-domain mixup loss, ε CDM represents the cross-domain mixup loss weight, represents the selective entropy optimization strategy loss, μ t represents the time-varying trade-off coefficient; Among them, cross-domain mixup loss for: Among them, n mix represents the total number of virtual samples, σ represents the softmax activation function, represents the i-th virtual sample. Similarly, represents the label of the i-th virtual sample; Among them, the selective entropy optimization strategy loss for: in, represents information entropy, {a1(x t ),a2(x t ),···,a k (x t )} represents the target domain sample x t k data augmentation versions of , i and j represent the indices of the last consistent and inconsistent augmentation versions, respectively.

Images