Semi-supervised enhanced intelligent fault diagnosis method for complex equipment assisted by digital multi-twins

Abstract

The present invention relates to the field of fault diagnosis technology, and discloses a digital multi-twin assisted semi-supervised enhanced complex equipment intelligent fault diagnosis method, comprising the following steps: mechanism simulation twin and deep data twin self-supervised data enhancement; lightweight multi-scale attention mechanism diagnostic network design; based on a semi-supervised three-stage diagnostic adjustment mechanism, the digital multi-twin assisted semi-supervised enhanced complex equipment intelligent fault diagnosis method introduces digital multi-twin technology on the basis of existing fault diagnosis methods, and combines deep data twin self-supervised data enhancement to significantly improve the robustness and accuracy of the model; at the same time, a lightweight multi-scale attention mechanism is added to the traditional diagnostic network structure to enhance the network's ability to extract multiple fault features and optimize the diagnostic accuracy.

Description

Technical Field

[0001] The present invention relates to the field of fault diagnosis technology, and specifically to a digital multi-twin assisted semi-supervised enhanced intelligent fault diagnosis method for complex equipment. Background Art

[0002] In today's rapidly evolving Industry 4.0 era, intelligent manufacturing, as a core component, is reshaping the global manufacturing landscape. Intelligent manufacturing leverages the deep integration of advanced information technology and manufacturing techniques to drive optimization and innovation in production processes through data-driven decision support. The widespread adoption of these technologies has made equipment status monitoring and fault prediction increasingly important.

[0003] In modern industrial environments, the effectiveness of fault diagnosis and failure analysis is directly related to equipment reliability, production efficiency, and economic benefits. Therefore, timely and accurate fault diagnosis is crucial for ensuring production safety and stable equipment operation. However, traditional fault diagnosis methods often rely on large amounts of labeled data, which is often difficult to implement in practical applications. Furthermore, traditional methods often exhibit low robustness when faced with complex faults and dynamic systems, making them unable to adapt to the ever-changing industrial environment. Against this backdrop, semi-supervised learning, as an emerging technology, demonstrates great potential. It can effectively learn from limited labeled data and provide new solutions for fault diagnosis.

[0004] Semi-supervised learning is a machine learning method that lies between supervised and unsupervised learning, aiming to fully utilize limited labeled data and abundant unlabeled data. The basic concept of this method is to construct a model that uses labeled samples to guide the learning process of unlabeled samples, thereby improving the model's generalization ability and learning efficiency. Current research methods have shown significant advantages in solving fault diagnosis problems with limited labeled data. Currently, a semi-supervised contrastive learning (SSCL) method based on a multi-scale attention (MSA) mechanism and multi-objective contrastive learning (MCL) has been developed. By integrating limited labeled data with a large amount of unlabeled data, it can jointly learn effective feature mappings, thereby achieving excellent diagnostic performance. The semi-supervised matrixed graph embedding machine (SMGEM) effectively alleviates the performance bottleneck caused by the scarcity of labeled samples by embedding unlabeled and labeled samples based on geometric similarity. Furthermore, a semi-supervised diagnosis method that combines a feature pre-extraction mechanism with an improved generative adversarial network (IGAN) further enhances the model's ability to utilize unlabeled data. Pseudo-label-assisted semi-supervised adversarial reinforcement learning methods achieve efficient and accurate diagnostic results in gearbox fault diagnosis by reducing the reliance on labeled data. Together, these methods provide an effective solution to the data scarcity problem under complex working conditions. The above research shows that semi-supervised learning can mine potential fault modes in a large amount of unlabeled operating data, thereby achieving more accurate fault prediction and diagnosis. However, how to effectively combine semi-supervised learning with other advanced technologies to further improve the effect of fault feature mining remains a challenge in current research.

[0005] The rapid development of data-driven approaches has driven the diversification of digital twin models, providing more possibilities for combining digital twin methods with other approaches. Furthermore, digital twin models are no longer limited to traditional simulation models but are increasingly incorporating deep learning models. Currently, a visual digital twin model has been developed that achieves image enhancement and improves the diagnostic accuracy of photovoltaic systems. Another deep learning time-series digital twin model has achieved accurate wind power prediction. Furthermore, deep digital twin methods based on adversarial training can achieve a balance of feature samples and improve fault diagnosis accuracy. A pseudo-label-assisted semi-supervised adversarial reinforcement learning method for gearbox fault diagnosis is proposed. This method effectively utilizes unlabeled data, reduces reliance on labeled samples, and improves diagnostic accuracy. The introduction of digital twin technology provides semi-supervised learning with rich fault feature information and data support, potentially enhancing the performance of semi-supervised fault diagnosis models. However, while deep learning-based digital twin models have achieved some success in fault diagnosis tasks, data-driven twin models still face challenges in terms of effectiveness and accuracy. To address this, we propose a semi-supervised enhanced intelligent fault diagnosis method for complex equipment assisted by digital multi-twins. Summary of the Invention

[0006] The purpose of the present invention is to provide a digital multi-twin assisted semi-supervised enhanced intelligent fault diagnosis method for complex equipment to solve the problems raised in the above background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a digital multi-twin assisted semi-supervised enhanced complex equipment intelligent fault diagnosis method, comprising the following steps:

[0008] S1. Mechanism simulation twin and deep data twin self-supervision data enhancement stage:

[0009] S11. In MATLAB Simscape TM A three-cylinder pump mechanism simulation model was built on the platform, including the parametric design of the plunger, seals, one-way valve, and bearings, simulating normal conditions and the following failure modes;

[0010] S12. During the simulation process, the flow and pressure data of steady-state operation are collected, and 1 second of data is recorded for each state to form a diverse sample data set;

[0011] S13. Introducing the deep data twin method, achieving data enhancement through generative adversarial networks (GANs), and constructing pseudo labels;

[0012] S14. Define the generator and discriminator loss functions:

[0013]

[0014] S15. To ensure the diversity and consistency of generated samples, define the Pearson similarity constraint:

[0015]

[0016] And construct the self-supervised data enhancement loss function:

[0017]

[0018] S16. Comprehensively optimize the objective functions of the generator, discriminator, and data augmentation modules:

[0019]

[0020] Among them, λ and β are weight hyperparameters, and the Adam algorithm is used to dynamically adjust the learning rate and update the parameters;

[0021] S2. Lightweight Multi-scale Attention Mechanism Diagnosis Network Design Phase: Construct a Lightweight Multi-scale Attention Mechanism Diagnosis Network (LMAN-SSFD), including a multi-scale convolution module and a channel attention module, to achieve feature extraction and fusion of input data;

[0022] The multi-scale convolution module uses convolution kernels of different sizes to extract details, local and global features, and outputs multi-scale feature representations:

[0023] The channel attention module optimizes feature representation by weight distribution and defines weighted fusion features:

[0024] f att (x) = f LMAN (x)⊙A(f LMAN (x));

[0025] Among them, A(f LMAN (x)) represents the feature weights distributed by the height and width direction attention weights;

[0026] S3, a three-stage diagnostic adjustment mechanism based on semi-supervision; by combining labeled and unlabeled data, a semi-supervised learning framework is constructed to enhance the robustness and diagnostic accuracy of the model.

[0027] Optionally, the S11 further includes:

[0028] S111, three single faults include: leakage (Leak P1), inlet blockage (Block P1) and bearing wear (Bearing Wear);

[0029] S112. Three combination faults include: leakage + inlet blockage (Leak P1+Block P1), leakage + bearing wear (Leak P1+Bearing Wear), and inlet blockage + bearing wear (Block P1+Bearing Wear).

[0030] Optionally, the S3 includes:

[0031] S31. Phase 1: Initial model training and pseudo-label generation; using labeled datasets Train the semi-supervised diagnosis network and build a basic model. Use this model to analyze the unlabeled dataset. Generate pseudo labels to form a pseudo label dataset Through the test set D test Evaluate the model and get the initial accuracy P initial , providing a reference for subsequent optimization;

[0032] S32, the second stage: DDT model generates enhanced samples; according to the labeled dataset D l Train the Deep Data Twin (DDT) model to generate an enhanced sample dataset Combine the generated samples with the labeled data into an extended dataset D combined =D l ∪Dgenerated ; The extended dataset is used to further train the model and pass the test set D test Get the second-stage accuracy P combined ;

[0033] S33, the third stage: final model optimization; using the model optimized in the second stage, the unlabeled dataset D u Generate pseudo labels again and build the final dataset The final dataset is used for full data training to optimize the model performance and the test set D test Get the final accuracy P final .

[0034] Optionally, the S1 includes:

[0035] Generate samples using a generative adversarial network. The optimization objectives of the generator and discriminator include authenticity, diversity, and consistency of the generated samples. Pearson similarity constraints are introduced to ensure the similarity between the generated samples and the real samples in the feature space while avoiding mode collapse. The comprehensive objective function is jointly optimized through adversarial training and self-supervised data enhancement.

[0036] Optionally, the S2 includes:

[0037] A combination of multi-scale convolution kernels is used to extract features of different scales. The channel attention module combines feature weighted allocation and fusion to enhance the ability to extract key features. Its core formula is:

[0038] f att (x) = f LMAN (x)⊙A(f LMAN (x)).

[0039] Optionally, the S13 further includes: a generator (DDT-G) generates synthetic samples using a random noise vector as input, and a discriminator (DDT-D) distinguishes between real samples and generated samples.

[0040] Compared with the existing technology, the present invention provides a digital multi-twin assisted semi-supervised enhanced intelligent fault diagnosis method for complex equipment, which has the following beneficial effects:

[0041] 1. Based on the existing fault diagnosis method, this invention introduces digital multi-twin technology, combined with deep data twin self-supervised data enhancement, which significantly improves the robustness and accuracy of the model; at the same time, it adds a lightweight multi-scale attention mechanism to the traditional diagnostic network structure, enhances the network's ability to extract multiple fault features, and optimizes the diagnostic accuracy.

[0042] 2. This invention effectively combines labeled and unlabeled data through a three-stage adjustment mechanism based on semi-supervised learning, improving the diagnostic system's generalization capabilities under low-label data conditions. Furthermore, the self-supervised data augmentation method employed can precisely optimize the consistency and diversity of generated samples within the generative adversarial network framework, improving model performance even in data-scarce conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a schematic diagram of the twinning and sampling simulation mechanism of the present invention;

[0044] Figure 2 This is a schematic diagram of seven operating state samplings of the present invention;

[0045] Figure 3 Schematic diagram of the self-supervision enhancement effect of DDT data in the present invention;

[0046] Figure 4 Schematic diagram of the FFT comparison results of labeled samples and pseudo samples of the present invention

[0047] Figure 5 Schematic diagram of the improvement effect of the three-stage semi-supervised diagnosis accuracy of the present invention

[0048] Figure 6 Schematic diagram showing the t-SNE results of the five states of the present invention;

[0049] Figure 7 Schematic diagram showing the confusion matrix results of the five states of the present invention. DETAILED DESCRIPTION

[0050] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0051] like Figure 1-Figure 7 As shown, the present invention provides a technical solution:

[0052] A digital multi-twin-assisted semi-supervised enhanced intelligent fault diagnosis method for complex equipment combines physical modeling and data enhancement technology to build a reliable fault diagnosis data foundation for low-label situations. The method includes the following steps:

[0053] S1, mechanism simulation twin and deep data twin self-supervision data enhancement stage;

[0054] Simscape based on MATLAB TM, build a mechanism simulation twin model of the three-cylinder pump, combined with Simulink Design Optimization TM Optimize parameters. Through pressure and flow sensors, collect flow and pressure data of plunger 1 (P1) in different states, such as Figure 1 The model supports modeling three typical failure modes: Leakage (Leak P1), modeled through plunger seal leakage; Blockage (Block P1), modeled by reducing the maximum area of ​​the Check Valve Inlet; and Bearing Wear, modeled by increasing friction torque. In addition to the three classic single fault states, three combined fault states are also designed: "Leak P1 + Block P1," "Leak P1 + Bearing Wear," and "Block P1 + Bearing Wear," for a total of six model fault states.

[0055] Each operation of the mechanism simulation twin requires a startup time of 0.5s to achieve stable operation. Therefore, in the operation of one normal state and six fault states in the present invention, the dual sensors collect data waveforms with a length of 0.7s-1.7s totaling 1s to avoid the impact of the startup state.

[0056] Deep data twinning and self-supervised data enhancement: To improve the performance of the model under limited labeled data, this paper designs a deep data twinning method. Through self-supervised data enhancement technology, it can effectively learn features from the existing small amount of labeled data and generate new synthetic samples to enrich the training set, realizing the modeling of working conditions through pressure and flow data. This not only increases the diversity of data but also improves the generalization ability of the model. The specific implementation process is as follows:

[0057] The adversarial generation process consists of a generator (DDT-G) and a discriminator (DDT-D), which are implemented through adversarial training. The goal of the generator is to generate samples that look as real as possible, while the goal of the discriminator is to distinguish between real samples and generated samples. The objective function L of DDT-G is G It can be expressed as:

[0058]

[0059] Among them L G is the loss function of the generator, z is a random noise vector drawn from the distribution p z (z); the objective function of DDT-D is:

[0060]

[0061] Among them L Dis the loss function of the discriminator, x is the real sample, from the data distribution p data (x). In self-supervised learning, new samples are generated by enhancement techniques. It should have sufficient diversity while retaining the original sample information. To this end, Pearson similarity is introduced to evaluate the quality and diversity of enhanced samples. Its expression is shown in formula (3):

[0062]

[0063] in is the evaluation sample x and the enhanced sample Pearson similarity of the correlation between is the mean of sample x, σ x is the standard deviation of sample x, and the mean and standard deviation of the enhanced sample are respectively expressed as and To ensure enhanced sample diversity, the present invention requires Pearson similarity The self-supervised loss function is defined as follows:

[0064]

[0065] is the self-supervised data enhancement loss function, N is the number of samples, λ r is a weight hyperparameter used to balance the contribution of each part, is an indicator function that returns 1 when the condition is met and 0 otherwise. τ is the threshold that controls the balance between diversity and consistency and is set empirically.

[0066] Combining generative adversarial networks with self-supervised data enhancement, the comprehensive objective function L is shown in formula (5):

[0067]

[0068] Here λ and β are weight hyperparameters. In the optimization process, the Adam optimization algorithm is used to minimize the comprehensive objective function. The first-order moment estimation and the second-order moment estimation are used to dynamically adjust the learning rate. First, the first-order moment m is initialized. t and the second moment v t , and its update process is shown in formulas (6) and (7):

[0069]

[0070]

[0071] β1, β2 are the decay rates of the first and second order moments. The deviation correction is shown in formulas (8) and (9):

[0072]

[0073]

[0074] The updated parameter θ is shown in formula (10).

[0075]

[0076] Where θ is the parameter to be optimized, η is the learning rate, and ∈ is a small constant to prevent division by zero. During the training process, the generator and the discriminator are updated alternately. The specific steps include calculating the generator loss L G and the discriminator loss L D , while calculating the self-supervised loss Generator parameters θ G , discriminator parameters θ D and the enhanced policy parameter θ aug The update formula is:

[0077]

[0078]

[0079]

[0080] The adversarial process and self-supervised data augmentation, and the introduction of Pearson similarity to control the quality and diversity of enhanced samples, can improve the performance and generalization ability of the model when processing limited labeled data.

[0081] The network structure of DDT mainly consists of two parts: a generator (DDT-G) and a discriminator (DDT-D). The network composition parameters are shown in the following table:

[0082]

[0083] In step S2, the lightweight multi-scale attention mechanism diagnosis network design phase involves designing and building the LMAN-SSFD network. Assuming that self-supervised sample generation was achieved in the previous step, a lightweight multi-scale attention mechanism semi-supervised fault diagnosis network (LMAN-SSFD) is designed to further process the samples.

[0084] In this framework, assume that the input sample is x and its label is y. The feature extraction part uses multi-scale convolution to output feature representation f LMAN (x) is as follows:

[0085]

[0086] Among them, K is the number of convolution kernels of different scales, CNN k (x) is the convolution operation of the kth scale. The multi-scale multi-channel attention mechanism EMA module

[24] is introduced into the diagnosis network to obtain the weighted feature representation f att (x):

[0087] f att (x) = f LMAN (x)⊙A(f LMAN (x)) (15)

[0088] Among them, the output A(f LMAN (x)) can be expressed as:

[0089] A(f LMAN (x))=σ(W h ⊙AvgPool(f LMAN (x))+W w ⊙AvgPool(f LMAN (x) T )) (16)

[0090] Here, W h is the height weight and W w is the width weight, σ is the Sigmoid activation function. The input feature x is grouped.

[0091] group x =reshape(x,b·g,-1,h,w) (17)

[0092] Where b is the number of samples in a single batch and g is the number of groups. Calculate the average pooling of height and width x h 、x w ;

[0093] x h =AvgPool h (group x ) (18)

[0094] x w =AvgPool w (group x ) (19)

[0095] Generate features through 1x1 convolution and 3x3 convolution;

[0096]

[0097] x1=GroupNorm(group x ·σ(hw)) (21)

[0098] x2=Conv3x3(group x ) (twenty two)

[0099] Calculate weights and generate weighted output A(f LMAN (x)) is as follows:

[0100] weights=Softmax(x1)·Softmax(x2) (23)

[0101] A(f LMAN (x))=group x ·weights (24)

[0102] There is a label loss L sup (f att (x l ),y l ) is shown in formula (25):

[0103]

[0104] where N l is the number of labeled data samples. This loss function is used to measure the model's performance on labeled data x. l The predictive power of y l,i is the true label of sample i, f att (x l ) i is the model for sample x l By minimizing this loss, the model will learn a more accurate label mapping.

[0105] Unlabeled loss As shown in formula (26):

[0106]

[0107] This loss function uses the unlabeled data x u and the pseudo labels generated by them Used to enhance the learning ability of the model. Minimizing the loss function helps the model utilize the information of unlabeled data, thereby improving the overall classification performance. In semi-supervised classification, constructing a comprehensive loss function L LMAN .

[0108]

[0109] Here, λ is a balancing parameter that controls the relative importance of labeled and unlabeled losses. By combining these two losses, the model can fully exploit the potential of unlabeled data while maintaining the accuracy of labeled data. This strategy is particularly effective when labels are scarce and can significantly improve the model's generalization and robustness. The Adam algorithm is used for optimization.

[0110] The LMAN-SSFD network effectively combines multi-scale feature extraction and EMA modules, thereby improving the model's learning ability under limited labeled data and its robustness and accuracy in practical applications. The network structure parameters of LMAN-SSFD are shown in the following table:

[0111]

[0112]

[0113] S3, a three-stage diagnostic adjustment mechanism based on semi-supervision;

[0114] In the step, a three-stage diagnosis adjustment mechanism under a semi-supervised learning framework is introduced, which aims to enhance the robustness and accuracy of the model in fault diagnosis by combining labeled and unlabeled data.

[0115] First, in the first stage, based on the labeled dataset Train the semi-supervised diagnosis network model and use the trained model to analyze the unlabeled dataset. Divide the data to obtain pseudo labels, and further train the model based on the divided data to build a basic semi-supervised strategy. Get the initial accuracy P initial .

[0116] In the second stage, based on the labeled dataset Train the DDT model to generate unlabeled datasets And design pseudo labels according to their types. The main goal of this stage is to improve the diversity of the model through features while maintaining the consistency of the model. combined =D l ∪D generated , using the model to D combined Training, through D test The two-stage accuracy P of DDT self-supervision enhancement is obtained combined .

[0117] In the third stage, the unlabeled dataset is analyzed based on the fault diagnosis model trained in the second stage. Divide to get pseudo labels, and use D final =Dl ∪D generated ∪D u , through D test Get the final accuracy P final , achieving a complete three-stage semi-supervised strategy.

[0118] The above generally describes the present invention in detail. However, it is obvious to those skilled in the art that modifications or improvements may be made based on the present invention. Therefore, modifications or improvements that do not depart from the spirit of the present invention are within the scope of protection of the present invention.

Claims

1. A digital multi-twin assisted semi-supervised enhanced intelligent fault diagnosis method for complex equipment, characterized by: The steps include: S1. Mechanism simulation twin and deep data twin self-supervision data enhancement stage: S11. In MATLAB Simscape TM A three-cylinder pump mechanism simulation model was built on the platform, including the parametric design of the plunger, seals, one-way valve, and bearings, simulating normal conditions and the following failure modes; S12. During the simulation process, the flow and pressure data of steady-state operation are collected, and 1 second of data is recorded for each state to form a diverse sample data set; S13. Introducing the deep data twin method, achieving data enhancement through generative adversarial networks (GANs), and constructing pseudo labels; S14. Define the generator and discriminator loss functions: S15. To ensure the diversity and consistency of generated samples, define the Pearson similarity constraint: And construct the self-supervised data enhancement loss function: S16. Comprehensively optimize the objective functions of the generator, discriminator, and data augmentation modules: Among them, λ and β are weight hyperparameters, and the Adam algorithm is used to dynamically adjust the learning rate and update the parameters; S2. Lightweight multi-scale attention mechanism diagnosis network design phase: Construct a lightweight multi-scale attention mechanism diagnosis network, including a multi-scale convolution module and a channel attention module, to achieve feature extraction and fusion of input data; The multi-scale convolution module uses convolution kernels of different sizes to extract details, local and global features, and outputs multi-scale feature representations: The channel attention module optimizes feature representation by weight distribution and defines weighted fusion features: f att (x)=f LMAN (x)⊙A(f LMAN (x)); Among them, A(f LMAN (x)) represents the feature weights distributed by the height and width direction attention weights; S3, a three-stage diagnostic adjustment mechanism based on semi-supervision; by combining labeled and unlabeled data, a semi-supervised learning framework is constructed to enhance the robustness and diagnostic accuracy of the model.

2. The digital multi-twin assisted semi-supervised enhanced intelligent fault diagnosis method for complex equipment according to claim 1 is characterized by: Said S11 further comprises: S111. Three single faults include: leakage, inlet blockage and bearing wear; S112. The three combined faults include: leakage + inlet blockage, leakage + bearing wear, and inlet blockage + bearing wear.

3. The digital multi-twin assisted semi-supervised enhanced intelligent fault diagnosis method for complex equipment according to claim 1 is characterized by: The S3 includes: S31. Phase 1: Initial model training and pseudo-label generation; using labeled datasets Train the semi-supervised diagnosis network and build a basic model; use this model to analyze the unlabeled dataset Generate pseudo labels to form a pseudo label dataset Through the test set D test Evaluate the model and get the initial accuracy P initial , providing a reference for subsequent optimization; S32, the second stage: DDT model generates enhanced samples; according to the labeled dataset D l Train the deep data twin model to generate enhanced sample datasets Combine the generated samples with the labeled data into an extended dataset D combined =D l ∪D generated ; The extended dataset is used to further train the model and pass the test set D test Get the second-stage accuracy P combined ; S33, the third stage: final model optimization; using the model optimized in the second stage, the unlabeled dataset D u Generate pseudo labels again and build the final dataset The final dataset is used for full data training to optimize the model performance and the test set D test Get the final accuracy P final .

4. The digital multi-twin assisted semi-supervised enhanced intelligent fault diagnosis method for complex equipment according to claim 1 is characterized by: Said S1 comprises: Generate samples using a generative adversarial network. The optimization objectives of the generator and discriminator include authenticity, diversity, and consistency of the generated samples. Pearson similarity constraints are introduced to ensure the similarity between the generated samples and the real samples in the feature space while avoiding mode collapse. The comprehensive objective function is jointly optimized through adversarial training and self-supervised data enhancement.

5. The digital multi-twin assisted semi-supervised enhanced intelligent fault diagnosis method for complex equipment according to claim 1 is characterized by: The S2 includes: A combination of multi-scale convolution kernels is used to extract features of different scales. The channel attention module combines feature weighted allocation and fusion to enhance the ability to extract key features. Its core formula is: f att (x)=f LMAN (x)⊙A(f LMAN (x))。 6. The digital multi-twin assisted semi-supervised enhanced intelligent fault diagnosis method for complex equipment according to claim 1 is characterized by: The S13 further includes: a generator generates a synthetic sample using a random noise vector as input, and a discriminator distinguishes between real samples and generated samples.

Images