Unsupervised domain adaptive fault diagnosis method based on time-frequency domain collaborative feature alignment

The unsupervised domain adaptive fault diagnosis method based on time-frequency domain collaborative feature alignment solves the problem of insufficient generalization ability of deep learning models under the data distribution offset in the target domain, realizes efficient cross-domain fault diagnosis and equipment condition assessment, and improves the stability and efficiency of industrial production.

CN121009461APending Publication Date: 2025-11-25CHINA UNIV OF MINING & TECH
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202510941869.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-07-09
Publication Date
2025-11-25

AI Technical Summary

Technical Problem

Existing deep learning models lack generalization ability in industrial scenarios where target domain data is unlabeled and has distributional bias, and fail to fully utilize the complementarity of time-frequency domain features, resulting in a decline in cross-domain fault diagnosis performance.

Method used

An unsupervised domain adaptive fault diagnosis method based on time-frequency domain collaborative feature alignment is adopted. Time-frequency features are extracted by sharing convolutional neural networks. Combined with the local maximum mean difference algorithm and gradient inversion mechanism, a domain adaptive learning module and a dual classifier collaborative decision module are constructed to achieve fine-grained alignment of time-frequency features and multimodal integrated decision.

Benefits of technology

It significantly improves the accuracy and robustness of cross-domain fault diagnosis in complex industrial environments, and is suitable for intelligent condition assessment and cross-scenario deployment of equipment such as rotating machinery, reducing equipment downtime risks and optimizing production efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121009461A_ABST
    Figure CN121009461A_ABST
Patent Text Reader

Abstract

The invention discloses an unsupervised domain adaptive fault diagnosis method based on time-frequency domain collaborative feature alignment. The method comprises the following steps: acquiring source domain data and target domain data; constructing a time-frequency feature extraction module by using a shared convolutional neural network, and performing time-frequency collaborative learning when time domain and frequency domain features are extracted; respectively carrying out distribution alignment in a fine-grained category on the time domain and frequency domain features; constructing a domain classification discriminator network by adopting a multi-layer full connection layer; training a time-frequency feature extraction module based on a unified domain adversarial training mechanism of time-frequency fusion features; constructing a dual-classifier collaborative decision module by using a time domain classifier and a frequency domain classifier; a cross-domain fault diagnosis model is formed based on a time-frequency feature extraction module, a domain adaptive learning module and a dual-classifier collaborative decision module; and carrying out online diagnosis operation by using the cross-domain fault diagnosis model. According to the method, the fault state of the complex industrial equipment under the variable working conditions can be accurately identified.
Need to check novelty before this filing date? Find Prior Art

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of industrial equipment state monitoring and cross-domain transfer learning, and specifically relates to an unsupervised domain adaptive fault diagnosis method based on time-frequency domain collaborative feature alignment. BACKGROUND

[0002] Modern industrial systems are accelerating towards automation and intelligentization, and equipment health monitoring and fault diagnosis technology has become an important guarantee for the safe and stable operation of industrial systems. The core goal of fault diagnosis is to achieve accurate prediction and timely maintenance before equipment failure occurs, to prevent equipment performance degradation, reduce maintenance costs, and reduce safety risks in the production process. Therefore, industrial equipment health monitoring and fault diagnosis technology has gradually attracted widespread attention.

[0003] In recent years, with the rapid development of intelligent manufacturing, advanced sensing technology and big data storage and analysis capabilities, industrial data acquisition and analysis methods have become increasingly rich, promoting the development of data-driven fault diagnosis methods. Although data-driven fault diagnosis has made significant progress in feature extraction and fault recognition, it still faces many challenges in practical applications. On the one hand, traditional fault diagnosis methods often rely on complex mathematical modeling and expert knowledge, and are highly dependent on feature engineering in high-dimensional data environments, limiting the model's ability to learn automatically and its generalization performance. On the other hand, the variability of equipment operating environments leads to inconsistent data distribution between source and target domains, and existing methods perform poorly when the target domain data distribution shifts, severely restricting their application in industrial settings.

[0004] To solve the above problems, deep learning technology has gradually become an important tool in the field of fault diagnosis. Deep learning has the ability to automatically extract features end-to-end, and can effectively handle complex high-dimensional data, significantly improving fault diagnosis performance. A variety of existing deep learning models combined with residual networks, long short-term memory networks, etc. have been successfully applied to mechanical equipment state monitoring, achieving efficient recognition of fault patterns. However, deep learning models usually rely on large-scale, complete and consistent labeled data, making it difficult to adapt to the actual industrial scene where target domain data is unlabeled and has distribution bias, resulting in insufficient model generalization ability and decreased diagnosis performance.

[0005] To address the aforementioned issues, unsupervised domain adaptation (UDA) has emerged as an effective technique for mitigating data distribution bias and improving cross-domain fault diagnosis performance. UDA methods enhance the model's generalization ability in the unlabeled target domain by learning shared feature representations using labeled source domain data. Current UDA methods primarily include statistical alignment methods based on difference metrics and adversarial training methods based on adversarial learning. While some methods have reduced inter-domain distribution discrepancies and improved diagnostic accuracy to some extent, most focus only on overall feature alignment, neglecting category decision boundaries, which can easily lead to misalignment of target domain samples and negative transfer. Furthermore, existing research largely focuses on domain adaptation of time-domain features, without fully considering the potential performance improvement of frequency-domain features in time-series domain adaptation.

[0006] In fact, time-domain and frequency-domain features are significantly complementary in fault diagnosis, and the distribution of time-frequency features varies significantly under different operating conditions and domains. Existing methods fail to fully align time-frequency modal features, making it difficult to comprehensively alleviate the cross-domain offset problem. Therefore, it is of great significance to provide an unsupervised domain adaptive fault diagnosis method based on time-frequency collaborative feature alignment, which aims to fully explore the complementary information of time-domain and frequency-domain features, clarify the decision boundary, and thus significantly improve the accuracy and robustness of cross-domain diagnosis in complex industrial environments. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides an unsupervised domain adaptive fault diagnosis method based on time-frequency domain collaborative feature alignment. This method can accurately identify the fault state of complex industrial equipment under varying operating conditions, achieve efficient diagnosis of unlabeled target domains, and has good transfer and generalization capabilities and fault identification robustness. It is suitable for intelligent condition assessment and cross-scenario deployment of industrial equipment such as rotating machinery.

[0008] To achieve the above objectives, this invention provides an unsupervised adaptive fault diagnosis method based on time-frequency domain collaborative feature alignment, comprising the following steps:

[0009] Step 1: Data sampling and preprocessing;

[0010] Data from the source domain and the target domain are collected and preprocessed to form a sample dataset. The sample dataset is then divided into a training set and a test set according to a set ratio.

[0011] Step 2: Construct a time-frequency feature extraction module;

[0012] A time-frequency feature extraction module is constructed using a shared convolutional neural network, and this module performs time-frequency co-learning when extracting time-domain and frequency-domain features. As input, the temporal features of the source domain Temporal characteristics of the target domain Frequency domain characteristics of the source domain Frequency domain characteristics of the target domain As the extracted output;

[0013] Step 3: Construct the domain adaptive learning module;

[0014] Local maximum mean difference (MMD) algorithms are introduced into the time and frequency domain feature spaces respectively to construct domain adaptation layers, and the MMD loss in the time and frequency domains is utilized. Optimization yields a domain adaptive learning module that performs fine-grained intra-category distribution alignment for both time-domain and frequency-domain features; the domain adaptive learning module uses... As input, the temporal feature space of the source domain Temporal feature space of the target domain Frequency domain feature space of the source domain Frequency domain feature space of the target domain As output;

[0015] Step 4: Construct the domain classification module;

[0016] A domain classification discriminator network is constructed using multiple fully connected layers, and the domain classification loss L is utilized. adv The optimization process yields a domain classification module, which is a source-domain and frequency-domain fused feature vector obtained by concatenating time-domain and frequency-domain features along the feature dimension. As input, the probability of the source domain or the target domain is output;

[0017] Step 5: Train the time-frequency feature extraction module based on the unified domain adversarial training mechanism of time-frequency fusion features;

[0018] Based on L adv A gradient reversal mechanism is introduced to conduct adversarial training between the time-frequency feature extraction module and the domain classification module, which prompts the time-frequency feature extraction module to learn the invariant features shared by the source domain and the target domain.

[0019] Step Six: Construct a dual-classifier collaborative decision-making module;

[0020] A dual-classifier collaborative decision-making module is constructed using a time-domain classifier and a frequency-domain classifier. For both the time-domain and frequency-domain classifiers, source-domain labeled samples are used to apply cross-entropy loss during time-domain classification. Cross-entropy loss in frequency domain classification Optimize the module to obtain a trained dual-classifier collaborative decision-making module;

[0021] Step 7: Construct a cross-domain fault diagnosis model;

[0022] S71: A cross-domain fault diagnosis model is formed based on the time-frequency feature extraction module, the domain adaptive learning module, and the dual classifier collaborative decision-making module;

[0023] S72: Construct the total loss L;

[0024] S73: A cross-domain fault diagnosis model is trained using the training set. During training, supervised learning is performed on the time-domain and frequency-domain classifiers, respectively, based on source domain labels, while unsupervised learning is performed on the unlabeled target domain classifiers. A weighted fusion strategy is then used to integrate the classification decisions of the two modalities. The total classification loss L is jointly optimized. D LMMD total loss L L and countermeasures against losses L D This allows the model parameters to be iteratively optimized until the overall loss is minimized, resulting in a cross-domain fault diagnosis model with high robustness and generalization ability.

[0025] Step 8: Perform real-time diagnostics of cross-domain faults;

[0026] S81: Collect target data X online using a sensor array deployed in an industrial system, and normalize and standardize X.

[0027] S82: The data is sampled using a sliding window to obtain a sequence of data of consistent length, and then the obtained time-domain signal is subjected to Fourier transform to obtain the frequency signal;

[0028] S83: The time-domain and frequency-domain signals of the unlabeled target domain are used as input data and fed into the cross-domain fault diagnosis model. In the cross-domain fault diagnosis model, the time-frequency feature extraction module is first used to extract high-dimensional features in the time-frequency domain. Then, the domain adaptation layer is used to perform LMMD measurement on the high-dimensional features in the time domain and frequency domain respectively to reduce the difference in feature distribution between the source domain and the target domain. Then, the dual classifier collaborative decision module is used for inference. In the inference stage, the output results of the time-domain classifier and the frequency-domain classifier are combined through a weighted fusion strategy to finally output the category prediction result.

[0029] As a preferred option, the process of collecting source domain data and target domain data in step one, and forming a sample dataset after preprocessing, is as follows:

[0030] S11: Acquire raw vibration signal data from the source region using a sensor array. and the original vibration signal data of the target domain in, and Data from the source domain and the target domain, respectively. For source domain tags, Let L be a real number, and L be the length of the time series for a single sample. s Nt These represent the number of samples in the source domain and the target domain, respectively.

[0031] S12: Perform data normalization and standardization processing on the source domain data and the target domain data;

[0032] S13: Use a sliding window to partition the source domain data, obtaining batch datasets (X) that serve as input data each time. N ,Y N ), where N represents the number of samples in a batch; the target domain data is processed in the same way as the source domain data; thus, the time domain signals of the source domain are obtained respectively. and the time-domain signal of the target domain

[0033] S14: Introducing frequency domain feature information, a Fast Fourier Transform is applied to each preprocessed sample to obtain the frequency domain signal from the source domain. Frequency domain signal of the target domain

[0034] S15: Obtain the time-domain signal from the source domain. Time-domain signal of the target domain Source domain frequency domain signal Frequency domain signal of the target domain As a sample dataset.

[0035] As a preferred option, in step S12 of step one, the source domain data and the target domain data are normalized according to formula (1), and the source domain data and the target domain data are standardized according to formula (2).

[0036]

[0037] In the formula, Z′ represents the data after normalization preprocessing; Z represents the original input data; min(Z) represents the minimum value of the input data; and max(Z) represents the maximum value of the input data.

[0038]

[0039] In the formula, Z″ represents the data after regularization preprocessing, μ represents the mean of the input data, and σ represents the standard deviation of the input data.

[0040] As a preferred option, in step S13 of step one, X is obtained according to formula (3). N Y is obtained according to formula (4) N ;

[0041]

[0042] As a preferred embodiment, in step S14 of step one, the frequency domain signal of the source domain is obtained according to formula (5). The frequency domain signal of the target domain is obtained according to formula (6).

[0043]

[0044] In the formula, These represent the time-domain signals from the source domain, respectively. These represent the time-domain signals of the target domain, respectively.

[0045] As a preferred embodiment, in step two, the temporal characteristic space of the source domain is obtained according to formula (7). The temporal feature space of the target domain is obtained according to formula (8). The frequency domain feature space of the source domain is obtained according to formula (9). The frequency domain feature space of the target domain is obtained according to formula (10).

[0046]

[0047]

[0048] In the formula, F(·) represents the mapping function of the shared convolutional neural network.

[0049] As a preferred embodiment, in step three, the local maximum mean difference loss in the time domain and frequency domain is obtained according to formulas (11) and (12), respectively.

[0050]

[0051] In the formula, Φ(·) is the kernel mapping function, and C is the total number of categories. Let c be the set of time-domain feature samples of category c in the source domain. Let c be the set of time-domain feature samples of category c in the target domain. Let c be the set of frequency domain feature samples of category c in the source domain. Let c be the set of frequency domain feature samples of category c in the target domain.

[0052] As a preferred option, in step four, the source domain and frequency domain fused feature vectors are obtained according to formulas (13) and (14), respectively. Meanwhile, the domain classification loss L is obtained according to formula (15). adv ;

[0053]

[0054] In the formula, ";" indicates that the splicing operation is performed in the column direction;

[0055]

[0056] In the formula, For the output of the domain classifier, d i These are real domain tags.

[0057] As a preferred embodiment, in step six, the temporal classification loss is obtained according to formulas (16) and (17), respectively. Frequency domain classification loss

[0058]

[0059] In the formula, C T (·) and C F (·) represent the time-domain classifier and the frequency-domain classifier, respectively.

[0060] As a preferred option, the process of constructing the total loss L in step seven is as follows:

[0061] S72-1: Obtain the total classification loss L according to formula (18) C And make the parameters of the time-frequency feature extractor F(x) Parameters of the domain adaptive layer A(x) Temporal classifier classifier C T The parameter of (x) Frequency domain classifier C F The parameter of (x) It satisfies formula (19);

[0062]

[0063] In the formula, λ T , λ F These are the decision weights for the time-domain and frequency-domain classifiers, respectively, λ. T +λ F =1;

[0064]

[0065] In the formula, For the temporal characteristics of the source domain, Labels for the temporal features of the source domain. For the frequency characteristics of the source domain, F(·) is the label for the frequency features of the source domain, F(·) is the time-frequency domain feature extractor, A(·) is the domain adaptation layer, and C(·) is the classifier.

[0066] S72-2: Obtain the total LMMD loss L according to formula (20) L And make the parameters of the time-frequency feature extractor F(x) Parameters of the domain adaptive layer A(x) Satisfies formula (21);

[0067]

[0068] In the formula,

[0069] S72-3: Obtain the adversarial loss L according to formula (22) D And make the parameters of the time-frequency feature extractor F(x) Parameters of the domain adaptive layer A(x) Parameters of the domain classifier D(x) Satisfies formula (23);

[0070]

[0071] In the formula, D is the domain classifier, and d is the binary domain label, indicating whether the sample x∈S or x∈T;

[0072]

[0073] S72-4: Obtain the total loss L according to formula (24);

[0074]

[0075] In the formula, δ is the balancing hyperparameter of the domain adaptive learning module, and β is the balancing hyperparameter of the domain classification module;

[0076] S72-5: Adjust the relative weights of δ and β to satisfy δ = 1 - β, and transform formula (24) into formula (25);

[0077]

[0078] This invention provides an unsupervised adaptive fault diagnosis method based on time-frequency domain collaborative feature alignment to address the degradation of cross-domain diagnostic performance caused by distribution misalignment between the source and target domains in complex industrial environments. First, raw vibration signals from the source and target domains are acquired and processed using standardization and Fast Fourier Transform (FFT) to obtain time-domain and frequency-domain data in a unified format. Second, a shared convolutional neural network (CNN) is used to extract multimodal high-order features from the time and frequency domains of the source and target domains, constructing a unified multimodal high-dimensional representation that effectively compresses redundancy and enhances time-frequency collaborative expression capabilities. Next, a Local Maximum Mean Difference (LMMD) algorithm is introduced, applying the LMMD criterion to both the time and frequency domain feature spaces. This aligns the time-domain and frequency-domain features with fine-grained class distribution within their respective intra-class subspaces, significantly reducing intra-modal mismatches between classes, minimizing cross-domain distribution differences, and improving cross-domain discrimination capabilities. Through deep collaboration and fine-grained distribution alignment of time-frequency domain information, the accuracy and stability of the diagnosis are significantly improved, making it suitable for intelligent monitoring of various rotating machinery equipment. Then, based on the fused time-frequency features, a unified domain discriminant network is constructed, and adversarial training is implemented using a gradient inversion layer (GRL) to enable the feature extractor to learn domain-indistinguishable shared feature representations. Furthermore, a dual-classifier structure is constructed to independently learn time-domain and frequency-domain features, and a weighted fusion strategy is used to achieve multimodal integrated decision-making. Finally, considering the complementarity of time-domain and frequency-domain information, an end-to-end trainable cross-domain diagnostic framework is constructed through four stages: multimodal feature extraction, intra-modal alignment, domain adversarial feature extraction, and multi-classifier fusion decision-making. Simultaneously, during training, supervised learning is performed on the time-domain and frequency-domain classifiers in conjunction with source domain labels, and the classification decisions of the two modalities are integrated through a weighted fusion strategy. By jointly optimizing the classification loss, LMMD loss, and adversarial loss, the model parameters are iteratively optimized until the overall loss is minimized, thereby obtaining a cross-domain fault diagnosis model with high robustness and generalization ability. This invention can fully exploit the complementarity of time-domain and frequency-domain features. Through multimodal feature extraction, fine-grained intra-class distribution alignment, domain adversarial training, and weighted fusion of dual classifiers, it constructs an end-to-end cross-domain diagnostic framework, which effectively improves the model's generalization performance and robustness in the target domain.

[0079] This method can accurately identify the fault status of complex industrial equipment under varying operating conditions, achieve efficient diagnostic operations in unlabeled target domains, and has good transfer and generalization capabilities and fault identification robustness. It is suitable for intelligent condition assessment and cross-scenario deployment of industrial equipment such as rotating machinery, and can provide strong support for real-time fault warning and maintenance decisions in industrial production processes. It effectively reduces the risk of equipment downtime, optimizes production efficiency, and promotes the intelligent development of industry. Attached Figure Description

[0080] Figure 1This is a flowchart of the domain adaptive fault diagnosis method in this invention;

[0081] Figure 2 This is a schematic diagram of the cross-domain fault diagnosis model in this invention;

[0082] Figure 3 This is the accuracy result of the cross-domain fault diagnosis model in this invention;

[0083] Figure 4 This is a schematic diagram of the confusion matrix of the model performance under a task (D→A) in this invention;

[0084] Figure 5 This is the feature visualization result based on the cross-domain fault diagnosis model in this invention;

[0085] Figure 6 This is a schematic diagram of the CWRU experimental device in this invention. Detailed Implementation

[0086] The invention will now be further described with reference to the accompanying drawings.

[0087] like Figures 1 to 5 As shown, this invention provides an unsupervised domain adaptive fault diagnosis method based on time-frequency collaborative feature alignment, comprising the following steps:

[0088] Step 1: Data sampling and preprocessing;

[0089] Data from the source domain and the target domain are collected and preprocessed to form a sample dataset. The sample dataset is then divided into a training set and a test set according to a set ratio.

[0090] Step 2: Construct a time-frequency feature extraction module;

[0091] A time-frequency feature extraction module is constructed using a shared convolutional neural network (CNN), and this module performs time-frequency co-learning when extracting features in both the time and frequency domains. As input, the temporal features of the source domain Temporal characteristics of the target domain Frequency domain characteristics of the source domain Frequency domain characteristics of the target domain As the extracted output;

[0092] Step 3: Construct the domain adaptive learning module;

[0093] Local Maximum Mean Discrepancy (LMMD) algorithms are introduced into the time and frequency domain feature spaces respectively to construct domain adaptation layers, and the LMMD loss in the time and frequency domains is utilized. Optimization yields a domain adaptive learning module that performs fine-grained intra-category distribution alignment for both time-domain and frequency-domain features; the domain adaptive learning module uses... As input, the temporal feature space of the source domain Temporal feature space of the target domain Frequency domain feature space of the source domain Frequency domain feature space of the target domain As output;

[0094] Step 4: Construct the domain classification module;

[0095] A domain classification discriminator network is constructed using multiple fully connected layers, and the domain classification loss L is utilized. adv The optimization process yields a domain classification module, which is a source-domain and frequency-domain fused feature vector obtained by concatenating time-domain and frequency-domain features along the feature dimension. As input, the probability of the source domain or the target domain is output;

[0096] Step 5: Train the time-frequency feature extraction module based on the unified domain adversarial training mechanism of time-frequency fusion features;

[0097] Based on L adv A gradient reversal layer (GRL) mechanism is introduced to conduct adversarial training between the time-frequency feature extraction module and the domain classification module, which prompts the time-frequency feature extraction module to learn the invariant features shared by the source domain and the target domain.

[0098] Step Six: Construct a dual-classifier collaborative decision-making module;

[0099] A dual-classifier collaborative decision-making module is constructed using a time-domain classifier and a frequency-domain classifier. For both the time-domain and frequency-domain classifiers, source-domain labeled samples are used to apply cross-entropy loss during time-domain classification. Cross-entropy loss in frequency domain classification Optimization is performed to obtain a well-trained dual-classifier collaborative decision-making module; thus, by combining the output results of the two classifiers through a dynamic fusion strategy during the inference stage, the complementarity of time-domain and frequency-domain features in fault category discrimination can be fully explored.

[0100] Step 7: Construct a cross-domain fault diagnosis model;

[0101] S71: A cross-domain fault diagnosis model is formed based on the time-frequency feature extraction module, the domain adaptive learning module, and the dual classifier collaborative decision-making module;

[0102] S72: Construct the total loss L;

[0103] S73: A cross-domain fault diagnosis model is trained using the training set. During training, supervised learning is performed on the time-domain and frequency-domain classifiers, respectively, based on source domain labels, while unsupervised learning is performed on the unlabeled target domain classifiers. A weighted fusion strategy is then used to integrate the classification decisions of the two modalities. The total classification loss L is jointly optimized. D LMMD total loss L L and countermeasures against losses L D This allows the model parameters to be iteratively optimized until the overall loss is minimized, resulting in a cross-domain fault diagnosis model with high robustness and generalization ability.

[0104] Step 8: Perform real-time diagnostics of cross-domain faults;

[0105] S81: Collect target data X online using a sensor array deployed in an industrial system, and normalize and standardize X.

[0106] S82: The data is sampled using a sliding window to obtain a sequence of data of consistent length, and then the obtained time-domain signal is subjected to Fourier transform (FFT) to obtain the frequency signal;

[0107] S83: The time-domain and frequency-domain signals of the unlabeled target domain are used as input data and fed into the cross-domain fault diagnosis model. In the cross-domain fault diagnosis model, the time-frequency feature extraction module is first used to extract high-dimensional features in the time-frequency domain. Then, the domain adaptation layer is used to perform LMMD measurement on the high-dimensional features in the time domain and frequency domain respectively to reduce the difference in feature distribution between the source domain and the target domain. Then, the dual classifier collaborative decision module is used for inference. In the inference stage, the output results of the time-domain classifier and the frequency-domain classifier are combined through a weighted fusion strategy to finally output the category prediction result.

[0108] As a preferred option, the process of collecting source domain data and target domain data in step one, and forming a sample dataset after preprocessing, is as follows:

[0109] S11: Acquire raw vibration signal data from the source region using a sensor array. and the original vibration signal data of the target domain in, and Data from the source domain and the target domain, respectively. For source domain tags, Let L be a real number, and L be the length of the time series for a single sample. s N tThese represent the number of samples in the source domain and the target domain, respectively.

[0110] S12: Perform data normalization and standardization processing on the source domain data and the target domain data;

[0111] S13: Use a sliding window to partition the source domain data, obtaining batch datasets (X) that serve as input data each time. N ,Y N ), where N represents the number of samples in a batch; the target domain data is processed in the same way as the source domain data; thus, the time domain signals of the source domain are obtained respectively. and the time-domain signal of the target domain

[0112] S14: Introducing frequency domain feature information, a Fast Fourier Transform is applied to each preprocessed sample to obtain the frequency domain signal from the source domain. Frequency domain signal of the target domain

[0113] S15: Obtain the time-domain signal from the source domain. Time-domain signal of the target domain Source domain frequency domain signal Frequency domain signal of the target domain As a sample dataset.

[0114] As a preferred option, in step S12 of step one, the source domain data and the target domain data are normalized according to formula (1), and the source domain data and the target domain data are standardized according to formula (2).

[0115]

[0116] In the formula, Z′ represents the data after normalization preprocessing; Z represents the original input data; min(Z) represents the minimum value of the input data; and max(Z) represents the maximum value of the input data.

[0117]

[0118] In the formula, Z″ represents the data after regularization preprocessing, μ represents the mean of the input data, and σ represents the standard deviation of the input data.

[0119] As a preferred option, in step S13 of step one, X is obtained according to formula (3). N Y is obtained according to formula (4) N ;

[0120]

[0121] As a preferred embodiment, in step S14 of step one, the frequency domain signal of the source domain is obtained according to formula (5). The frequency domain signal of the target domain is obtained according to formula (6).

[0122]

[0123] In the formula, These represent the time-domain signals from the source domain, respectively. These represent the time-domain signals of the target domain, respectively.

[0124] As a preferred embodiment, in step two, the temporal characteristic space of the source domain is obtained according to formula (7). The temporal feature space of the target domain is obtained according to formula (8). The frequency domain feature space of the source domain is obtained according to formula (9). The frequency domain feature space of the target domain is obtained according to formula (10).

[0125]

[0126] In the formula, F(·) represents the mapping function of the shared convolutional neural network.

[0127] As a preferred embodiment, in step three, the local maximum mean difference loss in the time domain and frequency domain is obtained according to formulas (11) and (12), respectively.

[0128]

[0129] In the formula, Φ(·) is the kernel mapping function (often using a Gaussian kernel), and C is the total number of categories. Let c be the set of time-domain feature samples of category c in the source domain. Let c be the set of time-domain feature samples of category c in the target domain. Let c be the set of frequency domain feature samples of category c in the source domain. Let c be the set of frequency domain feature samples of category c in the target domain.

[0130] As a preferred option, in step four, the source domain and frequency domain fused feature vectors are obtained according to formulas (13) and (14), respectively. Meanwhile, the domain classification loss L is obtained according to formula (15). adv ;

[0131]

[0132] In the formula, ";" indicates that the splicing operation is performed in the column direction;

[0133]

[0134] In the formula, For the output of the domain classifier, d i These are real domain tags.

[0135] As a preferred embodiment, in step six, the temporal classification loss is obtained according to formulas (16) and (17), respectively. Frequency domain classification loss

[0136]

[0137] In the formula, C T (·) and C F (·) represent the time-domain classifier and the frequency-domain classifier, respectively.

[0138] To ensure collaborative optimization among submodules and improve overall system performance, a reasonable total loss function needs to be designed, along with a corresponding joint training strategy. Specifically, as a preferred approach, the process of constructing the total loss L in step seven is as follows:

[0139] S72-1: Obtain the total classification loss L according to formula (18) C And make the parameters of the time-frequency feature extractor F(x) Parameters of the domain adaptive layer A(x) Temporal classifier classifier C T The parameter of (x) Frequency domain classifier C F The parameter of (x) It satisfies formula (19);

[0140]

[0141] In the formula, λ T , λ F These are the decision weights for the time-domain and frequency-domain classifiers, respectively, λ. T +λ F =1;

[0142]

[0143] In the formula, Temporal characteristics of the source domain Labels for the temporal features of the source domain. For the frequency characteristics of the source domain, F(·) is the label for the frequency features of the source domain, F(·) is the time-frequency domain feature extractor, A(·) is the domain adaptation layer, and C(·) is the classifier.

[0144] S72-2: Obtain the total LMMD loss L according to formula (20)L And make the parameters of the time-frequency feature extractor F(x) Parameters of the domain adaptive layer A(x) Satisfies formula (21);

[0145]

[0146] In the formula,

[0147] S72-3: Obtain the adversarial loss L according to formula (22) D And make the parameters of the time-frequency feature extractor F(x) Parameters of the domain adaptive layer A(x) Parameters of the domain classifier D(x) Satisfies formula (23);

[0148]

[0149] In the formula, D is the domain classifier, and d is the binary domain label, indicating whether the sample x∈S or x∈T;

[0150]

[0151] S72-4: Obtain the total loss L according to formula (24);

[0152]

[0153] In the formula, δ is the balancing hyperparameter of the domain adaptive learning module, and β is the balancing hyperparameter of the domain classification module;

[0154] S72-5: Adjust the relative weights of δ and β to satisfy δ = 1 - β, and transform formula (24) into formula (25);

[0155]

[0156] The technical solution of the present invention will be described in detail below with reference to the embodiments, and its feasibility will be verified.

[0157] Example:

[0158] The CWRU dataset is one of the most representative publicly available datasets in rolling bearing fault diagnosis research. The experimental setup used includes... Figure 6As shown, it mainly includes a motor, torque sensor, accelerometer, and power meter. Bearings are installed at the motor drive end and fan end, and vibration signals are recorded by the accelerometer. The bearing signal at the drive end is selected as the analysis object, with a sampling frequency of 12kHz. The CWRU dataset contains three types of faults (inner ring fault IF, outer ring fault OF, and rolling element fault BF), each type of fault includes three fault degrees (7, 14, 21 mils), for a total of 9 fault categories and 1 normal state (NA), as shown in Table 1. Different loads (0HP, 1HP, 2HP, 3HP) and speeds (1797, 1772, 1750, 1730rpm) constitute four domains (A, B, C, D), as listed in Table 1. In the table, NA represents a normal bearing; IF represents an internal fault; BF represents a ball fault; and OF represents an external fault.

[0159] Table 1: Detailed information about the CWRU dataset

[0160]

[0161]

[0162] 1) Data preprocessing

[0163] To improve model training stability and ensure fairness in comparative experiments, all vibration signals were first bandpass filtered (30Hz–6kHz) to remove low-frequency interference and high-frequency noise, and then normalized to ensure a uniform amplitude scale across different signals. Subsequently, the signals were sliced ​​into 1024-segment data blocks using a sliding window, while retaining the time-domain signal and calculating the frequency-domain amplitude spectrum using Fast Fourier Transform (FFT). The first 512 frequency points were retained to avoid image interference, forming the final multimodal sample input. Source domain samples were labeled for supervised training, while target domain samples were unlabeled for unsupervised adaptation and testing.

[0164] 2) Experimental parameter settings

[0165] To ensure the fairness and reproducibility of the experimental results, standardized network training parameters and settings were used in all experiments. This experiment used the PyTorch framework to build the model, and a progressively decaying learning rate function was used to train the network structure. Specific settings were as follows: initial learning rate of 0.00003, momentum of 0.9, learning rate decay step size of 0.5, and cross-entropy as the loss function. Each batch contained 220 samples, the number of training epochs was set to 50, and each experiment was repeated 10 times. The dataset was divided into training and test sets in an 8:2 ratio. Since domain adversarial and domain adaptation techniques act on different parts of the model—the former helps extract domain-invariant features by acting on the feature extractor—the importance of domain adversarial and domain adaptation techniques for model domain adaptation differs during training. In this case, maintaining the relative magnitudes of δ and β constant throughout the entire model training process is unreasonable. A more reasonable approach is to adjust the relative weights of δ and β during training. Therefore, δ = 1 - β. Through multiple experimental training adjustments, this invention determined a set of optimal parameter settings: α T =0.3 and β=0.3.

[0166] 2) Cross-domain fault diagnosis test based on TF-CADA

[0167] A schematic diagram of the TF-CADA model structure is shown below. Figure 2 As shown, it mainly consists of four parts: time-frequency feature extraction module, domain adaptation module, domain classification module, and dual classifier module. Figure 3 The accuracy results based on the TF-CADA model in the cross-domain fault diagnosis task on the CWRU dataset are presented. Figure 4 The confusion matrix of the TF-CADA model under task D→A shows that the diagonal elements are significantly concentrated, indicating that the model can achieve high-precision discrimination of various types of faults, especially maintaining good recognition ability for fine-grained damage categories, which reflects the effective transfer capability of the method.

[0168] Figure 5 The feature visualization results of task D→A are presented, showing clear feature distributions among different fault categories and obvious decision boundaries, further verifying the superior classification performance and robustness of the model for fault state identification. Therefore, this invention demonstrates high robustness for cross-domain fault diagnosis.

[0169] The simulation examples above demonstrate the effectiveness of this invention. Experimental results based on the CWRU dataset show that this method can achieve accurate and robust cross-domain fault diagnosis, providing important technical support and guidance for fault prediction and maintenance decisions in actual production.

[0170] This invention provides an unsupervised adaptive fault diagnosis method based on time-frequency domain collaborative feature alignment to address the degradation of cross-domain diagnostic performance caused by distribution misalignment between the source and target domains in complex industrial environments. First, raw vibration signals from the source and target domains are acquired and processed using standardization and Fast Fourier Transform (FFT) to obtain time-domain and frequency-domain data in a unified format. Second, a shared convolutional neural network (CNN) is used to extract multimodal high-order features from the time and frequency domains of the source and target domains, constructing a unified multimodal high-dimensional representation that effectively compresses redundancy and enhances time-frequency collaborative expression capabilities. Next, a Local Maximum Mean Difference (LMMD) algorithm is introduced, applying the LMMD criterion to both the time and frequency domain feature spaces. This aligns the time-domain and frequency-domain features with fine-grained class distribution within their respective intra-class subspaces, significantly reducing intra-modal mismatches between classes, minimizing cross-domain distribution differences, and improving cross-domain discrimination capabilities. Through deep collaboration and fine-grained distribution alignment of time-frequency domain information, the accuracy and stability of the diagnosis are significantly improved, making it suitable for intelligent monitoring of various rotating machinery equipment. Then, based on the fused time-frequency features, a unified domain discriminant network is constructed, and adversarial training is implemented using a gradient inversion layer (GRL) to enable the feature extractor to learn domain-indistinguishable shared feature representations. Furthermore, a dual-classifier structure is constructed to independently learn time-domain and frequency-domain features, and a weighted fusion strategy is used to achieve multimodal integrated decision-making. Finally, considering the complementarity of time-domain and frequency-domain information, an end-to-end trainable cross-domain diagnostic framework is constructed through four stages: multimodal feature extraction, intra-modal alignment, domain adversarial feature extraction, and multi-classifier fusion decision-making. Simultaneously, during training, supervised learning is performed on the time-domain and frequency-domain classifiers in conjunction with source domain labels, and the classification decisions of the two modalities are integrated through a weighted fusion strategy. By jointly optimizing the classification loss, LMMD loss, and adversarial loss, the model parameters are iteratively optimized until the overall loss is minimized, thereby obtaining a cross-domain fault diagnosis model with high robustness and generalization ability. This invention can fully exploit the complementarity of time-domain and frequency-domain features. Through multimodal feature extraction, fine-grained intra-class distribution alignment, domain adversarial training, and weighted fusion of dual classifiers, it constructs an end-to-end cross-domain diagnostic framework, which effectively improves the model's generalization performance and robustness in the target domain.

[0171] This method can accurately identify the fault status of complex industrial equipment under varying operating conditions, achieve efficient diagnostic operations in unlabeled target domains, and has good transfer and generalization capabilities and fault identification robustness. It is suitable for intelligent condition assessment and cross-scenario deployment of industrial equipment such as rotating machinery, and can provide strong support for real-time fault warning and maintenance decisions in industrial production processes. It effectively reduces the risk of equipment downtime, optimizes production efficiency, and promotes the intelligent development of industry.

Claims

1. An unsupervised adaptive fault diagnosis method based on time-frequency domain collaborative feature alignment, characterized in that, Includes the following steps: Step 1: Data sampling and preprocessing; Data from the source domain and the target domain are collected and preprocessed to form a sample dataset. The sample dataset is then divided into a training set and a test set according to a set ratio. Step 2: Construct a time-frequency feature extraction module; A time-frequency feature extraction module is constructed using a shared convolutional neural network, and the time-frequency feature extraction module performs time-frequency co-learning when extracting time-domain and frequency-domain features. Time-frequency feature extraction module As input, the temporal features of the source domain Temporal characteristics of the target domain Frequency domain characteristics of the source domain Frequency domain characteristics of the target domain As the extracted output; Step 3: Construct the domain adaptive learning module; Local maximum mean difference (MMD) algorithms are introduced into the time and frequency domain feature spaces respectively to construct domain adaptation layers, and the MMD loss in the time and frequency domains is utilized. Optimization yields a domain adaptive learning module that performs fine-grained intra-category distribution alignment for both time-domain and frequency-domain features; the domain adaptive learning module uses... As input, the temporal feature space of the source domain Temporal feature space of the target domain Frequency domain feature space of the source domain Frequency domain feature space of the target domain As output; Step 4: Construct the domain classification module; A domain classification discriminator network is constructed using multiple fully connected layers, and the domain classification loss L is utilized. adv The optimization process yields a domain classification module, which is a source-domain and frequency-domain fused feature vector obtained by concatenating time-domain and frequency-domain features along the feature dimension. As input, the probability of the source domain or the target domain is output; Step 5: Train the time-frequency feature extraction module based on the unified domain adversarial training mechanism of time-frequency fusion features; Based on L adv A gradient reversal mechanism is introduced to conduct adversarial training between the time-frequency feature extraction module and the domain classification module, which prompts the time-frequency feature extraction module to learn the invariant features shared by the source domain and the target domain. Step Six: Construct a dual-classifier collaborative decision-making module; A dual-classifier collaborative decision-making module is constructed using a time-domain classifier and a frequency-domain classifier. For both the time-domain and frequency-domain classifiers, source-domain labeled samples are used to apply cross-entropy loss during time-domain classification. Cross-entropy loss in frequency domain classification Optimize the module to obtain a trained dual-classifier collaborative decision-making module; Step 7: Construct a cross-domain fault diagnosis model; S71: A cross-domain fault diagnosis model is formed based on the time-frequency feature extraction module, the domain adaptive learning module, and the dual classifier collaborative decision-making module; S72: Construct the total loss L; S73: A cross-domain fault diagnosis model is trained using the training set. During training, supervised learning is performed on the time-domain and frequency-domain classifiers, respectively, based on source domain labels, while unsupervised learning is performed on the unlabeled target domain classifiers. A weighted fusion strategy is then used to integrate the classification decisions of the two modalities. The total classification loss L is jointly optimized. D LMMD total loss L L and countermeasures against losses L D This allows the model parameters to be iteratively optimized until the overall loss is minimized, resulting in a cross-domain fault diagnosis model with high robustness and generalization ability. Step 8: Perform real-time diagnostics of cross-domain faults; S81: Collect target data X online using a sensor array deployed in an industrial system, and normalize and standardize X. S82: The data is sampled using a sliding window to obtain a sequence of data of consistent length, and then the obtained time-domain signal is subjected to Fourier transform to obtain the frequency signal; S83: The time-domain and frequency-domain signals of the unlabeled target domain are used as input data and fed into the cross-domain fault diagnosis model. In the cross-domain fault diagnosis model, the time-frequency feature extraction module is first used to extract high-dimensional features in the time-frequency domain. Then, the domain adaptation layer is used to perform LMMD measurement on the high-dimensional features in the time domain and frequency domain respectively to reduce the difference in feature distribution between the source domain and the target domain. Then, the dual classifier collaborative decision module is used for inference. In the inference stage, the output results of the time-domain classifier and the frequency-domain classifier are combined through a weighted fusion strategy to finally output the category prediction result.

2. The unsupervised adaptive fault diagnosis method based on time-frequency domain collaborative feature alignment according to claim 1, characterized in that, In step one, the process of collecting source domain data and target domain data, and then preprocessing them to form a sample dataset is as follows: S11: Acquire raw vibration signal data from the source region using a sensor array. and the original vibration signal data of the target domain in, and Data from the source domain and the target domain, respectively. For source domain tags, Let L be a real number, and L be the length of the time series for a single sample. s N t These represent the number of samples in the source domain and the target domain, respectively. S12: Perform data normalization and standardization processing on the source domain data and the target domain data; S13: Use a sliding window to partition the source domain data, obtaining batch datasets (X) that serve as input data each time. N ,Y N ), where N represents the number of samples in a batch; the target domain data is processed in the same way as the source domain data; thus, the time domain signals of the source domain are obtained respectively. and the time-domain signal of the target domain S14: Introducing frequency domain feature information, a Fast Fourier Transform is applied to each preprocessed sample to obtain the frequency domain signal from the source domain. Frequency domain signal of the target domain S15: Obtain the time-domain signal from the source domain. Time-domain signal of the target domain Source domain frequency domain signal Frequency domain signal of the target domain As a sample dataset.

3. The unsupervised adaptive fault diagnosis method based on time-frequency domain collaborative feature alignment according to claim 2, characterized in that, In step S12 of step one, the source domain data and target domain data are normalized according to formula (1), and the source domain data and target domain data are standardized according to formula (2). In the formula, Z′ represents the data after normalization preprocessing; Z represents the original input data; min(Z) represents the minimum value of the input data; and max(Z) represents the maximum value of the input data. In the formula, Z″ represents the data after regularization preprocessing, μ represents the mean of the input data, and σ represents the standard deviation of the input data.

4. The unsupervised adaptive fault diagnosis method based on time-frequency domain collaborative feature alignment according to claim 3, characterized in that, In step S13 of step one, X is obtained according to formula (3). N Y is obtained according to formula (4) N ; 5. The unsupervised adaptive fault diagnosis method based on time-frequency domain collaborative feature alignment according to claim 4, characterized in that, In step S14 of step one, the frequency domain signal of the source domain is obtained according to formula (5). The frequency domain signal of the target domain is obtained according to formula (6). In the formula, These represent the time-domain signals from the source domain, respectively. These represent the time-domain signals of the target domain, respectively.

6. The unsupervised adaptive fault diagnosis method based on time-frequency domain collaborative feature alignment according to claim 5, characterized in that, In step two, the temporal feature space of the source domain is obtained according to formula (7). The temporal feature space of the target domain is obtained according to formula (8). The frequency domain feature space of the source domain is obtained according to formula (9). The frequency domain feature space of the target domain is obtained according to formula (10). In the formula, F(·) represents the mapping function of the shared convolutional neural network.

7. The unsupervised adaptive fault diagnosis method based on time-frequency domain collaborative feature alignment according to claim 6, characterized in that, In step three, the local maximum mean difference loss in the time domain and frequency domain is obtained according to formulas (11) and (12), respectively. In the formula, Φ(·) is the kernel mapping function, and C is the total number of categories. Let c be the set of time-domain feature samples of category c in the source domain. Let c be the set of time-domain feature samples of category c in the target domain. Let c be the set of frequency domain feature samples of category c in the source domain. Let c be the set of frequency domain feature samples of category c in the target domain.

8. The unsupervised adaptive fault diagnosis method based on time-frequency domain collaborative feature alignment according to claim 7, characterized in that, In step four, the source domain and frequency domain fused feature vectors are obtained according to formulas (13) and (14), respectively. Meanwhile, the domain classification loss L is obtained according to formula (15). adv ; In the formula, ";" indicates that the splicing operation is performed in the column direction; In the formula, For the output of the domain classifier, d i These are real domain tags.

9. The unsupervised adaptive fault diagnosis method based on time-frequency domain collaborative feature alignment according to claim 8, characterized in that, In step six, the temporal classification loss is obtained according to formulas (16) and (17), respectively. Frequency domain classification loss In the formula, C T (·) and C F (·) represent the time-domain classifier and the frequency-domain classifier, respectively.

10. The unsupervised adaptive fault diagnosis method based on time-frequency domain collaborative feature alignment according to claim 9, characterized in that, In step seven, the process of constructing the total loss L is as follows: S72-1: Obtain the total classification loss L according to formula (18) C And make the parameters of the time-frequency feature extractor F(x) Parameters of the domain adaptive layer A(x) Temporal classifier classifier C T The parameter of (x) Frequency domain classifier C F The parameter of (x) It satisfies formula (19); In the formula, λ T , λ F These are the decision weights for the time-domain and frequency-domain classifiers, respectively, λ. T +λ F =1; In the formula, For the temporal characteristics of the source domain, Labels for the temporal features of the source domain. For the frequency characteristics of the source domain, F(·) is the label for the frequency features of the source domain, F(·) is the time-frequency domain feature extractor, A(·) is the domain adaptation layer, and C(·) is the classifier. S72-2: Obtain the total LMMD loss L according to formula (20) L And make the parameters of the time-frequency feature extractor F(x) Parameters of the domain adaptive layer A(x) Satisfies formula (21); In the formula, S72-3: Obtain the adversarial loss L according to formula (22) D And make the parameters of the time-frequency feature extractor F(x) Parameters of the domain adaptive layer A(x) Parameters of the domain classifier D(x) Satisfies formula (23); In the formula, D is the domain classifier, and d is the binary domain label, indicating whether the sample x∈S or x∈T; S72-4: Obtain the total loss L according to formula (24); In the formula, δ is the balancing hyperparameter of the domain adaptive learning module, and β is the balancing hyperparameter of the domain classification module; S72-5: Adjust the relative weights of δ and β to satisfy δ = 1 - β, and transform formula (24) into formula (25);

Citation Information

Cited By

  • Unsupervised cross-working-condition fault diagnosis method based on edge guidance

    CN121614842A