Intelligent rolling bearing fault diagnosis method based on digital twinning and partial domain self-adaption

Abstract

The invention relates to a rolling bearing intelligent fault diagnosis method based on digital twinning and partial domain self-adaption, and the method comprises the following steps: building a bearing high-fidelity digital twinning model based on a geometric model, a physical model, a behavior model and a rule model according to the specific model of a target bearing; based on the constructed digital twinborn model, vibration simulation responses of the bearing under different preset faults are obtained; under the driving of an attention mechanism, in combination with a residual convolutional network of an improved CBAM module, domain discrimination features are extracted from the time-frequency image converted from the vibration simulation signal; based on a weight mechanism of a two-stage domain discriminator, adaptively measuring contribution of each source sample, and giving a lower weight to an outlier sample; a confrontation transfer strategy is introduced to reduce the distribution difference between different domains, and effective migration of fault diagnosis knowledge from a virtual space to a physical space is realized. According to the method, the problems of fault sample insufficiency and data label missing are effectively solved through a digital twin technology, and the problem of low fault diagnosis precision caused by poor generalization ability of a digital twin model and inconsistency of data distribution and data labels is solved through partial domain self-adaption.

Description

Technical Field

[0001] The present invention relates to the field of bearing fault diagnosis, and in particular to an intelligent fault diagnosis method for rolling bearings based on digital twins and partial domain adaptation. Background Art

[0002] Bearings are one of the key components of rotating machinery and are widely used in complex mechanical equipment. The operating status of bearings has a significant impact on the mechanical system. Once a bearing fails, it will disrupt production plans and reduce production efficiency. Severe failures may cause the collapse of the entire mechanical system, causing economic losses or even casualties. Therefore, effective fault diagnosis of bearings during operation is crucial to prevent serious accidents and reduce operation and maintenance costs.

[0003] In recent years, deep learning technology has been continuously developed and successfully applied to the field of bearing fault diagnosis. This technology can automatically evaluate the operating status of mechanical equipment based on the original monitoring data. Although deep learning methods can achieve good results in fault diagnosis, they mainly face two challenges: in order to obtain an accurate fault diagnosis model, a large amount of labeled data is required, which is often difficult to achieve in industrial applications. It is usually impractical to obtain sufficient fault data in various application scenarios; test data and training data need to meet the assumption that they come from the same data distribution, but this situation is not very satisfied in many industrial applications. The inconsistency between the distribution of training data and test data will lead to a significant decrease in diagnostic performance.

[0004] With the development of technologies such as industrial Internet of Things, virtual reality, and big data, the concept of digital twins has received widespread attention in various fields of intelligent manufacturing. Digital twins are considered to be an effective tool for realizing the interactive integration of the virtual world and the physical world. High-fidelity digital models of physical entities can generate simulation data that is close to the actual system performance. Therefore, digital twins are an emerging technology to solve the scarcity of labeled data in actual industrial scenarios, and have gained rapid development momentum in the field of mechanical health management research. By establishing a digital twin model of the bearing, simulation data of various faults that are consistent with the actual working conditions of the bearing can be obtained, effectively making up for the challenges of insufficient fault data and missing data labels.

[0005] Although the digital twin model simulates the operating status of the actual equipment as much as possible, there is still a large difference between the simulated data and the measured data, resulting in unsatisfactory diagnostic results. In order to solve the problem of data distribution differences, various domain adaptation technologies based on transfer learning are used for bearing fault diagnosis. Domain adaptation technology can transfer diagnostic knowledge learned from the supervised source domain to the unsupervised target domain by learning invariant feature representations. The role of domain adaptation learning is to derive invariant feature representations across two domains.

[0006] Although existing transfer learning research has achieved success in mechanical fault diagnosis, these studies largely rely on the source domain and the target domain having the same label space, that is, the fault categories of the two domains are consistent. However, this assumption is usually not true in industrial scenarios. For an unlabeled target domain, some fault types that exist in the source domain may never appear in the target domain. If the source and target domain data with different label spaces are directly matched, negative transfer may occur, which may reduce the diagnostic accuracy. In order to solve the problem of inconsistent data labels, partial domain adaptation methods have been proposed. Partial domain adaptation means that the label space of the target domain is a subset of the source domain. Classes that exist in both domains are shared classes, while classes that only exist in the source domain are outlier classes. Outlier samples refer to source samples whose labels do not exist in the target domain. Since the target samples are unlabeled, it is impossible to directly find outlier source samples, and the existence of outlier source samples will lead to a significant reduction in domain adaptation performance. Therefore, the key to partial domain adaptation is to extract domain-invariant features for knowledge transfer and suppress outlier samples during the transfer process to prevent negative transfer. Summary of the invention

[0007] In response to the above problems, the present invention aims to solve the problems of insufficient fault samples and missing data labels when training diagnostic models, poor generalization ability of digital twin models when producing fault samples, and low fault diagnosis accuracy caused by inconsistency between data distribution and data labels during fault diagnosis.

[0008] In order to achieve the above-mentioned purpose, the present invention proposes an intelligent fault diagnosis method for rolling bearings based on digital twins and partial domain adaptation, so as to perform fault diagnosis on bearings more accurately and efficiently. The implementation process of the method is as follows: according to the specific model of the target bearing, a high-fidelity digital twin model of the bearing is established based on the geometric model, physical model, behavioral model and rule model; by presetting the fault depth and width in the digital twin model, based on the actual working conditions of the bearing, the vibration simulation response of the bearing under different preset faults is obtained; driven by the attention mechanism, the residual convolution network combined with the improved CBAM module is used to extract domain discrimination features from the time-frequency image converted from the vibration simulation signal; based on the weight mechanism of the two-stage domain discriminator, the contribution of each source sample is adaptively measured, and a lower weight is given to the outlier sample; based on the adversarial transfer strategy, the distribution difference between different domains is reduced, and the effective migration of diagnostic knowledge from the virtual space to the physical space is realized.

[0009] Furthermore, the geometric model is constructed based on the geometric size and shape of the target bearing and the positional relationship between the components.

[0010] Furthermore, the physical model is constructed based on physical properties of the target bearing, such as material, elastic modulus, Poisson's ratio, friction coefficient, etc.

[0011] Furthermore, the behavior model is constructed based on the action response of the target bearing to the internal and external environment and system instructions.

[0012] Furthermore, the rule model is constructed based on the historical operating data of the target bearing and the occurrence mechanisms of various failure modes.

[0013] Furthermore, different fault depths and widths are set in the geometric model of the bearing to obtain the digital twin model of the bearing under different faults.

[0014] Furthermore, based on the actual operating conditions of the bearing, the vibration simulation response of the bearing under different preset faults is obtained.

[0015] Furthermore, continuous wavelet transform is used to convert the vibration simulation response time domain signal into time-frequency distribution.

[0016] Furthermore, the time-frequency distribution is used as the input of the network to improve the convergence stability and recognition accuracy of the classification network.

[0017] Furthermore, the residual convolutional network of the improved CBAM module is pre-trained on the source labeled data to extract domain discriminative features from the time-frequency images.

[0018] Furthermore, a two-stage discriminator consisting of a sample importance discriminator and a domain discriminator is used to achieve partial domain adaptation.

[0019] Furthermore, the sample importance discriminator calculates the importance weight of each source sample through a weighted mechanism, and the domain discriminator is adversarially trained with the target feature extractor to reduce domain distribution divergence.

[0020] Furthermore, after the model training is completed, the optimized target feature extractor and classifier are used to perform fault diagnosis on the target bearing.

[0021] Compared with the prior art, the present invention has the following advantages and effects:

[0022] 1. A method for unsupervised intelligent fault diagnosis of bearings based on digital twins and partial domain adaptation is proposed. It effectively compensates for the problems of insufficient fault samples and missing data labels, solves the shortcomings of poor generalization ability of digital twin models, low diagnostic accuracy caused by inconsistent data distribution and data labels, and relaxes the dependence on the assumption of identical data labels, thus establishing a more accurate and generalized diagnostic model.

[0023] 2. A universal high-fidelity digital twin model of rolling bearings was established. This model only needs to set the bearing's structure, material, operating conditions and other parameters as well as the fault depth and width to obtain the bearing's vibration response.

[0024] 3. An improved CBAM residual convolutional network based on adversarial training is proposed, which can effectively enhance feature representation, extract domain-invariant features, and facilitate knowledge transfer. This method can reduce the feature distribution difference between simulated data and measured data.

[0025] 4. A two-stage domain discriminator weighting mechanism is designed to quantify the importance and transferability of each source sample. This method simplifies the adaptive structure and can adaptively identify and constrain abnormal samples, filter out outlier source samples, and promote positive transfer. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0027] Figure 1 This is a flow chart of an intelligent fault diagnosis method for rolling bearings based on digital twins and partial domain adaptation of the present invention.

[0028] Figure 2 The overall structure diagram of the improved CBAM residual convolutional network.

[0029] Figure 3 The overall structure diagram of partial domain adaptation of the improved CBAM model based on weighted adversarial network. DETAILED DESCRIPTION

[0030] In order to make the objectives, technical solutions and advantages of the present invention more clear, the technical solutions of the present invention will be described in detail below with reference to the accompanying drawings.

[0031] like Figure 1 As shown, the present invention provides a rolling bearing intelligent fault diagnosis method based on digital twin and partial domain adaptation, and its process mainly consists of 5 steps:

[0032] 1. Data generation: Based on the digital twin model of the bearing under different preset faults, the same operating conditions as the actual operation of the bearing are used to obtain the vibration simulation response of the bearing under different faults.

[0033] 2. Data preprocessing: Continuous wavelet transform is used to convert the vibration simulation response time domain signal into time-frequency distribution as the input of the network.

[0034] 3. Diagnostic model pre-training: The improved CBAM residual convolutional network is used for pre-training on the source labeled data.

[0035] 4. Adversarial Partial Domain Adaptation: A two-stage discriminator consisting of a sample importance discriminator and a domain discriminator is adopted to achieve partial domain adaptation.

[0036] 5. Target domain fault diagnosis: Use the optimized target feature extractor and classifier to diagnose bearing faults.

[0037] The geometric model of the bearing is constructed based on the geometric size, shape and positional relationship between the components of the target bearing. The physical model of the bearing is constructed based on the physical properties of the target bearing, such as material, elastic modulus, Poisson's ratio, friction coefficient, etc. The behavioral model of the bearing is constructed based on the action response of the target bearing to the internal and external environment and system instructions. The rule model of the bearing is constructed based on the historical operating data of the target bearing and the occurrence mechanism of various failure modes.

[0038] Based on the geometric model, physical model, behavioral model and rule model, the digital twin model of the rolling bearing is obtained; the fault depth and fault width are set in the digital twin model of the bearing, and based on the actual operating conditions of the bearing, the vibration simulation response of the bearing under different preset faults is obtained.

[0039] The vibration signal of the bearing is non-stationary. Extracting features directly from the original vibration signal will result in low network training efficiency and large amount of calculation. Compared with using vibration data to train a one-dimensional convolutional neural network, using two-dimensional time-frequency domain features to train a two-dimensional convolutional neural network can further improve the convergence stability and recognition accuracy of the classification network. Therefore, the vibration signal is converted into a time-frequency distribution through continuous wavelet transform to fully explore the failure characteristics of the bearing.

[0040] The residual convolutional network of the improved CBAM is used as a feature extractor to extract high-dimensional feature representations from time-frequency images. The use of residual structures can overcome the problem of network degradation. In the process of model construction, the depth of the model layer should be properly designed. As the number of layers increases, the difficulty of model adjustment will increase, thereby hindering the learning ability of the model. In addition, developing and deploying deeper models requires more computing power. The most suitable architecture is the one that can complete a specific task with as few model layers as possible. Based on the above analysis, four residual blocks are selected in the model to construct the feature extractor. Since the samples in the source domain are labeled, the F s The training supervision of and C is as follows:

[0041]

[0042] Specifically, the cross entropy loss function is used to calculate the empirical loss of the source domain classification task. When the samples in the target domain are unlabeled, F s The initial parameters and F t shared.

[0043] Since the attention mechanism has a powerful feature extraction capability, an improved CBAM attention mechanism is inserted into the feature extractor to enhance the important features in the time-frequency image and suppress unnecessary features, thereby improving the feature representation capability of the network. CBAM consists of two connected submodules: the channel attention module (CAM) and the spatial attention module (SAM), which continuously generate attention feature maps in the channel and spatial dimensions respectively. The input feature map is multiplied by the attention map to adaptively refine the features. The detailed structure of the improved CBAM is shown in Figure 2 shown.

[0044] Two major improvements are made in CBAM to enhance the feature aggregation capability. First, a residual connection is added between the input and output of CBAM to satisfy a shorter effective gradient propagation path. Second, the global maximum pooling with a sigmoid function is used to enhance the channel mapping to improve CAM; it can enhance channel feature aggregation and reduce the calculation of the multi-layer perceptron (MLP). Given the input feature map F, the overall calculation process of the improved CBAM can be expressed as:

[0045]

[0046]

[0047] Among them, M eca ∈R C×1×1 It is the enhanced CAM (ECAM), M sa ∈R 1×H×W For SAM, and represents element-wise multiplication and summation, F new is the refined feature map through ECAM, F n ' ew is the final feature map passed through SAM. ECAM acts as a feature detector to highlight the meaningful channels of the input features. In order to efficiently calculate the channel attention, the global average pooling and the global maximum pooling are multiplied using the sigmoid function to squeeze the spatial dimension and enhance the channel attention. The processed features are then sent to the MLP to produce the channel attention feature map. The calculation of ECAM is as follows:

[0048]

[0049] Where σ represents the sigmoid function, GAP represents global average pooling, and GMP represents global maximum pooling.

[0050] As a complement, SAM assigns weights to each point in the feature map to highlight the informative part. To calculate the channel attention, the average pooling feature map and the maximum pooling feature map are concatenated. Then a convolution operation is used to generate a spatial attention map. The calculation process of SAM is as follows:

[0051]

[0052] Among them, Conc represents the cascade operation, CAP represents the channel average pooling, CMP represents the channel maximum pooling, and f 7×7 Represents a 7×7 convolution kernel. Through this ECAM and SAM process, the feature extraction network can adaptively adjust the input feature map and focus on the important parts of it.

[0053] In order to solve the negative transfer problem caused by outlier classes and transfer useful knowledge to the target domain, a weighting mechanism is used, which measures the importance of each source sample through the sample importance discriminator D0. Optimizing the probability output of D0 can determine whether the source sample is from the shared class or the outlier class. The source samples are then weighted and domain adapted with the target samples to promote positive transfer, thereby enhancing the reliability of the source data and effectively suppressing the influence of outlier samples.

[0054] Given the extracted feature representation F s,t (x s,t ), the similarity between source and target domain features can be obtained as follows:

[0055] D0(F s,t (x s,t ))=p(y s,t =1|F s,t (x s,t ))=δ(α(F s,t (x s,t )))

[0056] Where δ is the sigmoid function, α refers to the network parameters of D0, and y s,t is the domain label. D0 is a binary classifier that labels source samples as 1 and target samples as 0. For any F s (x) and F t (x), the objection of D0 is maximized as:

[0057]

[0058] The best D0 is obtained at:

[0059]

[0060] The optimized D0 can give the probability that the input sample comes from the source domain. If Then the sample is likely to come from the anomaly class in the source domain, because its distribution area rarely or not covers the target sample, and it can be well distinguished from the target distribution by D0. These samples should be given little attention because they contribute little to transferring useful classification knowledge to the target domain. On the contrary, if If is close to 0, the input sample is likely to come from the shared class between domains and should be assigned a larger weight during the transfer process. Therefore, the function of the weighting mechanism should be designed to be similar to Inversely proportional:

[0061]

[0062] Therefore, the source samples from the shared class will have a larger weight than the outlier class, and the negative impact of the outlier samples can be suppressed. In order to obtain the relative importance weight of each source sample, the weight function is normalized as:

[0063]

[0064] Although the sample importance discriminator D0 is trained to discriminate features from the source and target domains, it does not participate in the domain adversarial learning process, which means that F t is not trained to confuse D0. After D0 gives the weights to each source sample, the weighted source features and target features are sent to the domain discriminator D, which is compared with F t A minimum-maximum game is played to achieve partial domain adaptation and reduce the distance between shared classes. The objective function of the weighted adversarial network can be written as:

[0065]

[0066] where F t is trained to minimize L2 to make the distribution difference between the two domains smaller, while D is trained to maximize L2 to make the bound on the domain distribution difference tighter. In order to better constrain F t , the conditional entropy loss H(.) is further used to minimize the prediction uncertainty when processing the target sample:

[0067]

[0068] The overall structure of a partial domain adaptive bearing unsupervised fault diagnosis scheme based on an improved CBAM model with weighted adversarial network is as follows: Figure 3 shown.

[0069] Different from most existing domain adaptation models that use a shared feature extractor, this model uses two residual convolutional networks combined with an improved CBAM with the same structure to capture domain-specific features, denoted as F and s and F tThe two-stage discriminator contains a sample importance discriminator D0 and a domain discriminator D. D0 measures the importance of source samples according to the proposed weighting mechanism. D is related to the target feature extractor F t Perform a minimum-maximum game. Use the gradient reversal layer for back propagation and parameter optimization. After training, the target feature extractor F t And the classifier C can be used for target bearing fault diagnosis.

[0070] The proposed diagnostic model is trained through a multi-task loss function, which consists of three parts: (1) source domain pre-training loss L s ; (2) Sample importance discriminator loss L1; (3) Adversarial training loss L2. The overall objective function can be constructed as:

[0071]

[0072] γ and λ are the coefficients of the equilibrium loss function. The optimization problem is to obtain parameter values ​​that satisfy the following conditions:

[0073]

[0074] Iteratively train F by stochastic gradient descent algorithm s ,C,D0,F t and the parameters of D The gradient reversal layer is inserted in F t Between and D.

[0075] The above are only exemplary embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. An intelligent fault diagnosis method for rolling bearings based on digital twins and partial domain adaptation, characterized by: A high-fidelity digital twin model of the bearing was established, which was used to efficiently obtain the vibration simulation response of the bearing under different preset faults, solving the problems of insufficient fault samples and missing data labels. Based on the partial domain adaptation method, the effective migration of fault diagnosis knowledge from virtual space to physical space was achieved, solving the problem of low fault diagnosis accuracy caused by poor generalization ability of the digital twin model and inconsistency between data distribution and data labels.

2. The rolling bearing intelligent fault diagnosis method based on digital twin and partial domain adaptation according to claim 1 is characterized by: The high-fidelity digital twin model of the bearing is constructed based on the bearing's geometric model, physical model, behavioral model, and rule model. By presetting the fault depth and width in the digital twin model and based on the actual operating conditions of the bearing, the vibration simulation response of the bearing under different preset faults is obtained, providing a sufficient fault data set for partial domain adaptation.

3. The rolling bearing intelligent fault diagnosis method based on digital twin and partial domain adaptation according to claim 1 is characterized by: The partial domain adaptation method, driven by the attention mechanism, combines the residual convolutional network of the improved CBAM module to extract domain discriminant features from the time-frequency image converted from the vibration simulation signal; Based on the weight mechanism of the two-stage domain discriminator, the contribution of each source sample is adaptively measured, and outlier samples are given lower weights; based on the adversarial transfer strategy, the distribution differences between different domains are reduced, and the migration of diagnostic knowledge from virtual space to physical space is realized.

Images