A rotating machinery health assessment method based on adversarial metric convolutional network

Through the adversarial metric convolution network and attention weight module, the problem of the difference in the data distribution of source and target domains in rotating machinery is solved, and the health assessment under different working conditions is realized, which improves the accuracy and reliability of mechanical health assessment. Great.

CN113947012BActive Publication Date: 2024-11-08SOUTHEAST UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111067866.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-09-13
Publication Date
2024-11-08
Estimated Expiration
2041-09-13

AI Technical Summary

Technical Problem

When predicting the remaining service life of rotating machinery, the prior art faces the problem of the difference in the distribution of data characteristics of the source domain and the target domain, which leads to the model lacking domain generalization ability and being unable to effectively adapt to mechanical health assessment under different working conditions.

Method used

Adversarial metric convolution network is used to maximize the mutual information between the target domain data and mechanical degradation characteristics through adversarial discriminator and information comparison device. Combined with the attention weight module, robust domain invariant features are extracted to achieve the implementation of different operating conditions. health assessment.

Benefits of technology

The generalization of traditional domain adaptation methods was improved, the distribution differences of degradation characteristics were successfully aligned, and the accuracy and robustness of mechanical health assessments were improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113947012B_ABST
    Figure CN113947012B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for health assessment of rotating machinery of adversarial metric convolutional network, comprising the following steps: using continuous wavelet transform to convert the collected vibration signal of rotating machinery into a time-frequency diagram, and dividing the data set and marking the life value; constructing an adversarial metric convolutional network; inputting data into the network; using a source feature extractor to extract source features; inputting the source features into a regressor to predict the life value; migrating the parameters of the source feature extractor to a target feature extractor; using a target feature extractor to extract target features; inputting the source and target features into an adversarial discriminator, and inputting the target features and target data into an information comparator; updating parameters; combining the target feature extractor and the regressor to realize health assessment of rotating machinery. The present invention can monitor the health status of rotating machinery more accurately.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of state monitoring and health assessment of rotating machinery equipment, and in particular to a rotating machinery health assessment method based on an adversarial metric convolutional network. Background Art

[0002] Industrial machinery cannot avoid failures in long-term operation, such as aircraft engines and wind turbines, which can lead to catastrophic accidents. Therefore, it is very important to use fault prognosis and health management (PHM) technology to accurately diagnose the current health status of the machine. Through PHM technology, maintenance plans can be made in advance based on the future behavior of the mechanical system, shortening unnecessary downtime, reducing costs, and improving the reliability of the mechanical system. Among them, accurate prediction of the remaining useful life (RUL) of the mechanical system is a key task of PHM technology, and the time series regression problem is challenging. Rolling bearings are widely used parts in rotating machinery, and they are often subjected to complex time-varying conditions during operation. It is reported that more than 40% of mechanical system failures are related to bearings. The use of historical data to predict the RUL of rolling bearings is of great value to the fault prediction and health management of rotating machinery, and has attracted more and more attention from researchers in recent years. Therefore, the remaining life value of the bearing is estimated based on the vibration signal during the operation of the bearing.

[0003] Many existing RUL prediction methods can be divided into two categories: model-based and data-driven. Model-based methods use statistical models based on degradation phenomena (caused by wear, fatigue, corrosion, etc.) to predict RUL. These methods require expert knowledge to accurately model the system dynamics and evaluate its health trends, and are usually unable to simulate complex systems. Data-driven methods aim to convert the data provided by sensors into parametric models of degradation behavior. These methods use available data to extract hidden features and accurately predict RUL. In general, data-driven methods usually include machine learning and deep learning models. Machine learning methods need to extract various degradation features (such as mean, skewness, kurtosis) from the data provided by different sensors (such as vibration, force, temperature) and use learning algorithms to regress RUL. However, machine learning methods are heavily dependent on the extracted degradation features and often require prior knowledge. In recent years, deep learning has become more attractive in dealing with time series regression, such as convolutional neural networks, deep belief networks, recurrent neural networks, and autoencoders. Deep learning methods achieve layer-by-layer nonlinear transformation of data features through a multi-hidden layer structure to extract the best feature representation. Unfortunately, these methods only work effectively under the assumption that the source and target domain data are independent and identically distributed. This assumption is unrealistic for many real-world rolling bearing applications. Since the run-to-failure process of a bearing is often subjected to time-varying conditions, this can lead to differences in feature distributions. Pre-trained models that achieve the best performance under the current conditions usually do not work well when the conditions change. Due to the different data distributions for different conditions, the model lacks domain generalization capabilities, resulting in limited performance. Summary of the invention

[0004] In order to solve the above problems, the present invention proposes a rotating machinery health assessment method of an adversarial metric convolutional network on the basis of domain adaptation; at the same time, when the adversarial discriminator in adversarial domain adaptation is used to learn domain invariant features, the mutual information between the target domain data and the target mechanical degradation features is maximized, thereby maintaining the inherent manifold structure of the target domain data. An attention weight module is embedded in the feature extractor to reduce the model's focus on global features, so that it pays more attention to fine-grained information in local related subdomains, and then learns more robust domain invariant features, thereby realizing health assessment of mechanical equipment under different working conditions. To achieve this purpose, the present invention provides a rotating machinery health assessment method of an adversarial metric convolutional network, which is characterized by comprising the following steps:

[0005] Step 1: Use an acceleration sensor to collect horizontal vibration signals of bearings in rotating machinery under different working conditions when the bearings run to failure, and use continuous wavelet transform to obtain time-frequency diagrams of the vibration signals, divide all time-frequency diagrams into source domain and target domain data sets, and finally mark the remaining life value of the rotating machinery corresponding to the source domain data;

[0006] Step 2, construct an adversarial metric convolutional network, including a source feature extractor, a target feature extractor, a regressor, an adversarial discriminator, and an information comparator;

[0007] Step 3, input the labeled source domain data and the unlabeled target domain data into the adversarial metric convolutional network;

[0008] Step 4, using a source feature extractor to extract feature vectors in the source domain time-frequency diagram and reconstruct source mechanical degradation features;

[0009] Step 5, input the source mechanical degradation features into the regressor, predict the remaining life value, and use the regression loss to update the parameters of the source feature extractor and the regressor;

[0010] Step 6, migrating the parameters of the source feature extractor to the target feature extractor;

[0011] Step 7, using a target feature extractor to extract feature vectors in the target domain time-frequency graph and reconstruct target mechanical degradation features;

[0012] Step 8, inputting the source and target mechanical degradation features into the adversarial discriminator, and inputting the target mechanical degradation features and the target domain data into the information comparator;

[0013] Step 9, using the adversarial loss of the adversarial discriminator and the information loss of the information comparator to update the parameters of the target feature extractor;

[0014] Step 10, combining the optimized target feature extractor and regressor to predict the remaining life value of the target domain data, thereby realizing the health assessment of the rotating machinery.

[0015] Beneficial effects of the present invention:

[0016] (1) This paper introduces the idea of ​​adversarial discrimination into the domain adaptation method and proposes an adversarial domain adaptation architecture with information loss and adversarial loss, which can greatly improve the generalization of traditional domain adaptation problems and successfully align the distribution differences of degraded features.

[0017] (2) The present invention proposes an attention weight layer that can fully consider the global and local contextual semantic information while the model extracts the degraded features well. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flow chart of the present invention;

[0019] Figure 2 is the vibration signal in the horizontal direction of the experiment of the present invention;

[0020] Figure 3It is the time-frequency diagram of the vibration signal in the horizontal direction of the experiment of the present invention;

[0021] Figure 4 This is a diagram of the adversarial metric convolutional network structure of the present invention;

[0022] Figure 5 These are the results obtained from the experiments of the present invention. DETAILED DESCRIPTION

[0023] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments:

[0024] The present invention provides a rotating machinery health assessment method based on a deep temporal convolutional network, aiming to highlight the extraction of transferable degradation features in different domains and achieve higher-precision machinery health assessment.

[0025] Figure 1 The steps of the present invention are described in detail below in conjunction with the flow chart.

[0026] Step 1: Use an acceleration sensor to collect horizontal vibration signals of bearings in rotating machinery under different working conditions when the bearings run to failure, and use continuous wavelet transform to obtain time-frequency diagrams of the vibration signals, divide all time-frequency diagrams into source domain and target domain data sets, and finally mark the remaining life value of the rotating machinery corresponding to the source domain data;

[0027] Step 2, construct an adversarial metric convolutional network, including a source feature extractor, a target feature extractor, a regressor, an adversarial discriminator, and an information comparator;

[0028] Step 3, input the labeled source domain data and the unlabeled target domain data into the adversarial metric convolutional network;

[0029] Step 4, using a source feature extractor to extract feature vectors in the source domain time-frequency diagram and reconstruct source mechanical degradation features;

[0030] Step 5, input the source mechanical degradation features into the regressor, predict the remaining life value, and use the regression loss to update the parameters of the source feature extractor and the regressor;

[0031] Step 6, migrating the parameters of the source feature extractor to the target feature extractor;

[0032] Step 7, using a target feature extractor to extract feature vectors in the target domain time-frequency graph and reconstruct target mechanical degradation features;

[0033] Step 8, inputting the source and target mechanical degradation features into the adversarial discriminator, and inputting the target mechanical degradation features and the target domain data into the information comparator;

[0034] Step 9, using the adversarial loss of the adversarial discriminator and the information loss of the information comparator to update the parameters of the target feature extractor;

[0035] Step 10, combining the optimized target feature extractor and regressor to predict the remaining life value of the target domain data, thereby realizing the health assessment of the rotating machinery.

[0036] Furthermore, the time-frequency diagrams under different working conditions in step 1 are divided into source domain data and target domain data in represents the source domain time-frequency diagram, express The corresponding remaining life value of the rotating machinery, n s express The number of samples, represents the target domain time-frequency diagram, n T express The number of samples, Figure 2 The mechanical vibration signal collected by the present invention, Figure 3 The time-frequency diagram converted by the present invention; specifically, the time-frequency data of the rotating machinery under one working condition is divided into source domain data, and the time-frequency data under another working condition is divided into target domain data.

[0037] Furthermore, the adversarial metric convolutional network in step 2 consists of a source feature extractor, a target feature extractor, a regressor, a adversarial discriminator and an information comparator. Figure 4 It is the adversarial metric convolutional network of the present invention; specifically, the structure of the source feature extractor and the target feature extractor is consistent, consisting of 2 layers of temporal residual blocks, 2 layers of attention weight layers and 2 layers of maximum pooling layers; specifically, the regressor and the adversarial discriminator are both composed of 3 layers of fully connected layers; the information comparator is composed of 2 layers of encoders.

[0038] Furthermore, the temporal residual block consists of two layers of dilated atrous convolution and one layer of residual convolution, where the dilation rate d of the first layer of dilated atrous convolution is 1, and the dilation rate d of the second layer of dilated atrous convolution is 2; multivariate signal After inputting the temporal residual block, each layer of dilated dilated convolution is performed separately, and the calculation method is as follows:

[0039]

[0040] Where i represents the i-th time-frequency graph signal element, k represents the kernel size, represents the bias, f(k) represents the filter, and σ represents the activation function; the multivariate feature reconstruction calculation method of the two-layer temporal convolutional network is as follows:

[0041]

[0042] In the formula, Φ d represents the dilated dilated convolution, m 1 and m 2 Represents the feature mask of dropout, represents residual convolution, represents matrix addition, w l represents the residual convolution weight, b l Represents the residual convolution bias.

[0043] Specifically, the source feature extractor E S and target feature extractor E T From input to output, it includes the first layer of temporal residual blocks, the first layer of attention weight layer, the first layer of maximum pooling layer, the second layer of temporal residual blocks, the second layer of attention weight layer and the second layer of maximum pooling layer; specifically, the source domain data x S and target domain data x T Input source feature extractor E S and target feature extractor E T , respectively extract the source mechanical degradation characteristics f S and the target mechanical degradation characteristic f T .

[0044] Further, the attention weight calculation decomposition steps are:

[0045] Specifically, in step 5.1, the input X is subjected to three linear transformations to obtain the corresponding outputs Q, K and V;

[0046] Q=w q X

[0047] K=w k X

[0048] V=w v X

[0049] In the formula, w k , w q and w v are the weight matrices of Q, K and V respectively;

[0050] Step 5.2, using the weight matrices of Q, K and V and the softmax function, calculate the attention weights

[0051]

[0052] Where, d k is the dimension of the weight matrix K, and T is the matrix transpose operation.

[0053] Furthermore, the number of neurons in the three fully connected layers of the regressor in step 5 are 32, 16, and 1 respectively;

[0054] The source mechanical degradation characteristic f S Input to the regressor to predict the remaining life value Calculating regression loss

[0055]

[0056] In the formula, Represents the source domain data x S The remaining lifetime label value, n s Represents the source domain data x S The number of samples.

[0057] Furthermore, the number of neurons in the three fully connected layers of the adversarial discriminator in step 8 are 64, 32, and 1 respectively;

[0058] Specifically, the source mechanical degradation characteristics f S and the target mechanical degradation characteristic f T Input to the adversarial discriminator, and adversarially discriminate the degraded features, so that the adversarial discriminator cannot correctly distinguish f as much as possible S and f T , to minimize f S and f T The difference between them can reduce the difference between data under different working conditions. The adversarial training process is as follows:

[0059]

[0060] In the formula, x S represents the source domain data, x T represents the target domain data, E S represents the source feature extractor, E T represents the target feature extractor, represents the adversarial loss, D represents the adversarial discriminator, represents the source domain feature distribution, represents the target domain feature distribution, and log(·) represents the logarithmic function.

[0061] Furthermore, in step 8, the information comparator uses the target mechanical degradation feature f T and target domain data x T As input, the target mechanical degradation feature f T Input to the first encoder f θ1 Generate different negative samples, where different feature masks are used to generate different negative samples; target data x T Input to the second encoder f θ2 Generate 1 positive sample; thus forming a positive and negative sample pair where x iis a positive sample, is a negative sample; using information loss Compare the target mechanical degradation characteristics f T and target domain data x T information;

[0062]

[0063]

[0064] In the formula, h i and Refers to x i and The batch expression of , τ represents the scale factor, Θ represents E T parameter.

[0065] By minimizing information loss To maximize x T and f T The mutual information between them.

[0066] Further, the target feature extractor E in step 9 T The update process is described as follows:

[0067]

[0068] In the formula, Indicates a total loss. Represents resistance to loss, represents information loss, and λ represents the regularization coefficient, which is used to balance the actual contribution of adversarial loss and information loss to the model;

[0069] By the total loss For the target feature extractor E T Update the parameters, specifically:

[0070]

[0071] In the formula, θ f Represents the target feature extractor parameters, η represents the model learning rate, and through the above gradient calculation, a model with optimal parameters can be obtained;

[0072] By combining the optimized target feature extractor and regressor and combining the corresponding vibration signals of rolling bearings, the corresponding rotating machinery health assessment tasks can be completed.

[0073] Figure 5 The test results obtained in the experiment show that the predicted value of the present invention successfully fits the actual remaining life value, thereby verifying the effectiveness of the method proposed in the present invention in the health assessment of rotating machinery.

[0074] The above description is only a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent change made based on the technical essence of the present invention still falls within the scope of protection required by the present invention.

Claims

1. A rotating machinery health assessment method based on adversarial metric convolutional network, characterized by The steps include: Step 1: Use an acceleration sensor to collect horizontal vibration signals of bearings in rotating machinery under different working conditions when they are running to failure, and use continuous wavelet transform to obtain time-frequency diagrams of the vibration signals. All time-frequency diagrams are divided into source domain and target domain data sets, and finally the remaining life value of the rotating machinery corresponding to the source domain data is marked; Step 2: Construct an adversarial metric convolutional network, including a source feature extractor, a target feature extractor, a regressor, an adversarial discriminator, and an information comparator. Step 3: Input the labeled source domain data and unlabeled target domain data into the adversarial metric convolutional network; Step 4: Use the source feature extractor to extract the feature vectors in the source domain time-frequency graph and reconstruct the source mechanical degradation features; Step 5: Input the source mechanical degradation features into the regressor to predict the remaining life value, and use the regression loss to update the parameters of the source feature extractor and regressor; Step 6: Migrate the parameters of the source feature extractor to the target feature extractor; Step 7: Use the target feature extractor to extract the feature vectors in the target domain time-frequency graph and reconstruct the target mechanical degradation features; Step 8: Input the source and target mechanical degradation features into the adversarial discriminator, and input the target mechanical degradation features and target domain data into the information comparator; Step 9: Update the parameters of the target feature extractor using the adversarial loss of the adversarial discriminator and the information loss of the information comparator; Step 10: Combine the optimized target feature extractor and regressor to predict the remaining life value of the target domain data, thereby realizing the health assessment of the rotating machinery; The adversarial metric convolutional network in step 2 consists of a source feature extractor, a target feature extractor, a regressor, a adversarial discriminator, and an information comparator; the source feature extractor and the target feature extractor have the same structure, consisting of two layers of temporal residual blocks, two layers of attention weight layers, and two layers of maximum pooling layers; the regressor and the adversarial discriminator both consist of three layers of fully connected layers; the information comparator consists of two layers of encoders; The time residual block consists of two layers of dilated dilated convolution and one layer of residual convolution, where the dilation rate d of the first layer of dilated dilated convolution is 1, and the dilation rate d of the second layer of dilated dilated convolution is 2; After inputting the temporal residual block, each layer of dilated convolution performs dilated convolution separately, and the calculation method is as follows: Where i represents the i-th time-frequency graph signal element, k represents the kernel size, Denotes bias, f(k) denotes filter, and σ denotes activation function. The multivariate feature reconstruction calculation method of the two-layer temporal convolutional network is as follows: Where, Φ d represents the dilated hole convolution, m1 and m2 represent the feature mask of dropout, represents residual convolution, represents matrix addition, w l represents the residual convolution weight, b l Represents the residual convolution bias; Source feature extractor E S and target feature extractor E T From input to output, including the first layer of temporal residual blocks, the first layer of attention weight layer, the first layer of maximum pooling layer, the second layer of temporal residual blocks, the second layer of attention weight layer and the second layer of maximum pooling layer; source domain data x S and target domain data x T Input source feature extractor E respectively S and target feature extractor E T , respectively extract the source mechanical degradation characteristics f S and the target mechanical degradation characteristic f T ; Note that the weight calculation steps are as follows: Step 5.1: Perform three linear transformations on the input X to obtain the corresponding outputs Q, K, and V. Q=w q X K=w k X V=w v X Where w k , w q and w v are the weight matrices of Q, K and V respectively; Step 5.2, using the weight matrices of Q, K and V and the softmax function, calculate the attention weight Where, d k is the dimension of the weight matrix K, and T is the matrix transpose operation; The target feature extractor E in step 9 T The update process is described as follows: Where, Indicates a total loss. Represents resistance to loss, represents information loss, Θ represents E T Parameter, λ represents the regularization coefficient, which is used to balance the actual contribution of adversarial loss and information loss to the model; By the total loss For target feature extractor E T Update the parameters, specifically: Where θ f Represents the target feature extractor parameters, η represents the model learning rate, and through the above gradient calculation, a model with optimal parameters can be obtained; By combining the optimized target feature extractor and regressor and combining the corresponding vibration signals of rolling bearings, the corresponding rotating machinery health assessment task can be completed.

2. The rotating machinery health assessment method based on an adversarial metric convolutional network according to claim 1 is characterized in that: The time-frequency diagrams under different working conditions in step 1 are divided into source domain data and target domain data in represents the source domain time-frequency diagram, express The corresponding remaining life value of the rotating machinery, n s express The number of samples, represents the target domain time-frequency diagram, n T express The number of samples; among them, the time-frequency data of the rotating machinery under one working condition is divided into the source domain data, and the time-frequency data under the other working condition is divided into the target domain data.

3. The rotating machinery health assessment method based on an adversarial metric convolutional network according to claim 1 is characterized in that: The number of neurons in the three fully connected layers of the regressor in step 5 is 32, 16, and 1 respectively; The source mechanical degradation characteristic f S Input to the regressor to predict the remaining life value Calculating regression loss Where, Represents the source domain data x S The remaining lifetime label value, n s Represents the source domain data x S The number of samples.

4. The rotating machinery health assessment method based on an adversarial metric convolutional network according to claim 1, characterized in that: In step 8, the number of neurons in the three fully connected layers of the adversarial discriminator are 64, 32, and 1 respectively; Source mechanical degradation characteristics f S and the target mechanical degradation characteristic f T Input to the adversarial discriminator, and adversarially discriminate the degraded features, so that the adversarial discriminator cannot correctly distinguish f S and f T , to minimize f S and f T The difference between them can be reduced, thereby reducing the difference between data under different working conditions. The adversarial training process is as follows: Where x S represents the source domain data, x T represents the target domain data, E S represents the source feature extractor, E T represents the target feature extractor, represents the adversarial loss, D represents the adversarial discriminator, represents the source domain feature distribution, represents the target domain feature distribution, and log(·) represents the logarithmic function.

5. The rotating machinery health assessment method based on an adversarial metric convolutional network according to claim 1, characterized in that: In step 8, the information comparator uses the target mechanical degradation feature f T and target domain data x T As input, the target mechanical degradation feature f T Input to the first encoder f θ1 Generate different negative samples, where different feature masks are used to generate different negative samples; target data x T Input to the second encoder f θ2 Generate 1 positive sample; thus forming a positive and negative sample pair where x i is a positive sample, is a negative sample; using information loss Compare the target mechanical degradation characteristics f T and target domain data x T information; Where h i and Refers to x i and The batch expression, τ represents the scale factor, E T represents the target feature extractor, Θ represents E T parameter.