KR102387663B1 - Apparatus for fault diagnosis using domain adaptation with semantic clustering algorithm and method for fault diagnosis using the same - Google Patents

Abstract

The present invention relates to a diagnosis apparatus applied with a domain adaptation and semantic clustering algorithm. The apparatus comprises: a feature extractor unit learning a machine learning-based feature calculation model that calculates a first feature and a second feature from a first data set including labeled time series signal data and a second data set including non-labeled time series signal data; a label classifier unit learning a machine learning-based state classification model by using the first data set and calculating classifier loss (LC); a domain discrepancy unit learning a machine learning-based domain adaptation model that performs domain adaptation and calculating domain-related loss (LD); and a semantic clustering unit calculating semantic clustering loss (LSC). According to the present invention, the apparatus can efficiently diagnose a state of a target system without label information based on diagnosis data acquired from a source system.

Description

A diagnostic device and method to which domain adaptation and semantic clustering algorithms are applied

The present invention relates to a system failure diagnosis technology, and more particularly, to a diagnosis apparatus and method to which a domain adaptation and semantic clustering algorithm for efficiently diagnosing the state of a target system based on abundant diagnostic data of a source system is applied. .

Various types of mechanical parts and systems are used in many industries, and defects in these parts or systems cause enormous social and financial damage. In the past, it relied heavily on experts and domain knowledge to check the actual system state one by one and judge its performance according to a set rule.

In order to overcome this, research on a technology that can predict the state of a system in advance is being actively conducted, and in particular, a technology that predicts the state of a system using a deep learning method is increasing the stability and efficiency of the system. In order to further increase the accuracy of the prediction, it is necessary to analyze a large number of signal data according to each state of the system. In other words, to develop robust fault diagnosis methods using deep learning, sufficient labeled data sets in the target system must be prepared for all system states of interest.

However, it has the disadvantage that it is difficult to achieve high target diagnostic accuracy when the amount of label data obtained from the system is not sufficient or when the distribution types found in the training and test data sets are inconsistent.

In order to overcome this problem and improve target diagnostic performance, an unsupervised domain adaptation (UDA)-based error diagnosis method that can transfer knowledge or information obtained from labeled source data to an unlabeled target domain has become popular. The main concept of UDA is to learn common features that can be shared by two domains, which is achieved by reducing the domain mismatch between two different but related domains. That is, a classifier trained using source data with label information can be adapted for unlabeled target domain diagnosis. Several types of neural network-based UDA methods are available depending on the learning strategy.

First, the inconsistency-based method employs specific metrics including Correlation Alignment (CORAL) and Maximum Mean Discrepancy (MMD). These metrics quantify the discrepancy between the target and source domains. Then, by reducing these inconsistency metrics, domain invariant properties can be obtained.

Second, adversarial adaptation approaches obtain common characteristics based on adversarial training methods in which domain classifiers and feature extractors compete with each other.

However, most of these UDA methods focus only on minimizing the discrepancy in the marginal distributions of two different regions. In this case, even if the marginal distribution is well-aligned, the diagnostic performance for each class of the target region may not generalize satisfactorily. This is because the target samples of each class may not cluster well or align well with the source data with the same class label. As a result, samples located near the boundaries of each class can be easily misclassified.

Metric learning is a learning strategy that can learn good representations based on comparison of distance or similarity metrics between samples. A better representation can be obtained by mapping similar samples closer based on these distance metrics and keeping dissimilar samples apart from each other. In other words, the generalization performance of the model can be improved by learning good features that better separate samples into classes. Therefore, recently, metric learning to improve diagnostic performance for fault diagnosis has been extensively studied.

Therefore, in order to solve the above-mentioned problem in UDA and improve the diagnostic performance for the target domain, an efficient diagnostic technology that can grasp the state of the system from the limited measurement signal using a new domain adaptation technology using the semantic clustering method is needed.

In order to effectively understand the technology of the present invention, a Convolutional Neural Network (CNN) and an Unsupervised Domain Adaptation (UDA) will be described.

First, CNN will be described. Among the existing deep learning methods, CNN (Convolutional Neural Network) is one of the most widely used methods. It consists of a combination of multiple layers, such as convolutional and pooling layers, and unlike traditional fully connected neural networks, CNNs use a weight matrix with a smaller dimension than the input data called a kernel. Therefore, CNNs have local connectivity characteristics that can learn local patterns within a small area of input data. In the convolutional layer, an output layer called a feature map is generated by sliding a kernel within the input data and performing a convolution operation. Based on these operations, features are extracted from the input data. Also, CNN uses a kernel with equal weight values for convolution operations on all parts of the input data. This is called parameter sharing. As a result, CNN can greatly minimize the number of parameters and increase computational efficiency. The pooling layer pools a specific value (ie, the maximum value or the average value) in the sub-region. Therefore, the feature map of the previous layer is reduced. Also, like other conventional neural networks, nonlinear activation functions such as Rectified Linear Unit (ReLU) are used to learn nonlinear features. There are different types of CNNs depending on the input data dimension. For fault diagnosis of rotating machines, one-dimensional CNN (1D CNN) is widely used because a one-dimensional time-series vibration signal is generally input data. In the case of 1D CNN, a 1D kernel is used to extract appropriate feature representations from 1D vibration signals. A fully connected neural network that acts as a classifier is provided at the end of the CNN.

Next, UDAs will be described. For a real rotating machine, obtaining a sufficient amount of labeled data can be expensive and not easy. Additionally, the distribution of the test data is often different from the distribution of the data used for training due to noise, changes in operating conditions, or other factors (for example, if the data was obtained from a different system). In this situation, transfer learning can be utilized to develop a defect diagnosis method. For clarity on transfer learning, notations and definitions are presented here. First, the domain (D) is composed of a marginal probability (P(X)) and a feature space (x), where X, where X∈x, means input data. The task (τ) consists of a prediction function (P(Y|X)) and a label space (y), where Y ∈y={1, ......, C} means the health state and C represents the number of possible cases. Transfer learning refers to a learning strategy that uses knowledge or information obtained from another (but related) source domain to generate a diagnostic model applicable to a target domain. On the other hand, the technology that is the background of the present invention is disclosed in Korean Patent Registration Nos. 10-2015417 and 10-2014820.

SUMMARY OF THE INVENTION An object of the present invention is to solve the conventional problems described above, and a diagnostic device to which a domain adaptation and semantic clustering algorithm for efficiently diagnosing the state of a target system based on abundant diagnostic data of a source system simulating a target system is applied; to provide a method.

The above object is, according to the present invention, in a diagnosis apparatus to which a domain adaptation and semantic clustering algorithm performed by a state diagnosis system of a mechanical system without label information is applied, including time series signal data measured and labeled in the first system A machine learning-based feature calculation model that calculates the first features and the second features from the first dataset and the second dataset including unlabeled time-series signal data measured in the second system. learning feature extractor unit; a label classifier unit for learning a machine learning-based state classification model for classifying the state of the first system using the first dataset, and calculating a classifier loss ( _LC ); Learning a machine learning-based domain adaptation model that performs domain adaptation by setting the first dataset as a source domain and setting the second dataset as a target domain as a target domain, a domain dispensing unit for calculating domain-related loss ( _LD ); and a semantic clustering unit for learning a machine learning-based clustering model for clustering by state based on the label of the first dataset, and calculating a semantic clustering loss (L _SC ), wherein the classification loss is, It is calculated based on Equation 1, the domain adaptation loss is calculated based on Equation 2, and the semantic clustering loss is calculated based on Equation 3. achieved by

delete

{Formula 1}

는 상기 분류손실, 상기 C는 상기 제1시스템의 상태의 수, 상기 Y는 라벨 벡터, 상기

는 상기 Y의 k번째 노드의 값, 상기 p는 상기 Y의 k번째 노드에 해당하는 확률벡터)(remind

is the classification loss, C is the number of states of the first system, Y is the label vector, and

is the value of the k-th node of Y, and p is a probability vector corresponding to the k-th node of Y)

{Formula 2}

는 상기 도메인적응손실, 상기 H는 universal RKHS(reproducing kernel Hilbert space), 상기 n_S와 상기 n_T는 각각 상기 제1데이터셋과 상기 제2데이터셋의 수, 상기

와 상기

는 각각 상기 제1피쳐와 상기 제2피쳐)(remind

is the domain adaptation loss, H is a universal reproducing kernel Hilbert space (RKHS), n _S and n _T are the numbers of the first and second datasets, respectively, and

and above

is the first feature and the second feature, respectively)

{Formula 3}

는 상기 의미군집화손실, 상기

는 상기 시멘틱클러스터링부의 특성층의 수, 상기

는 k번째 밸런싱 파라미터(balancing parameter), 상기

는 상기 제1데이터셋의 수, 상기

는 k번째 피쳐 레벨에서

번째 상기 제1데이터셋의 피쳐 벡터, 상기

는

번째 데이터셋과

번째 데이터셋이 같은 클래스에 속하면 1이고 다른 클래스라면 0인 행렬

의 값, 상기 d는 서로 다른 상태의 두 데이터셋 사이의 최소 거리를 제어하는 값)(remind

is the semantic clustering loss,

is the number of characteristic layers of the semantic clustering part, the

is the kth balancing parameter (balancing parameter), the

is the number of the first dataset, the

is at the kth feature level

th feature vector of the first dataset, the

Is

the second dataset and

A matrix that is 1 if the second dataset belongs to the same class and 0 if it is a different class.

value of , where d is a value controlling the minimum distance between two datasets in different states)

In addition, the feature extractor unit, the domain discrepancy unit, and the semantic clustering unit are trained in mini-batch units of the first dataset and the second dataset, and the first dataset and the When learning is performed on all small groups of the second dataset, small groups may be randomly re-extracted from the first dataset and the second dataset to be re-learned.

Also, the feature extractor unit may calculate the first feature and the second feature based on Equation (4).

{Formula 4}

는 상기 제1데이터셋 또는 상기 제2데이터셋, 상기

는 상기

으로부터 계산된 상기 제1피쳐 또는 상기 제2피쳐, 상기

는 활성화함수(activation function), 상기

는 웨이트(weight), 상기 bias는 상수)(remind

is the first data set or the second data set, the

is said

The first feature or the second feature calculated from

is an activation function, the

is a weight, the bias is a constant)

)을 계산하는 연산부를 더 포함하며, 상기 연산부는, 수식 6 및 경사하강법(gradient descent)에 기초하여 상기 웨이트를 수정하고, 상기 피쳐익스트랙터부와 상기 도메인디스크리펜시부와 상기 시멘틱클러스터링부는, 수정된 상기 웨이트를 기초로 재학습되는 것을 특징으로 하며, 상기 연산부는, 기설정된 횟수로 상기 웨이트가 수정됨에도 불구하고 상기 전체손실의 변화가 기설정된 비율보다 작은 경우 상기 피쳐익스트랙터부와 상기 도메인디스크리펜시부와 상기 시멘틱클러스터링부의 학습을 중단시켜 최종 진단 알고리즘을 도출할 수 있다.In addition, the present invention, based on Equation 5, the total loss (

. , characterized in that re-learning is performed based on the corrected weight, and the calculating unit includes the feature extractor unit and the A final diagnostic algorithm may be derived by stopping the domain discrepancy unit and the semantic clustering unit learning.

{Formula 5}

은 전체손실, 상기

는 상기 분류손실, 상기

는 상기 도메인적응손실, 상기

는 상기 의미군집화손실, 상기

와

는 사용자가 기설정한 계수들)(remind

is the total loss, above

is the classification loss,

is the domain adaptation loss, the

is the semantic clustering loss,

Wow

is the coefficients set by the user)

{Formula 6}

은 수정된 상기 웨이트, 상기

는 수정전 상기 웨이트, 상기

는 사용자가 기설정한 상수, 상기

는 상기 웨이트에 대한 상기 전체손실의 미분값)(remind

is the modified said weight, said

is the weight before fertilization,

is a constant set by the user,

is the derivative of the total loss with respect to the weight)

In addition, the differential value of the total loss with respect to the weight may be calculated according to a back-propagation algorithm and a chain rule.

In addition, the label classifier unit corrects the weight based on Equation 7 and gradient descent, and is re-learned based on the modified weight, wherein the label classifier unit, Even though the weight is corrected a preset number of times, if the change in the classification loss is smaller than a preset ratio, learning is stopped, and a final machine learning-based state classification model may be derived.

{Formula 7}

은 수정된 상기 웨이트, 상기

는 수정전 상기 웨이트, 상기

는 사용자가 기설정한 상수, 상기

는 상기 웨이트에 대한 상기 분류손실의 미분값)(remind

is the modified said weight, said

is the weight before fertilization,

is a constant set by the user,

is the differential value of the classification loss with respect to the weight)

In addition, the differential value of the classification loss with respect to the weight may be calculated according to a back-propagation algorithm and a chain rule.

In addition, the present invention may include an output unit for outputting the state of the second system diagnosed using the final diagnosis algorithm.

In addition, the feature extractor unit uses a convolution neural network (CNN), and the label classifier unit may be provided as a fully connected layer.

The above object is, according to the present invention, a diagnosis method to which a domain adaptation and semantic clustering algorithm performed by a state diagnosis system of a mechanical system without label information is applied, including time series signal data measured and labeled in the first system A machine learning-based machine learning-based system in which the feature extractor calculates first features and second features from a second dataset including unlabeled time-series signal data measured in a second system a feature extractor step of learning a feature calculation model; A label classifier step in which the label classifier unit learns a machine learning-based state classification model that classifies the state of the first system using the first dataset, and calculates a classifier loss ( _LC ) ; Machine learning-based domain adaptation in which a domain dispensing unit performs domain adaptation by setting the first dataset as a source domain and setting the second dataset as a target domain Learning the model, domain-related loss (domain-related loss, _LD ) domain dispensing step of calculating; and a semantic clustering step of learning a machine learning-based clustering model in which a semantic clustering unit clusters by state based on the label of the first dataset, and calculating a semantic clustering loss (L _SC ), The classification loss is calculated based on Equation 1, the domain adaptation loss is calculated based on Equation 2, and the semantic clustering loss is calculated based on Equation 3. This is achieved by the applied diagnostic method.

delete

{Formula 1}

는 상기 분류손실, 상기 C는 상기 제1시스템의 상태의 수, 상기 Y는 라벨 벡터, 상기

는 상기 Y의 k번째 노드의 값, 상기 p는 상기 Y의 k번째 노드에 해당하는 확률벡터)(remind

is the classification loss, C is the number of states of the first system, Y is the label vector, and

is the value of the k-th node of Y, and p is a probability vector corresponding to the k-th node of Y)

{Formula 2}

와 상기

는 각각 상기 제1피쳐와 상기 제2피쳐)(wherein H is a universal reproducing kernel Hilbert space (RKHS), and n _S and n _T are the number of the first dataset and the second dataset, respectively, and

and above

is the first feature and the second feature, respectively)

{Formula 3}

는 상기 의미군집화손실, 상기

는 상기 시멘틱클러스터링단계의 특성층의 수, 상기

는 k번째 밸런싱 파라미터(balancing parameter), 상기

는 상기 제1데이터셋의 수, 상기

는 k번째 피쳐 레벨에서

번째 상기 제1데이터셋의 피쳐 벡터, 상기

는

번째 데이터셋과

번째 데이터셋이 같은 클래스에 속하면 1이고 다른 클래스라면 0인 행렬

의 값, 상기 d는 서로 다른 상태의 두 데이터셋 사이의 최소 거리를 제어하는 값)(remind

is the semantic clustering loss,

is the number of characteristic layers in the semantic clustering step,

is the kth balancing parameter (balancing parameter), the

is the number of the first dataset, the

is at the kth feature level

th feature vector of the first dataset, the

Is

the second dataset and

A matrix that is 1 if the second dataset belongs to the same class and 0 if it is a different class.

value of , where d is a value controlling the minimum distance between two datasets in different states)

In addition, in the feature extractor step, the domain discrepancy step, and the semantic clustering step, the first dataset and the second dataset are learned in mini-batch units, and the first dataset and When learning is performed on all small groups of the second dataset, small groups may be randomly re-extracted from the first dataset and the second dataset and re-learned.

In addition, the feature extractor step may calculate the first feature and the second feature based on Equation (4).

{Formula 4}

는 상기 제1데이터셋 또는 상기 제2데이터셋, 상기

는 상기

으로부터 계산된 상기 제1피쳐 또는 상기 제2피쳐, 상기

는 활성화함수(activation function), 상기

는 웨이트(weight), 상기 bias는 상수)(remind

is the first data set or the second data set, the

is said

The first feature or the second feature calculated from

is an activation function, the

is a weight, the bias is a constant)

)을 계산하는 연산단계를 더 포함하며, 상기 연산단계는, 수식 6 및 경사하강법(gradient descent)에 기초하여 상기 웨이트를 수정하고, 상기 피쳐익스트랙터단계와 상기 도메인디스크리펜시단계와 상기 시멘틱클러스터링단계는, 수정된 상기 웨이트를 기초로 재학습되는 것을 특징으로 하며, 상기 연산단계는, 기설정된 횟수로 상기 웨이트가 수정됨에도 불구하고 상기 전체손실의 변화가 기설정된 비율보다 작은 경우 상기 피쳐익스트랙터단계와 상기 도메인디스크리펜시단계와 상기 시멘틱클러스터링단계의 학습을 중단시켜 최종 진단 알고리즘을 도출할 수 있다.In addition, the present invention, based on Equation 5, the total loss (

. The clustering step is characterized in that re-learning is performed based on the modified weight, and the calculating step is performed when the change in the total loss is smaller than a predetermined ratio despite the weight being corrected a predetermined number of times. The final diagnostic algorithm can be derived by stopping the learning of the tractor step, the domain discrepancy step, and the semantic clustering step.

{Formula 5}

은 전체손실, 상기

는 상기 분류손실, 상기

는 상기 도메인적응손실, 상기

는 상기 의미군집화손실, 상기

와

는 사용자가 기설정한 계수들)(remind

is the total loss, above

is the classification loss,

is the domain adaptation loss, the

is the semantic clustering loss,

Wow

is the coefficients set by the user)

{Formula 6}

은 수정된 상기 웨이트, 상기

는 수정전 상기 웨이트, 상기

는 사용자가 기설정한 상수, 상기

는 상기 웨이트에 대한 상기 전체손실의 미분값)(remind

is the modified said weight, said

is the weight before fertilization,

is a constant set by the user,

is the derivative of the total loss with respect to the weight)

In addition, the differential value of the total loss with respect to the weight may be calculated according to a back-propagation algorithm and a chain rule.

In addition, the label classifier step is characterized in that the weight is corrected based on Equation 7 and gradient descent, and re-learning is performed based on the modified weight, the label classifier step Although the weight is modified a preset number of times, when the change in the classification loss is smaller than a preset ratio, learning is stopped, and a final machine learning-based state classification model can be derived.

{Formula 7}

은 수정된 상기 웨이트, 상기

는 수정전 상기 웨이트, 상기

는 사용자가 기설정한 상수, 상기

는 상기 웨이트에 대한 상기 분류손실의 미분값)(remind

is the modified said weight, said

is the weight before fertilization,

is a constant set by the user,

is the differential value of the classification loss with respect to the weight)

In addition, the differential value of the classification loss with respect to the weight may be calculated according to a back-propagation algorithm and a chain rule.

In addition, the present invention may include an output step of outputting the state of the second system diagnosed using the final diagnosis algorithm by an output unit.

In addition, the feature extractor step uses a convolution neural network (CNN), and the label classifier step may be provided as a fully connected layer.

According to the present invention, it is possible to efficiently diagnose the state of the target system without label information based on the diagnostic data acquired from the source system.

On the other hand, the effects of the present invention are not limited to the above-mentioned effects, and various effects may be included within the range obvious to those skilled in the art from the description below.

1 shows the architecture of a diagnosis method of a diagnosis apparatus to which domain adaptation and semantic clustering algorithms are applied according to an embodiment of the present invention;
2 is a diagram illustrating electrical coupling between components of a diagnostic apparatus to which domain adaptation and semantic clustering algorithms are applied according to an embodiment of the present invention;
3 is a diagram illustrating a distribution conceptual diagram of source domain and target domain data in a diagnostic apparatus to which a domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention;
4 is a flowchart illustrating a diagnosis method to which a domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention;
5 is a diagram illustrating a diagnosis flow of a diagnosis method to which a domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention;
6 is a diagram illustrating a mechanical system for diagnosing the performance of a diagnostic method to which a domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention;
7 is a comparison of the diagnostic accuracy of Scenario I for CWRU of another diagnostic method with the diagnostic method to which the domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention;
8 is a comparison of the diagnostic accuracy of Scenario I for XJTU-SY of the diagnostic method to which the domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention and another diagnostic method;
9 is a comparison of the diagnostic accuracy of Scenario II for CWRU and IMS of the diagnostic method to which the domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention and other diagnostic methods;
10 is a comparison of the diagnostic accuracy of Scenario II for IMS and XJTU-SY of the diagnostic method to which the domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention and other diagnostic methods;
11 is a diagram illustrating a visualization result of the operation of Scenario I for a CWRU of a diagnosis method different from the diagnosis method to which the domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention;
12 is a diagram showing the visualization results of the operation of Scenario II for CWRU and IMS of the diagnostic method to which the domain adaptation and semantic clustering algorithm is applied and other diagnostic methods according to an embodiment of the present invention;
13 is a diagram showing the diagnosis accuracy results improved by considering classification loss, domain adaptation loss, and semantic clustering loss in the diagnosis method to which the domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention;
14 is a diagram illustrating average SCI values of a diagnosis method to which a domain adaptation and semantic clustering algorithm is applied and another diagnosis method according to an embodiment of the present invention.

Hereinafter, some embodiments of the present invention will be described in detail with reference to exemplary drawings. In adding reference numerals to the components of each drawing, it should be noted that the same components are given the same reference numerals as much as possible even though they are indicated on different drawings.

In the description of the embodiment of the present invention, if it is determined that a detailed description of a related known configuration or function interferes with the understanding of the embodiment of the present invention, the detailed description thereof will be omitted.

In addition, in describing the components of the embodiment of the present invention, terms such as first, second, A, B, (a), (b), etc. may be used. These terms are only for distinguishing the elements from other elements, and the essence, order, or order of the elements are not limited by the terms.

Hereinafter, a diagnosis apparatus to which a domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied will be described in detail with reference to the accompanying drawings.

1 shows the architecture of a diagnostic method of a diagnostic apparatus to which domain adaptation and semantic clustering algorithms are applied according to an embodiment of the present invention, and FIG. 2 shows domain adaptation and semantic clustering algorithms according to an embodiment of the present invention. The electrical coupling between each component of the diagnostic apparatus is illustrated, and FIG. 3 is a conceptual diagram illustrating the distribution of source domain and target domain data in the diagnostic apparatus to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied.

As shown in FIG. 1 , the diagnosis apparatus 100 to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied includes a feature extractor unit 110 , a label classifier unit 120 , and a domain discriminator. It includes a pensi unit 130 , a semantic clustering unit 140 , an operation unit 150 , and an output unit 160 .

Here, the feature extractor unit 110 , the label classifier unit 120 , the domain discrepancy unit 130 , the semantic clustering unit 140 , and the operation unit 150 are one terminal such as a personal computer (PC). may be implemented as a separate device, and may be provided as separate devices according to functions and purposes.

The feature extractor unit 110 is configured from a first dataset including labeled time-series signal data measured in the first system and a second dataset including unlabeled time-series signal data measured in the second system. As a configuration for learning a machine learning-based feature calculation model that calculates a first feature and a second feature, it can be learned in mini-batch units of the first dataset and the second dataset described above. there is.

Here, the time series signal data may be provided as an acceleration signal or a vibration signal. However, as long as it is a signal that can distinguish the state of the system, it does not matter what kind of signal is provided.

In addition, when learning is performed on all subgroups of the aforementioned first dataset and the aforementioned second dataset, the feature extractor unit 110 randomly selects small groups from the aforementioned first dataset and the aforementioned second dataset. This can be re-extracted and re-learned.

Here, the feature extractor unit 110 calculates the above-described first feature and the above-described second feature based on Equation (4).

{Formula 4}

는 상기 제1데이터셋 또는 상기 제2데이터셋을 나타내고, 상술한

는 상술한

으로부터 계산된 상술한 제1피쳐 또는 상술한 제2피쳐를 나타내고, 상술한

는 활성화함수(activation function)를 나타내고, 상술한

는 웨이트(weight)를 나타내고, 상기 bias는 상수를 나타낸다.Here, the above

represents the first data set or the second data set, and

is the above

represents the above-mentioned first feature or the above-mentioned second feature calculated from

represents an activation function, and

denotes a weight, and the bias denotes a constant.

In addition, the feature extractor unit 110 may be re-learned based on the above-described weight modified by the operation unit 150 to be described later. Here, the operation unit 150, which will be described later, may also correct the bias, and in this case, the feature extractor unit 110 may be re-learned based on the modified weight and bias.

Also, the feature extractor unit 110 may be trained based on a convolution neural network (CNN).

As shown in FIG. 2 , the feature extractor unit 110 is electrically connected to the label classifier unit 120 , the domain discrepancy unit 130 , and the semantic clustering unit 140 , which will be described later. The feature may be transmitted, and it may be re-learned by being connected to the operation unit 150 and receiving the modified weight and bias. In addition, the label classifier unit 120 , the domain dispensing unit 130 , and the semantic clustering unit 140 , which will be described later, are electrically connected to each other to exchange signals. In addition, the operation unit 150 to be described later may be electrically connected to the feature extractor unit 110 and the label classifier unit 120 to be described later, the domain declarative unit 130 and the semantic clustering unit 140, The output unit 160 to be described later may be electrically connected to a calculation unit 150 to be described later and a label classifier unit 120 to be described later.

The label classifier unit 120 is configured to learn a machine learning-based state classification model for classifying the state of the first system by using the above-described first data set, and based on Equation 1, the classification loss (classifier) loss, L _C ).

{Formula 1}

는 상기 분류손실을 나타내고, 상술한 C는 상술한 제1시스템의 상태의 수를 나타내고, 상술한 Y는 라벨 벡터를 나타내고, 상술한

는 상술한 Y의 k번째 노드의 값을 나타내고, 상술한 p는 상술한 Y의 k번째 노드에 해당하는 확률벡터를 나타낸다.Here, the above

denotes the classification loss, the aforementioned C denotes the number of states of the aforementioned first system, the aforementioned Y denotes the label vector, and the aforementioned

denotes a value of the k-th node of Y described above, and p denotes a probability vector corresponding to the k-th node of Y described above.

Here, FIG. 3(a) is a conventional approach, FIG. 3(b) is a UDA approach, and FIG. 3(c) is a diagnosis apparatus 100 to which a domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied. A conceptual diagram of the distribution of source and target domain data with domain mismatch in the feature space trained using the DASC approach is shown, and the blue dashed line is a label classifier, which will be described later, trained using labeled source domain data in each feature space. Indicates the state determination boundary of the unit 120 . According to this, the label classifier unit 120 learns a criterion for classifying two different states corresponding to the blue dashed line, and diagnoses the state of the system accordingly. Also, circles and rectangles represent different classes, and colors represent label information. Unlabeled target domains are grayed out.

In addition, as shown in FIG. 3 , in many cases, the classification of target domain data is not performed correctly in the conventional approach. However, when the UDA method through the domain dispensing unit 130 is used, the source domain data and the target domain data are Classification accuracy is increased by having common features. However, classification of the data at the boundary between the two labels is still not done accurately. When the semantic clustering unit 140 of the diagnostic apparatus 100 to which the domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention is applied, the distance between the same labels is reduced and the distance between different labels is increased, as shown in FIG. As shown in (c), there is an effect that can significantly improve the classification accuracy.

Here, the label classifier unit 120 may correct the above-described weight based on Equation 7 and gradient descent, and re-learn based on the modified weight.

{Formula 7}

은 수정된 상술한 웨이트를 나타내고, 상술한

는 수정전 상술한 웨이트를 나타내고, 상술한

는 사용자가 기설정한 상수를 나타내고, 상술한

는 상술한 웨이트에 대한 상기 분류손실의 미분값을 나타낸다.Here, the above

denotes the above-mentioned weight modified, and

represents the above-mentioned weight before modification, and

represents a constant preset by the user, and

denotes the differential value of the classification loss with respect to the above-described weight.

Here, the differential value of the classification loss with respect to the weight may be calculated according to a back-propagation algorithm and a chain rule.

In addition, the label classifier unit 120 stops learning when the change in the classification loss is smaller than a preset rate even though the weight is corrected a preset number of times, so that a final machine learning-based state classification model can be derived. there is.

In addition, the label classifier unit 120 may be provided as a fully connected layer.

The domain dispensing unit 130 performs domain adaptation by setting the above-described first dataset as a source domain and setting the above-described second dataset as a target domain. As a configuration for learning a machine learning-based domain adaptation model, a domain-related loss ( _LD ) is calculated based on Equation 2. Here, domain adaptation is a research field that deals with the learning problem of a classifier or predictor when there is a difference between the learning distribution and the test distribution. That is, it aims to make the label classifier unit 120 learned in the source domain operate effectively in the target domain through the mapping between the source domain and the target domain.

{Formula 2}

는 상기 도메인적응손실을 나타내고, 상술한 H는 universal RKHS(reproducing kernel Hilbert space)를 나타내고, 상술한 n_S와 상술한 n_T는 각각 상기 제1데이터셋과 상기 제2데이터셋의 수를 나타내고, 상술한

와 상기

는 각각 상술한 제1피쳐와 상술한 제2피쳐를 나타낸다.Here, the above

denotes the domain adaptation loss, the above-mentioned H denotes a universal reproducing kernel Hilbert space (RKHS), the above-mentioned n _S and the above-mentioned n _T denote the number of the first dataset and the second dataset, respectively, the above

and above

denotes the aforementioned first feature and the aforementioned second feature, respectively.

In addition, the domain dispensing unit 130 is learned in units of a mini-batch of the above-described first dataset and the above-described second dataset, When learning is performed on all small groups, the small groups may be re-extracted randomly from the above-described first dataset and the above-described second dataset and re-learned.

Also, the domain description unit 130 may be re-learned based on the above-described weight modified by the operation unit 150 to be described later. Here, the operation unit 150 to be described later may also correct the bias, and in this case, the domain dispensing unit 130 may be re-learned based on the modified weight and bias.

The semantic clustering unit 140 is configured to learn a machine learning-based clustering model for clustering by state based on the labels of the first dataset described above, and based on Equation 3, semantic clustering loss (L _SC ) to calculate

{Formula 3}

는 상기 의미군집화손실을 나타내고, 상술한

는 상술한 시멘틱클러스터링의 특성층의 수를 나타내고, 상술한

는 k번째 밸런싱 파라미터(balancing parameter)를 나타내고, 상술한

는 상기 제1데이터셋의 수를 나타내고, 상술한

는 k번째 피쳐 레벨에서

번째 상술한 제1데이터셋의 피쳐 벡터를 나타내고, 상술한

는

번째 데이터셋과

번째 데이터셋이 같은 클래스에 속하면 1이고 다른 클래스라면 0인 행렬

의 값을 나타내고, 상술한 d는 서로 다른 상태의 두 데이터셋 사이의 최소 거리를 제어하는 값을 나타낸다.Here, the above

represents the semantic clustering loss, and

represents the number of characteristic layers of the above-described semantic clustering, and

represents the k-th balancing parameter (balancing parameter),

represents the number of the first data set, and

is at the kth feature level

th represents the feature vector of the first dataset described above,

Is

the second dataset and

A matrix that is 1 if the second dataset belongs to the same class and 0 if it is a different class.

represents a value of , and the above-mentioned d represents a value that controls the minimum distance between two datasets in different states.

Here, the semantic clustering unit 140 is trained in units of mini-batch of the aforementioned first dataset and the aforementioned second dataset, and all subgroups of the aforementioned first dataset and the aforementioned second dataset. When learning is performed, small groups are randomly re-extracted from each of the above-described first dataset and the above-described second dataset to be re-learned.

In addition, the semantic clustering unit 140 may be re-learned based on the above-described weight modified by the operation unit 150 to be described later. Here, the calculation unit 150, which will be described later, may also correct the bias, and in this case, the semantic clustering unit 140 may be re-learned based on the modified weight and bias.

)을 계산하는 구성으로, 수식 6 및 경사하강법(gradient descent)에 기초하여 상술한 웨이트를 수정한다.The calculating unit 150 calculates the total loss (

), the above-mentioned weight is corrected based on Equation 6 and gradient descent.

{Formula 5}

은 전체손실을 나타내고, 상술한

는 상술한 분류손실을 나타내고, 상술한

는 상술한 도메인적응손실을 나타내고, 상술한

는 상술한 의미군집화손실을 나타내고, 상술한

와

는 사용자가 기설정한 계수들을 나타낸다.Here, the above

represents the total loss, and

represents the above-described classification loss, and

represents the aforementioned domain adaptation loss, and

represents the above-mentioned semantic clustering loss, and

Wow

denotes coefficients preset by the user.

{Formula 6}

은 수정된 상기 웨이트를 나타내고, 상술한

는 수정전 상술한 웨이트를 나타내고, 상술한

는 사용자가 기설정한 상수를 나타내고, 상술한

는 상술한 웨이트에 대한 상술한 전체손실의 미분값을 나타낸다.Here, the above

denotes the modified weight, and

represents the above-mentioned weight before modification, and

represents a constant preset by the user, and

denotes the derivative of the above-mentioned total loss with respect to the above-mentioned weight.

Here, the differential value of the above-described total loss with respect to the above-described weight may be calculated according to a back-propagation algorithm and a chain rule.

In addition, when the change in the total loss is smaller than a preset ratio despite the weight being corrected a preset number of times, the calculating unit 150 includes the feature extractor unit 110 and the domain discrepancy unit 130 described above. And by stopping the learning of the above-described semantic clustering unit 140, it is possible to derive a final diagnostic algorithm.

The output unit 160 outputs the state of the second system diagnosed by using the final diagnosis algorithm and the final machine learning-based state classification model. Here, the output unit 160 may be provided in the form of a monitor, or may be provided in the form of a speaker that outputs a signal in sound.

As described above, including the feature extractor unit 110, the label classifier unit 120, the domain delimiter unit 130, the semantic clustering unit 140, the operation unit 150, and the output unit 160 According to the diagnosis apparatus 100 to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied, it is possible to efficiently diagnose the state of the target system without label information based on the diagnosis data obtained from the source system.

Hereinafter, the diagnosis method S100 to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied will be described in detail with reference to the accompanying drawings.

4 is a flowchart of a diagnosis method S100 to which a domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention, and FIG. 5 is a domain adaptation and semantic clustering algorithm to which an embodiment of the present invention is applied. The flow of diagnosis of the diagnosis method S100 is shown.

As shown in Fig. 4, the diagnosis method (S100) to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied is a feature extractor step (S110), a label classifier step (S120), and a domain discriminator. It includes a pensi step (S130) and a semantic clustering step (S140). Here, the label classifier step (S120), the domain discrepancy step (S130), and the semantic clustering step (S140) may be simultaneously performed. In addition, the feature extractor step (S110), the label classifier step (S120), the domain discrepancy step (S130), and the semantic clustering step (S140) may be learned, respectively, or may be separately learned.

The feature extractor step S110 is performed from a first dataset including labeled time-series signal data measured in the first system and a second dataset including unlabeled time-series signal data measured in the second system. The feature extractor unit 110 is a step of learning a machine learning-based feature calculation model for calculating a first feature and a second feature. -batch) unit can be learned.

In addition, in the feature extractor step ( S110 ), when learning is performed on all subgroups of the above-described first dataset and the above-described second dataset, randomly small groups in the above-described first dataset and the above-described second dataset, respectively This can be re-extracted and re-learned.

Here, in the feature extractor step ( S110 ), the above-described first feature and the above-described second feature are calculated based on Equation (4).

{Formula 4}

는 상기 제1데이터셋 또는 상기 제2데이터셋을 나타내고, 상술한

는 상술한

으로부터 계산된 상술한 제1피쳐 또는 상술한 제2피쳐를 나타내고, 상술한

는 활성화함수(activation function)를 나타내고, 상술한

는 웨이트(weight)를 나타내고, 상기 bias는 상수를 나타낸다.Here, the above

represents the first data set or the second data set, and

is the above

represents the above-mentioned first feature or the above-mentioned second feature calculated from

represents an activation function, and

denotes a weight, and the bias denotes a constant.

In addition, the feature extractor step ( S110 ) may be re-learned based on the above-described weight modified by an operation step ( S150 ) to be described later. Here, in the calculation step S150 to be described later, the bias may also be corrected, and in this case, the feature extractor step S110 may be re-learned based on the modified weight and bias.

Also, the feature extractor step S110 may be learned based on a convolution neural network (CNN).

As shown in Fig. 2, the feature extractor step (S110) is electrically connected to the label classifier step (S120), the domain discrepancy step (S130), and the semantic clustering step (S140), which will be described later. The feature can be delivered, and it can be re-learned by receiving the modified weight and bias in connection with the calculation step (S150). In addition, the label classifier step ( S120 ), the domain dispensing step ( S130 ), and the semantic clustering step ( S140 ), which will be described later, are electrically connected to each other to exchange signals. In addition, the calculation step (S150) to be described later may be electrically connected to the feature extractor step (S110), the label classifier step (S120), the domain dispensing step (S130), and the semantic clustering step (S140) to be described later. , an output step (S160) to be described later may be electrically connected to an operation step (S150) to be described later and a label classifier step (S120) to be described later.

The label classifier step (S120) is a step of learning a machine learning-based state classification model for classifying the state of the first system by the label classifier unit 120 using the above-described first dataset, Calculate the classifier loss ( _LC ) based on Equation 1.

{Formula 1}

는 상기 분류손실을 나타내고, 상술한 C는 상술한 제1시스템의 상태의 수를 나타내고, 상술한 Y는 라벨 벡터를 나타내고, 상술한

는 상술한 Y의 k번째 노드의 값을 나타내고, 상술한 p는 상술한 Y의 k번째 노드에 해당하는 확률벡터를 나타낸다.Here, the above

denotes the classification loss, the aforementioned C denotes the number of states of the aforementioned first system, the aforementioned Y denotes the label vector, and the aforementioned

denotes a value of the k-th node of Y described above, and p denotes a probability vector corresponding to the k-th node of Y described above.

Here, Fig. 3(a) is a conventional approach, Fig. 3(b) is a UDA approach, and Fig. 3(c) is a diagnosis method (S100) to which a domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied. A conceptual diagram of the distribution of source and target domain data with domain mismatch in the feature space trained using the DASC approach is shown, and the blue dashed line is a label classifier, which will be described later, trained using labeled source domain data in each feature space. Indicates the state determination boundary of step S120. According to this, the label classifier step (S120) learns a criterion for classifying two different states corresponding to the blue dashed line and diagnoses the state of the system accordingly. Also, circles and rectangles represent different classes, and colors represent label information. Unlabeled target domains are grayed out.

Here, the label classifier step S120 may correct the above-described weight based on Equation 7 and gradient descent, and re-learn based on the modified weight.

{Formula 7}

은 수정된 상술한 웨이트를 나타내고, 상술한

는 수정전 상술한 웨이트를 나타내고, 상술한

는 사용자가 기설정한 상수를 나타내고, 상술한

는 상술한 웨이트에 대한 상기 분류손실의 미분값을 나타낸다.Here, the above

denotes the above-mentioned weight modified, and

represents the above-mentioned weight before modification, and

represents a constant preset by the user, and

denotes the differential value of the classification loss with respect to the above-described weight.

Here, the differential value of the classification loss with respect to the weight may be calculated according to a back-propagation algorithm and a chain rule.

In addition, in the label classifier step (S120), even though the weight is corrected a preset number of times, if the change in the classification loss is smaller than a preset ratio, learning is stopped, so that a final machine learning-based state classification model can be derived. there is.

In addition, the label classifier step (S120) may be provided as a fully connected layer.

In the domain dispensing step (S130), the domain dispensing unit 130 sets the above-described first data set as a source domain and the above-described second data set as a target domain. As a step of learning a machine learning-based domain adaptation model performing domain adaptation, a domain-related loss ( _LD ) is calculated based on Equation 2.

{Formula 2}

는 상기 도메인적응손실을 나타내고, 상술한 H는 universal RKHS(reproducing kernel Hilbert space)를 나타내고, 상술한 n_S와 상술한 n_T는 각각 상기 제1데이터셋과 상기 제2데이터셋의 수를 나타내고, 상술한

와 상기

는 각각 상술한 제1피쳐와 상술한 제2피쳐를 나타낸다.Here, the above

denotes the domain adaptation loss, the above-mentioned H denotes a universal reproducing kernel Hilbert space (RKHS), the above-mentioned n _S and the above-mentioned n _T denote the number of the first dataset and the second dataset, respectively, the above

and above

denotes the aforementioned first feature and the aforementioned second feature, respectively.

In addition, in the domain dispensing step ( S130 ), the above-described first dataset and the above-described second dataset are learned in mini-batch units, and the above-described first dataset and the above-described second dataset are learned. When learning is performed on all small groups, the small groups may be re-extracted randomly from the above-described first dataset and the above-described second dataset and re-learned.

In addition, the domain discrepancy step ( S130 ) may be re-learned based on the above-described weight modified by the operation step ( S150 ) to be described later. Here, in the calculation step ( S150 ) to be described later, the bias may also be corrected, and in this case, the domain dispensing step ( S130 ) may be re-learned based on the modified weight and bias.

The semantic clustering step (S140) is a step in which the semantic clustering unit 140 learns a machine learning-based clustering model that clusters by state based on the label of the first dataset described above. Based on Equation 3, the semantic clustering loss ( Calculate the semantic clustering loss, L _SC ).

{Formula 3}

는 상기 의미군집화손실을 나타내고, 상술한

는 상술한 시멘틱클러스터링의 특성층의 수를 나타내고, 상술한

는 k번째 밸런싱 파라미터(balancing parameter)를 나타내고, 상술한

는 상기 제1데이터셋의 수를 나타내고, 상술한

는 k번째 피쳐 레벨에서

번째 상술한 제1데이터셋의 피쳐 벡터를 나타내고, 상술한

는

번째 데이터셋과

번째 데이터셋이 같은 클래스에 속하면 1이고 다른 클래스라면 0인 행렬

의 값을 나타내고, 상술한 d는 서로 다른 상태의 두 데이터셋 사이의 최소 거리를 제어하는 값을 나타낸다.Here, the above

represents the semantic clustering loss, and

represents the number of characteristic layers of the above-described semantic clustering, and

represents the k-th balancing parameter (balancing parameter),

represents the number of the first data set, and

is at the kth feature level

th represents the feature vector of the first dataset described above,

Is

the second dataset and

A matrix that is 1 if the second dataset belongs to the same class and 0 if it is a different class.

represents a value of , and the above-mentioned d represents a value that controls the minimum distance between two datasets in different states.

Here, the semantic clustering step S140 is learned in units of mini-batch of the aforementioned first dataset and the aforementioned second dataset, and all subgroups of the aforementioned first dataset and the aforementioned second dataset When learning is performed, small groups are randomly re-extracted from each of the above-described first dataset and the above-described second dataset to be re-learned.

In addition, the semantic clustering step ( S140 ) may be re-learned based on the above-described weight modified by the operation step ( S150 ) to be described later. Here, in the calculation step S150 to be described later, the bias may also be corrected, and in this case, the semantic clustering step S140 may be re-learned based on the modified weight and bias.

)을 계산하는 단계로, 수식 6 및 경사하강법(gradient descent)에 기초하여 상술한 웨이트를 수정한다.Calculation step (S150) is based on Equation 5, the total loss (

), the above-mentioned weight is corrected based on Equation 6 and gradient descent.

{Formula 5}

은 전체손실을 나타내고, 상술한

는 상술한 분류손실을 나타내고, 상술한

는 상술한 도메인적응손실을 나타내고, 상술한

는 상술한 의미군집화손실을 나타내고, 상술한

와

는 사용자가 기설정한 계수들을 나타낸다.Here, the above

represents the total loss, and

represents the above-described classification loss, and

represents the aforementioned domain adaptation loss, and

represents the above-mentioned semantic clustering loss, and

Wow

denotes coefficients preset by the user.

{Formula 6}

은 수정된 상기 웨이트를 나타내고, 상술한

는 수정전 상술한 웨이트를 나타내고, 상술한

는 사용자가 기설정한 상수를 나타내고, 상술한

는 상술한 웨이트에 대한 상술한 전체손실의 미분값을 나타낸다.Here, the above

denotes the modified weight, and

represents the above-mentioned weight before modification, and

represents a constant preset by the user, and

denotes the derivative of the above-mentioned total loss with respect to the above-mentioned weight.

Here, the differential value of the above-described total loss with respect to the above-described weight may be calculated according to a back-propagation algorithm and a chain rule.

In addition, in the calculation step S150, the above-described feature extractor step S110 and the above-described domain dispensing step S130 are performed when the change in the total loss is smaller than a predetermined ratio even though the weight is corrected a predetermined number of times. ) and the above-described semantic clustering step (S140) can be stopped to derive a final diagnostic algorithm.

The output step S160 is a step in which the state of the second system diagnosed by using the final diagnosis algorithm and the final machine learning-based state classification model is output by the output unit 160 . Here, the output unit 160 may be provided in the form of a monitor, or may be provided in the form of a speaker that outputs a signal in sound.

Next, the principle of the diagnosis method S100 to which the domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention will be described in detail.

6 shows a mechanical system for diagnosing the performance of the diagnostic method S100 to which the domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention, and FIG. The diagnostic accuracy of Scenario I for CWRU of another diagnostic method is compared with the diagnostic method (S100) to which the semantic clustering algorithm is applied and FIG. 8 is a diagnostic method to which the domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention ( S100) and the diagnostic accuracy of Scenario I for XJTU-SY of other diagnostic methods are compared. A comparison of the diagnostic accuracy of Scenario II for CWRU and IMS of A comparison of the diagnostic accuracy of Scenario II for The visualization results are shown, and FIG. 12 is a visualization result for the operation of Scenario II for CWRU and IMS of the diagnostic method (S100) to which the domain adaptation and semantic clustering algorithm is applied and other diagnostic methods according to an embodiment of the present invention. 13 shows the diagnostic accuracy results in which the diagnosis method S100 to which the domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention is improved by considering classification loss, domain adaptation loss, and semantic clustering loss. 14 shows the average SCI values of the diagnostic method S100 to which the domain adaptation and semantic clustering algorithm is applied and other diagnostic methods according to an embodiment of the present invention.

)만 있는 타겟 도메인에 대한 진단 모델은 다른(그러나 관련이 있는) 소스 도메인에서 얻은 레이블이 지정된 데이터(

)의 정보를 전송하여 개선된다. 여기서

는 소스 샘플의 수를 나타내고,

는 타겟 샘플의 수를 나타낸다. 다른 도메인은 다른 시스템이나 다른 작동 조건에서 얻은 데이터로 생각할 수 있다. 이러한 다른 도메인의 경우 도메인 불일치 또는 도메인 이동(예: D_S ≠ D_T)이 있다. 본 발명에서의 두 도메인은 동일한 레이블 공간(y (= y_S = y_T))을 갖는다. 이것은 두 도메인이 동일한 유형의 작동 상태를 가지고 있음을 의미한다. 동일한 유형의 오류 모드를 가진 유사한 시스템의 데이터가 전송되고 활용되기 때문에 이는 합리적이다. 이러한 종류의 전이 학습 문제를 UDA 문제라고 한다. 도메인 불일치로 인해 D_S에서 직접 얻은 레이블이 지정된 데이터 세트로 훈련된 진단 모델을 D_T에 적용할 수 없다. 이러한 문제를 해결하기 위해 UDA 알고리즘을 활용하여 도메인 불일치를 최소화하고 D_T에서 일반화 성능을 극대화할 수 있다. 즉, UDA 알고리즘의 주요 목표는 두 도메인 간의 불일치를 최소화하여 두 도메인에서 공유할 수 있는 피쳐의 공통 표현을 결정하는 것이다. 그런 다음 이러한 공유 피쳐를 기반으로 D_S의 레이블 정보를 사용하여 두 도메인 모두에서 잘 작동하는 결함 진단을 위한 모델을 개발할 수 있다. 이 학습 전략은 다음 공식으로 확인할 수 있다.The diagnostic method (S100) to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied is a target domain (D _T ) in which label information cannot be obtained by using the label data obtained from the source domain ( _DS ). It aims to improve diagnostic performance. That is, unlabeled data (

), a diagnostic model for a target domain with only labeled data (

) is improved by sending the information of here

represents the number of source samples,

represents the number of target samples. Different domains can be thought of as data obtained from different systems or different operating conditions. For these other domains, there is a domain mismatch or domain shift (eg D _S ≠ D _T ). Both domains in the present invention have the same label space (y (= y _S = y _T )). This means that both domains have the same type of operational state. This makes sense because data from similar systems with the same type of failure mode is transmitted and utilized. This kind of transfer learning problem is called a UDA problem. A diagnostic model _trained on labeled datasets obtained directly from DS cannot be applied to _DT due to domain mismatch. To solve this problem, the UDA algorithm can be utilized to minimize the domain mismatch and _maximize the generalization performance in DT. In other words, the main goal of the UDA algorithm is to determine a common representation of features that can be shared by both domains by minimizing the mismatch between the two domains. Then, based on these shared features, we can use the label information in _DS to develop a model for fault diagnosis that works well in both domains. This learning strategy can be confirmed by the following formula.

는

에서 라벨클래시파이어부(120) η를 사용한 타겟 리스크를 나타내며,

로 표현할 수 있다.

는

에서 라벨클래시파이어부(120) η를 사용한 소스 리스크를 나타낸다.

는 도메인 불일치를 의미하며,

와

에서 파생된 샘플에 대한 경험적 발산을 나타낸다. 그리고 Const.는 모델 복잡도와 표본 크기에 의해 결정되는 상수항을 나타낸다. 즉, 타겟 영역에 대해 잘 작동하는 진단 모델을 결정하려면 소스 위험과 영역 불일치를 동시에 최소화해야 한다.where η denotes a trained classifier belonging to the hypothesis class (H).

Is

represents the target risk using the label classifier unit 120 η,

can be expressed as

Is

In the label classifier 120 represents the source risk using η.

means domain mismatch,

Wow

represents the empirical divergence for samples derived from And Const. represents a constant term determined by model complexity and sample size. In other words, to determine a diagnostic model that works well for the target area, the source risk and area mismatch must be minimized simultaneously.

In the diagnosis method S100 to which the domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention, a diagnosis technique using a DASC (Domain Adaptation with Semantic Clustering) approach is proposed. This approach is designed to improve the generalization performance in the target domain by learning more discriminated domain invariant properties.

) 또는 레이블이 지정된 훈련 데이터의 예측 레이블과 실제 레이블 간의 오류를 최소화하기 위해 지도 학습을 수행한다. 본 발명의 일실시예에 따른 도메인적응 및 의미군집화 알고리즘이 적용된 진단 방법(S100)에서는 시스템의 결함을 진단하는 지능적인 방법을 얻기 위해 CNN을 채택했다. CNN은 합성곱 계층과 풀링 계층의 집합을 기반으로 하는 피쳐익스트랙터부(110)와 완전 연결 계층을 기반으로 하는 라벨클래시파이어부(120)로 구성된다. 피쳐익스트랙터부(110)와 라벨클래시파이어부(120)를 통해 입력 데이터(X)는 다음과 같이 출력 벡터(

)로 변환된다.The general goal of diagnostic techniques is to obtain a good fault diagnosis model in which the input data (X) and the corresponding label vector (Y) exactly match. To develop a diagnostic model, the classification loss (

) or perform supervised learning to minimize the error between the predicted labels in the labeled training data and the actual labels. In the diagnosis method S100 to which the domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention, CNN is adopted to obtain an intelligent method for diagnosing system defects. CNN is composed of a feature extractor unit 110 based on a set of convolutional layers and pooling layers and a label classifier unit 120 based on a fully connected layer. Through the feature extractor unit 110 and the label classifier unit 120, the input data (X) is an output vector (

) is converted to

및

는 각각 피쳐익스트랙터부(110) 및 라벨클래시파이어부(120) 내의 매개변수를 나타낸다. 또한,

와

는 피쳐익스트랙터부(110)와 라벨클래시파이어부(120)의 변환 함수를 나타낸다.here

and

denotes parameters in the feature extractor unit 110 and the label classifier unit 120, respectively. In addition,

Wow

denotes a conversion function of the feature extractor unit 110 and the label classifier unit 120 .

의 차원은 진단할 상태의 수량과 동일해야 한다. 그런 다음 softmax 함수를 사용하여 확률 벡터(p)를 다음과 같이 출력한다.For multi-class diagnostic models, it represents the number of nodes in the output layer.

The dimension of should be equal to the quantity of the condition to be diagnosed. Then, using the softmax function, the probability vector (p) is output as follows.

와

는 벡터 p와

의 j번째 노드를 의미한다. 여기서,

값은 데이터 X가 C 클래스들 중 j번째 상태 조건에 속할 확률로 해석될 수 있다.where C represents the number of states of the target system. In addition,

Wow

is the vector p and

It means the j-th node of here,

The value can be interpreted as the probability that data X belongs to the jth state condition among C classes.

)을 추정하기 위해 교차 엔트로피 손실을 사용하며, 이는 다음과 같이 나타낼 수 있다.In order to train a diagnostic model, the classification loss (

) is used to estimate the cross-entropy loss, which can be expressed as

는 타겟 라벨 Y의 k번째 노드의 값을 나타낸다. 이와 같이 정의된 분류손실을 줄임으로써 레이블이 지정된 소스 도메인 데이터를 올바른 레이블로 잘 분류하는 좋은 진단 모델을 개발할 수 있다.here,

denotes the value of the k-th node of the target label Y. By reducing the classification loss defined in this way, a good diagnostic model can be developed that classifies the labeled source domain data well into the correct label.

와

에서 공유할 수 있는 공통 피쳐를 학습하기 위해 불일치 기반 방법을 채택했다. 이 학습 방식의 전략은 먼저 두 도메인 간의 차이를 수량화하는 메트릭을 정의한 다음 해당 메트릭을 최소화하는 방식으로 피쳐를 학습한다. 여기서, MMD(Maximum Mean Disrepancy)는 서로 다른 도메인의 분포에 따른 불일치를 추정하는 데 사용된다. 이 비모수적 측정은 RKHS(Universal reproducing kernel Hilbert space)에서 함수에 대한 기대값의 가장 큰 차이로 결정될 수 있으며, 다음과 같이 표현할 수 있다.A domain invariant representation is needed to solve the UDA problem by sending the learned label classifier using the source data to the target domain.

Wow

We adopted a discrepancy-based method to learn common features that can be shared in The strategy of this learning method is to first define a metric that quantifies the difference between two domains, and then learn the features in a way that minimizes that metric. Here, MMD (Maximum Mean Disrepancy) is used to estimate the discrepancy according to the distribution of different domains. This non-parametric measurement can be determined as the largest difference between expected values for functions in RKHS (Universal reproducing kernel Hilbert space), and can be expressed as follows.

는 H의 단위 공 함수 클래스 내의 함수를 나타낸다. x_S와 x_T는 확률 분포(p_S 및 p_T)의 임의의 변수를 나타낸다. 모집단 기대치를 샘플에 대해 계산된 경험적 기대치로 대체하면 MMD의 경험적 추정치를 찾을 수 있다.Here, H stands for universal RKHS.

denotes a function within the unit empty function class of H. x _S and x _T represent arbitrary variables of the probability distribution (p _S and p _T ). An empirical estimate of the MMD can be found by replacing the population expectation with the empirical expectation calculated for the sample.

와

는 분포 p_S 및 p_T에서 추출된 샘플을 나타낸다. n_S와 n_T는 표본 X_S와 X_T의 개수를 나타낸다. RKHS의 커널 평균 임베딩을 기반으로 MMD 제곱의 경험적 추정은 다음과 같이 커널 함수로써 계산될 수 있다.here,

Wow

represents samples extracted from the distributions p _S and p _T . n _S and n _T represent the number of samples X _S and X _T . Based on the kernel mean embedding of RKHS, the empirical estimate of the MMD square can be calculated as a kernel function as follows.

는 x부터 H까지의 피쳐 매핑(

)간 내적(inner product)으로 정의되는 커널 함수를 나타낸다. 본 발명의 일실시예에 따른 도메인적응 및 의미군집화 알고리즘이 적용된 진단 방법(S100)에서는 범용 커널로 검증된 가우시안 커널(

)을 사용하여 MMD 제곱을 계산한다. MMD는 UDA 결과와 마찬가지로

의 값에 따라 다르다. 본 발명의 일실시예에 따른 도메인적응 및 의미군집화 알고리즘이 적용된 진단 방법(S100)에서는 MMD 기반 UDA 방법의 견고성과 성능을 향상시키기 위해 서로 다른

값을 가진 서로 다른 커널을 활용하는 다중 커널 MMD를 사용한다. 마지막으로 이 MMD 메트릭을 사용하여 두 도메인(소스 및 타겟) 간의 도메인적응손실(L_D)을 얻을 수 있다. 도메인적응손실은 두 도메인의 데이터에 대한 피쳐익스트랙터단계(S110)의 출력(

와

) 간의 MMD로 계산할 수 있다. 이 도메인적응손실은 UDA 작업을 수행하기 위해 도메인 불변 피쳐를 학습하는 데 필수적이다.here,

is the feature mapping from x to H (

) represents a kernel function defined as an inner product. In the diagnosis method (S100) to which the domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention, the Gaussian kernel (

) to calculate the MMD squared. MMD as well as UDA results

depends on the value of In the diagnosis method S100 to which the domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention, in order to improve the robustness and performance of the MMD-based UDA method, different

Use a multi-kernel MMD that utilizes different kernels with different values. Finally, we can use this MMD metric to obtain the domain adaptation loss ( _LD ) between the two domains (source and target). The domain adaptation loss is the output (

Wow

) can be calculated as the MMD of the liver. This domain adaptation loss is essential for learning domain invariant features to perform UDA tasks.

The existing UDA method aims to find domain invariant properties by reducing both classification loss and domain adaptation loss. However, these existing UDA methods only consider the alignment of the marginal distributions of different domain data. That is, the separability of features according to classes is not considered. Therefore, it is not possible to consistently learn identification features with significant inter-class differences and small intra-class variations. This results in low target domain generalization performance. To cope with this, the semantic clustering step (S140) of the diagnosis method (S100) to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied is a metric for learning a discriminant expression that semantically clusters samples of each class well. We propose a learning-based approach. This can be achieved by sufficiently segregating differently classified data while mapping the same class of data closely to each other. To this end, we propose an additional loss term called semantic clustering loss (L _SC ) that can be expressed as follows.

는 k번째 피쳐 레벨에서

번째 소스 샘플(

)의 피쳐 벡처를 나타낸다.

는

번째 데이터셋과

번째 데이터셋이 같은 클래스에 속하면 1이고 다른 클래스라면 0인 행렬

의 값을 나타낸다.

는 시멘틱클러스터링단계(S140)의 특성층의 수를 나타낸다.

는 k번째 밸런싱 파라미터(balancing parameter)를 나타낸다.

는 제1데이터셋의 수를 나타낸다. d는 서로 다른 상태의 두 데이터셋 사이의 최소 거리를 제어하는 값을 나타낸다.here,

is at the kth feature level

second source sample (

) represents the feature vector.

Is

the second dataset and

A matrix that is 1 if the second dataset belongs to the same class and 0 if it is a different class.

represents the value of

denotes the number of characteristic layers in the semantic clustering step (S140).

denotes a k-th balancing parameter (balancing parameter).

denotes the number of the first data set. d represents a value that controls the minimum distance between two datasets in different states.

=1)의 샘플 쌍에 대해서 상호간 거리는 줄어든다. 그러나 다른 클래스(

=0)의 샘플 쌍에 대한 상호간 거리는 증가한다. UDA 문제에서는 소스 도메인 레이블 정보만 사용할 수 있기 때문에 이 유사성 메트릭 기반 학습은 소스 데이터를 사용하여 수행된다. 두 도메인이 관련되어 있으므로 소스 데이터에 대한 의미군집화손실을 사용하는 본 발명의 일실시예에 따른 도메인적응 및 의미군집화 알고리즘이 적용된 진단 방법(S100)은 타겟 데이터도 각 클래스별로 잘 클러스터링되도록 한다. 또한, 본 발명의 일실시예에 따른 도메인적응 및 의미군집화 알고리즘이 적용된 진단 방법(S100)에서 의미군집화손실은 다중 피쳐 레벨에서 적용된다. 결과적으로 더 구별되는 특징 표현을 얻을 수 있으므로 시스템 상태에 따라 샘플을 의미적으로 더 잘 클러스터링할 수 있다. 이러한 시멘틱클러스터링단계(S140)에 의하면, 각 클래스의 경계 근처에 위치한 타겟 샘플에 대한 오분류율이 감소하는 효과가 있다. 결론적으로, 본 발명의 일실시예에 따른 도메인적응 및 의미군집화 알고리즘이 적용된 진단 방법(S100)은 타겟 시스템의 진단 성능을 향상시킬 수 있다.L _SC is calculated based on the similarity distance between all source samples to obtain robust and semantically well clustered features. By reducing L _SC in the learning process, the same class (

=1), the distance between them decreases. However, other classes (

= 0) the inter-reciprocal distance for the sample pair increases. Since only the source domain label information is available in the UDA problem, this similarity metric-based learning is performed using the source data. Since the two domains are related, the diagnosis method ( S100 ) to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention using the semantic clustering loss for the source data is applied also allows the target data to be well clustered for each class. In addition, in the diagnosis method S100 to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied, the semantic clustering loss is applied at the multi-feature level. As a result, more distinct feature representations can be obtained, so that samples can be semantically better clustered according to the state of the system. According to this semantic clustering step (S140), there is an effect of reducing the misclassification rate of the target sample located near the boundary of each class. In conclusion, the diagnosis method S100 to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied can improve the diagnostic performance of the target system.

According to the diagnosis method ( S100 ) to which the domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention, an accurate and powerful diagnostic model for the target domain can be obtained without label information. The data flow and architecture of the DASC proposed by the diagnosis method S100 to which the domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention is shown in FIG. 1 .

First, the feature extractor step S110 consists of two groups of convolutional and pooling layers. A 1D kernel with zero padding and ReLU is used for all convolutional layers. Here, maximum pooling is used for all pooling layers (layers).

In the label classifier step (S120), a fully connected network composed of two layers having dimensions of 256 and C is used, where C represents the number of states. The final objective function of DASC, that is, the total loss, is expressed as

은 전체손실,

는 분류손실,

는 도메인적응손실,

는 의미군집화손실을 의미한다.

와

는 사용자가 기설정한 계수들, 즉, balancing parameter를 나타낸다.here,

is the total loss,

is the classification loss,

is the domain adaptation loss,

is the semantic clustering loss.

Wow

represents the coefficients preset by the user, that is, the balancing parameter.

Based on the labeled source data, the classification loss is calculated only in the output layer of the label classifier step S120 as shown in FIG. 1 , and the formula is as follows.

In addition, the semantic clustering loss is calculated in all pooling layers of the feature extractor step S110 as shown in FIG. 1 , and the formula is as follows.

와

에서의 제1피쳐와 제2피쳐 간의 MMD로 계산되며, 도 1에 도시된 바와 같이, 마지막 컨볼루션 및 풀링 계층의 결과로 발견되고, 수식은 아래와 같다.The domain adaptation loss that must be minimized to obtain domain invariant properties is

Wow

It is calculated as the MMD between the first feature and the second feature in , and is found as a result of the last convolution and pooling layer, as shown in FIG. 1 , the formula is as follows.

의 레이블 정보를 이용하여 학습한 결함 진단 모델을

에 사용할 수 있으며, 피쳐를 클래스에 따라 의미적으로 잘 클러스터링하여 보다 구별적인 파쳐를 학습할 수 있다.by minimizing the total loss.

The fault diagnosis model learned using the label information of

can be used, and more distinct features can be learned by semantically clustering features according to classes.

Here, the classification loss and the semantic clustering loss use the label data information of the source domain, but the purpose and effect of the classification loss and the semantic clustering loss are different. In the case of classification loss, only the location of the labeled data is considered based on the hyperplane of the label classifier step (S120). That is, it aims to correctly classify the source domain label-specified data in the diagnostic method S100 to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied.

However, semantic clustering loss takes into account the distribution of labeled data, which is the relative distance between samples according to class, to produce semantically better clustered features. As a result, based on the DASC method proposed in the diagnosis method (S100) to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied in consideration of classification loss, domain adaptation loss, and semantic clustering loss, defect diagnosis with high target generalization performance model can be achieved.

로 설정한다. 또한, 다중 커널 MMD를 계산하기 위한

의 값으로 1e-6에서 1e6 사이의 10의 거듭제곱을 사용한다. 또한, 분류손실과 도메인적응손실과 의미군집화손실을 계산하기 위한 balancing parameter인

와

와

는 1e-5와 1e-1 사이의 값으로 선택할 수 있다. 본 발명의 일실시예에 따른 도메인적응 및 의미군집화 알고리즘이 적용된 진단 방법(S100)를 위한 더 상세한 정보는 아래 표와 같이 마련될 수 있다.The diagnostic method S100 to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied trains a diagnostic model using a mini-batch gradient descent method. Here, the learning rate decreases as p increases from 0 to 1.

set to In addition, to calculate the multi-kernel MMD,

Use a power of 10 between 1e-6 and 1e6 as the value of . In addition, it is a balancing parameter for calculating classification loss, domain adaptation loss, and semantic clustering loss.

Wow

Wow

can be selected as a value between 1e-5 and 1e-1. More detailed information for the diagnosis method S100 to which the domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention may be provided as shown in the table below.

A flowchart of a DASC-based error diagnosis method of the diagnosis method S100 to which the domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention is shown in FIG. 5 . As shown in FIG. 5 , the source dataset and the target dataset respectively measured in the source domain and the target domain are prepared as a training dataset. Thereafter, the user starts learning by inputting a set value to be written and inputting the training dataset into the diagnosis method S100 to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied. Classification loss, domain adaptation loss, and semantic clustering loss are calculated through the diagnostic method (S100) to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied, and the sum total loss and classification loss are minimized. CNN learning is performed by correcting the above-described weights and biases. Here, the direction of minimizing the total loss and the classification loss can be provided by learning based on Equation 7 described above. In addition, when the calculated total loss and classification loss are maintained to be smaller than the preset ratio despite the preset number of times of learning, the diagnosis method to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied (S100) ) can end learning and derive the final learning model. The state of the target system can be diagnosed by inputting the target signal to the derived final DASC learning model.

As described above, the diagnosis method S100 to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied is an error diagnosis model that can be applied to other unlabeled domains using labeled source data. to get By minimizing the final objective function, an invariant representation of the discriminant region can be obtained. Accordingly, there is an effect that the target domain diagnosis performance is improved.

The diagnosis method S100 to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied was applied to the bearing data obtained in the system shown in FIG. 6 . Figure 6 (a) shows a ball bearing system of CWRU (Case Western Reserve University), Figure 6 (b) shows a spherical roller bearing system provided by IMS (Center for Intelligent Maintenance Systems), Figure 6(c) shows the ball bearing system of Xi'an Jiaotong University, Institute of Design Science and Basic Component, Changxing Sumyoung Technology Co., Ltd. (XJTU-SY). Acceleration signals were measured in four types of diagnostic conditions: normal (N), ball defect (B), internal orbital defect (IR), and external orbital defect (OR). For the CWRU, 0.007, 0.014, and 0.021 inch diameter defects were applied to the bearings through electrical discharge machining. For IMS and XJTU-SY, error data were obtained through run-to-failure experiments. In addition, CWRU and XJTU-SY data were acquired under various operating conditions. The information of the data obtained from each system is shown in the table below.

In order to prove the effectiveness of the diagnostic method S100 to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied, two UDA-based error diagnosis scenarios were considered. First, we considered the problem of UDA between domains with inconsistencies caused by different operating conditions (Scenario I). Next, we considered when the source data and the target data were collected on separate rotating machines (Scenario II).

In all experiments, the data set for each diagnostic condition consisted of 4000 samples of raw vibration signals with a length of 1200 points. Among them, 70% of the data sample was used for training, and 30% was allocated for use in the test phase to estimate diagnostic performance. Data obtained from both domains were used to train a UDA-based defect diagnosis model. The target test sample was used to evaluate the target domain generalization performance of these UDA-based diagnostic models.

In order to verify the efficiency of the diagnosis method (S100) to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied, the target region diagnosis performance of the DASC and the existing method was investigated. First, for validation of the UDA algorithm, CNNs trained using only source data were compared. Next, we compared the results with several existing neural network-based UDA algorithms used for fault diagnosis. DANN was adopted as a representative method of the adversarial adaptive approach, and Deep CORAL, which learns desirable features by reducing the difference in covariance of features in two domains, was compared as a discrepancy-based method.

In addition, methods for minimizing multi-kernel MMD between source and target features were compared. Methods using both adversarial and disagreement-based strategies were also compared. This method is abbreviated as Adv+Disc in the remainder of the text. The same CNN architecture was adopted as the backbone model of all comparison methods in order to fairly and effectively compare the generalization ability, relying only on the UDA method used. In addition, the domain classifier for the adversarial adaptation strategy consists of a 2-layer fully connected network containing 256 and 2 nodes. The balancing parameter for DANN was selected from {0,0.5,1.0,1.5,2.0}, and the parameter for MMD was selected within the range of 1e-5 to 1e-1. Adversarial strategy and disagreement basis For the method using strategies simultaneously, the same parameters as DANN and MMD were used, and for Deep CORAL, parameters were selected in the range of 1e-6 to 1e-2. In addition, for fair comparison, all comparative UDA methods made the same optimizer, learning rate, training epoch, batch size, and momentum as well as training data and test data. Also, testing of each method was performed on the same computer with NVIDIA GeForce RTX 2080 Ti, Intel i7-8700, 16GB RAM.

First, a diagnosis (Scenario I) between domains with different operating conditions (eg load and rotational speed) is described. To verify the performance of the DASC under various operating conditions, we use the CWRU and XJTU-SY data sets obtained with vibration signals under various operating conditions. First, 12 UDA operations (1→2/3/4, 2→1/3/4, 3→1/2/4, 4→1,2) using different source and target domains for the CWRU data set. ,3, where the left means the source domain and the right means the target domain), the proposed method was evaluated and compared. Here, 1, 2, 3, and 4 mean different operating conditions of the CWRU, and the results are as shown in FIG. 7 . Here, FIG. 7(a) shows 1→2/3/4, FIG. 7(b) shows 2→1/3/4, FIG. 7(c) shows 3→1/2/4, and FIG. 7( d) means 4→1,2,3. Deviations due to defect size within each state condition were neglected to focus on discrepancies due to different operating conditions in the experiment. That is, samples of the same state with different defect sizes were grouped into the same classification category. There are four types of states (N, B, IR, OR) in this data set. Therefore, C, meaning the number of states, is 4. The diagnostic accuracy for the target sample under various operating conditions using the data set provided by the CWRU is shown in FIG. 7 . A value in the graph shown in FIG. 7 means an average of diagnostic accuracy in three different target domains, and error bars indicate standard error values.

As shown in FIG. 7 , it can be seen that the DASC method applied in the diagnosis method S100 to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied shows higher diagnostic accuracy than other diagnostic methods.

Next, the diagnostic performance of various UDA methods was evaluated using the XJTU-SY data set. Six diagnoses were performed in this dataset, including 1→2/3, 2→1/3, and 3→1/2. In the XJTU-SY1 and XJTU-SY2 data sets, there are only three types of health states (that is, N, IR, OR), so the number of states, C, is set to 3. The diagnosis result of XJTU-SY is shown in FIG. 8 . Here, FIG. 8(a) means 1→2/3, FIG. 8(b) means 2→1/3, and FIG. 8(c) means 3→1/2. As in FIG. 7 , averages of diagnostic accuracies for two different target domains are shown, and error bars indicate standard error values. It can be seen that even in the XJTU-SY dataset, the target diagnosis performance of DASC with semantic clustering loss is the highest.

Next, we describe the results of UDA operations (Scenario II) between domains obtained from different rotating machines. First, we considered the UDA operation between CWRU and IMS dataset. Each data set has 4 states, so C is 4. In order to confirm the generalization performance and robustness of DASC, verification of UDA operation between CWRU data and IMS data obtained for four operating states was performed. Here, a total of 8 UDA tasks (CWRU1→IMS, CWRU2→IMS, CWRU3→IMS, CWRU4→IMS, IMS→CWRU1, IMS→CWRU2, IMS→CWRU3, IMS→CWRU4) were performed. The average diagnostic accuracy of CWRU→IMS and IMS→CWRU is shown in FIG. 9 . As a result, it can be seen that DASC shows better diagnostic accuracy than the conventional method in all cases.

Next, we performed UDA tasks between two different data sets obtained through run-to-failure experiments, which are IMS and XJTU-SY. At this time, XJTU-SY3 with 4 types of failure (same as IMS data) was used to match the number of soundness between the two datasets. The target diagnosis result between the IMS and the XJTU-SY data set is shown in FIG. 10 . Even in this case, the DASC of the diagnosis method S100 to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied shows the best performance among other UDA methods.

In other words, it can be seen that the UDA operation can improve the diagnostic performance of rotating machines when the domains are different. In addition, the DASC method of the diagnostic method (S100) to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied adopts an additional loss term called semantic clustering loss and reduces it at various characteristic levels, so it is different UDA in all cases. It has been shown that the performance is superior to the method. This is the result of learning semantically well-clustered discriminant features with sufficiently large inter-class distances and small intra-class distances in DASC.

The efficacy of DASC is confirmed through visualization of the learned feature distribution as follows. 11 and 12 are diagrams showing the feature distribution learned in the diagnosis method (S100) to which the domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention is visualized using the t-dispersion t-SNE (Stochastic Neighboring Embedding) method. show that

11 shows a visualization result for the operation of Scenario I for CWRU. 11(a) is a CNN method, FIG. 11(b) is a DANN method, FIG. 11(c) is an MMD method, FIG. 11(d) is an Adv+Disc method, FIG. 11(e) is a deep CORAL method, FIG. (f) shows the visualization result by the DASC method. 11, it shows that the feature distribution obtained from DASC best separates the data by class. This is because the DASC method learns the most discriminant features by applying the semantic clustering loss. As a result, it can be seen that the overlap between different classes is minimal, and the same classes are well clustered.

12 shows a visualization result for the operation of Scenario II for CWRU→IMS. 12(a) is a CNN method, FIG. 12(b) is a DANN method, FIG. 12(c) is an MMD method, FIG. 12(d) is an Adv+Disc method, FIG. 12(e) is a deep CORAL method, FIG. (f) shows the visualization result by the DASC method. In the case of a CNN using only source data as shown in FIG. 12( a ), the domain gap is large because the relationship with the target data is not considered. On the other hand, when the UDA method is used, the mismatch between the source domain and the target domain is reduced in order to learn domain invariant properties that can be shared by both domains as shown in FIGS. 12(b) to 12(f). In addition, in the case of DASC, both discriminant and domain invariant representations can be obtained by learning semantically well clustered features according to states. Therefore, the learned features have the characteristics of significant inter-class variability and small intra-class variability, and as a result, the best generalization performance in the target region can be obtained based on these well-learned powerful features.

Next, through an ablation study, the effect of the DASC method of the diagnosis method (S100) to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied was verified by comparing it with other methods. The first is CNN using only classification loss, the second is UDA using domain adaptation loss, third is Last_SC using semantic clustering loss only in the last layer of the feature extractor, and fourth is an embodiment of the present invention This is the DASC method of the diagnosis method (S100) to which the domain adaptation and semantic clustering algorithm according to Diagnosis accuracy results for all tasks of scenario I and scenario II are shown in FIG. 13 .

As shown in FIG. 13 , when domain adaptation loss is additionally considered, it showed higher diagnostic accuracy than CNN using only classification loss. In addition, when the semantic clustering loss is considered, the classification accuracy is further improved. Furthermore, when the semantic clustering loss was applied at the multi-feature level, the highest diagnostic accuracy was shown.

Finally, in order to evaluate how well the target feature, which is a property to be achieved by the DASC method of the diagnosis method ( S100 ) to which the domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention, is semantically clustered, SCI (Semantic An index called Clustering Index) was defined. Here, the SCI is defined as the ratio between the average inter-class distance to the target data and the average intra-class distance, and may be expressed as follows.

은 상태

의 샘플 수를 나타내고,

는 피쳐익스트랙터단계(S110)에 의해 도출된 샘플 x_i의 피쳐벡터를 나타내며,

은 추출된 피쳐 공간에서 레이블

인 샘플의 중심을 나타낸다.where C denotes the number of states of the target rotating machine,

silver state

represents the number of samples in

represents the feature vector of the sample x _i derived by the feature extractor step (S110),

is the label in the extracted feature space

represents the center of the sample.

A larger SCI value means that the target sample is semantically better clustered according to the corresponding class. That is, the target sample feature has a small intra-class variation and a larger inter-class distance. Therefore, the UDA method with a large SCI value can have high target generalization performance, especially for data located near the boundary of each class.

The average SCI values of all UDA methods compared for the UDA tasks of scenario I and scenario II are shown in FIG. 14 . As shown in FIG. 14 , it shows that DASC has the largest SCI value among comparative UDAs. That is, it means that the features learned through DASC have the most discriminant distribution. This is because the DASC method learns semantically well clustered features by applying additional semantic clustering loss at multiple feature levels. This result well explains why the DASC of the diagnosis method S100 to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied shows the highest generalization performance in the target domain. That is, it can be confirmed that by considering the semantic clustering loss based only on the labeled source domain data, it is possible to improve the class-specific clustering performance of different but related target domain data. Accordingly, the DASC of the diagnosis method S100 to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied improves target domain generalization performance.

In addition, as shown in FIGS. 7 to 10 , the DASC method of the diagnosis method S100 to which the domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied shows the smallest performance deviation, so it is the most effective for the target domain data. It can be seen that stable and powerful diagnostic performance can be provided. That is, according to the diagnosis method (S100) to which the domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention, it is possible to provide the most accurate and powerful UDA-based diagnostic model for a rotating machine in an environment where it is difficult to obtain label information. It works.

As described above, the feature extractor step (S110), the label classifier step (S120), the domain discrepancy step (S130), the semantic clustering step (S140), the calculation step (S150), the output step (S160). According to the diagnostic method (S100) to which the domain adaptation and semantic clustering algorithm is applied according to an embodiment of the present invention, the state of the target system without label information can be efficiently diagnosed based on the diagnostic data obtained from the source system. .

In the above, even though all the components constituting the embodiment of the present invention are described as being combined or operated in combination, the present invention is not necessarily limited to this embodiment. That is, within the scope of the object of the present invention, all the components may operate by selectively combining one or more.

In addition, terms such as "comprises", "comprises" or "have" described above mean that the corresponding component may be inherent, unless otherwise specified, excluding other components. Rather, it should be construed as being able to further include other components. All terms, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs, unless otherwise defined. Terms commonly used, such as those defined in the dictionary, should be interpreted as being consistent with the contextual meaning of the related art, and are not interpreted in an ideal or excessively formal meaning unless explicitly defined in the present invention.

And, the above description is merely illustrative of the technical idea of the present invention, and various modifications and variations will be possible without departing from the essential characteristics of the present invention by those skilled in the art to which the present invention pertains.

Accordingly, the embodiments disclosed in the present invention are not intended to limit the technical spirit of the present invention, but to explain, and the scope of the technical spirit of the present invention is not limited by these embodiments. The protection scope of the present invention should be construed by the following claims, and all technical ideas within the scope equivalent thereto should be construed as being included in the scope of the present invention.

100: Diagnostic device to which domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied
110: feature extractor part
120: label classifier unit
130: domain disk refresh unit
140: semantic clustering unit
150: arithmetic unit
160: output unit
S100: Diagnosis method to which domain adaptation and semantic clustering algorithm according to an embodiment of the present invention is applied
S110: feature extractor stage
S120: Label classifier stage
S130: domain disk refresh step
S140: Semantic clustering step
S150: Calculation step
S160: output stage

Claims (20)

In the diagnosis apparatus to which a domain adaptation and semantic clustering algorithm performed by a state diagnosis system of a mechanical system without label information is applied,
A first feature and a first data set from a first dataset including time-series signal data measured and labeled in the first system and a second dataset including time-series signal data that is measured and unlabeled in the second system a feature extractor unit for learning a machine learning-based feature calculation model for calculating two features;
a label classifier unit for learning a machine learning-based state classification model for classifying the state of the first system using the first dataset, and calculating a classifier loss ( _LC );
Learning a machine learning-based domain adaptation model that performs domain adaptation by setting the first dataset as a source domain and setting the second dataset as a target domain as a target domain, a domain dispensing unit for calculating domain-related loss ( _LD ); and
A semantic clustering unit for learning a machine learning-based clustering model for clustering by state based on the label of the first dataset, and calculating a semantic clustering loss (L _SC ),
The classification loss is
It is calculated based on Equation 1,
The domain adaptation loss is
It is calculated based on Equation 2,
The semantic clustering loss is
A diagnosis apparatus to which domain adaptation and semantic clustering algorithms are applied, characterized in that the calculation is based on Equation 3.
{Formula 1}

(remind

is the classification loss, C is the number of states of the first system, Y is the label vector, and

is the value of the k-th node of Y, and p is a probability vector corresponding to the k-th node of Y)
{Formula 2}

(remind

is the domain adaptation loss, H is a universal reproducing kernel Hilbert space (RKHS), n _S and n _T are the numbers of the first and second datasets, respectively, and

and above

is the first feature and the second feature, respectively)
{Formula 3}

(remind

is the semantic clustering loss,

is the number of characteristic layers of the semantic clustering part, the

is the kth balancing parameter (balancing parameter), the

is the number of the first dataset, the

is at the kth feature level

th feature vector of the first dataset, the

Is

the second dataset and

A matrix that is 1 if the second dataset belongs to the same class and 0 if it is a different class.

value of , where d is a value controlling the minimum distance between two datasets in different states)
delete The method according to claim 1,
The feature extractor unit, the domain discrepancy unit, and the semantic clustering unit,
The first dataset and the second dataset are trained in mini-batch units, and when learning is performed on all the small groups of the first dataset and the second dataset, the first dataset and the second dataset A diagnostic device to which a domain adaptation and semantic clustering algorithm is applied, characterized in that small groups are re-extracted and re-learned at random from the second dataset.
4. The method according to claim 3,
The feature extractor unit,
A diagnosis apparatus to which a domain adaptation and semantic clustering algorithm is applied, characterized in that the first feature and the second feature are calculated based on Equation (4).
{Formula 4}

(remind

is the first data set or the second data set, the

is said

The first feature or the second feature calculated from

is an activation function, the

is a weight, the bias is a constant)
5. The method according to claim 4,
Based on Equation 5, the total loss (

) further comprising an operator to calculate,
The calculation unit,
Modify the weight based on Equation 6 and gradient descent,
The feature extractor unit, the domain discrepancy unit, and the semantic clustering unit,
It is characterized in that it is re-learned based on the modified weight,
The calculation unit,
When the change in the total loss is smaller than a preset ratio even though the weight is corrected a preset number of times, the learning of the feature extractor unit, the domain discrepancy unit, and the semantic clustering unit is stopped to derive a final diagnostic algorithm A diagnostic device to which domain adaptation and semantic clustering algorithms are applied.
{Formula 5}

(remind

is the total loss, above

is the classification loss,

is the domain adaptation loss, the

is the semantic clustering loss,

Wow

is the coefficients set by the user)
{Formula 6}

(remind

is the modified said weight, said

is the weight before fertilization,

is a constant set by the user,

is the derivative of the total loss with respect to the weight)
6. The method of claim 5,
The derivative of the total loss with respect to the weight is,
A diagnostic device to which a domain adaptation and semantic clustering algorithm is applied, characterized in that it is calculated according to a back-propagation algorithm and a chain rule.
5. The method according to claim 4,
The label classifier unit,
It is characterized in that the weight is corrected based on Equation 7 and gradient descent, and re-learning is performed based on the modified weight,
The label classifier unit,
Even though the weight is modified a preset number of times, if the change in the classification loss is smaller than a preset ratio, learning is stopped and a final machine learning-based state classification model is derived. A domain adaptation and semantic clustering algorithm is applied diagnostic device.
{Formula 7}

(remind

is the modified said weight, said

is the weight before fertilization,

is a constant set by the user,

is the differential value of the classification loss with respect to the weight)
8. The method of claim 7,
The differential value of the classification loss with respect to the weight is,
A diagnostic device to which a domain adaptation and semantic clustering algorithm is applied, characterized in that it is calculated according to a back-propagation algorithm and a chain rule.
7. The method of claim 6,
and an output unit for outputting the state of the second system diagnosed by using the final diagnosis algorithm.
7. The method of claim 6,
The feature extractor unit,
It uses a convolution neural network (CNN),
The label classifier unit,
A diagnostic device to which a domain adaptation and semantic clustering algorithm is applied, characterized in that it is provided as a fully connected layer.
In the diagnosis method to which a domain adaptation and semantic clustering algorithm performed by a state diagnosis system of a mechanical system without label information is applied,
From a first dataset including time-series signal data that is measured and labeled in the first system and a second dataset that includes time-series signal data that is not measured and labeled in the second system, the feature extractor unit includes the first feature ( feature) and a feature extractor step of learning a machine learning-based feature calculation model for calculating a second feature;
A label classifier step in which the label classifier unit learns a machine learning-based state classification model that classifies the state of the first system using the first dataset, and calculates a classifier loss ( _LC ) ;
Machine learning-based domain adaptation in which a domain dispensing unit performs domain adaptation by setting the first dataset as a source domain and setting the second dataset as a target domain Learning the model, domain-related loss (domain-related loss, _LD ) domain dispensing step of calculating; and
The semantic clustering unit learns a machine learning-based clustering model that clusters by state based on the label of the first dataset, and includes a semantic clustering step of calculating semantic clustering loss (L _SC ),
The classification loss is
It is calculated based on Equation 1,
The domain adaptation loss is
It is calculated based on Equation 2,
The semantic clustering loss is
A diagnosis method to which a domain adaptation and semantic clustering algorithm is applied, characterized in that it is calculated based on Equation 3.
{Formula 1}

(remind

is the classification loss, C is the number of states of the first system, Y is the label vector, and

is the value of the k-th node of Y, and p is a probability vector corresponding to the k-th node of Y)
{Formula 2}

(wherein H is a universal reproducing kernel Hilbert space (RKHS), and n _S and n _T are the number of the first dataset and the second dataset, respectively, and

and above

is the first feature and the second feature, respectively)
{Formula 3}

(remind

is the semantic clustering loss,

is the number of characteristic layers in the semantic clustering step,

is the kth balancing parameter (balancing parameter), the

is the number of the first dataset, the

is at the kth feature level

th feature vector of the first dataset, the

Is

the second dataset and

A matrix that is 1 if the second dataset belongs to the same class and 0 if it is a different class.

value of , where d is a value controlling the minimum distance between two datasets in different states)
delete 12. The method of claim 11,
The feature extractor step, the domain discrepancy step, and the semantic clustering step are
The first dataset and the second dataset are trained in mini-batch units, and when learning is performed on all the small groups of the first dataset and the second dataset, the first dataset and the second dataset A diagnostic method to which a domain adaptation and semantic clustering algorithm is applied, characterized in that small groups are re-extracted at random from the second dataset and re-learned.
14. The method of claim 13,
The feature extractor step is
A diagnosis method to which a domain adaptation and semantic clustering algorithm is applied, characterized in that the first feature and the second feature are calculated based on Equation (4).
{Formula 4}

(remind

is the first data set or the second data set, the

is said

The first feature or the second feature calculated from

is an activation function, the

is a weight, the bias is a constant)
15. The method of claim 14,
Based on Equation 5, the total loss (

) further comprising an operation step of calculating
The calculation step is
Modify the weight based on Equation 6 and gradient descent,
The feature extractor step, the domain discrepancy step, and the semantic clustering step are
It is characterized in that it is re-learned based on the modified weight,
The calculation step is
When the change in the total loss is smaller than a preset ratio even though the weight is corrected a preset number of times, the learning of the feature extractor step, the domain discrepancy step, and the semantic clustering step is stopped to derive a final diagnostic algorithm A diagnostic method to which domain adaptation and semantic clustering algorithms are applied.
{Formula 5}

(remind

is the total loss, above

is the classification loss,

is the domain adaptation loss, the

is the semantic clustering loss,

Wow

is the coefficients set by the user)
{Formula 6}

(remind

is the modified said weight, said

is the weight before fertilization,

is a constant set by the user,

is the derivative of the total loss with respect to the weight)
16. The method of claim 15,
The derivative of the total loss with respect to the weight is,
A diagnostic method to which a domain adaptation and semantic clustering algorithm is applied, characterized in that it is calculated according to a back-propagation algorithm and a chain rule.
15. The method of claim 14,
The label classifier step is,
It is characterized in that the weight is corrected based on Equation 7 and gradient descent, and re-learning is performed based on the modified weight,
The label classifier step is,
Even though the weight is modified a preset number of times, if the change in the classification loss is smaller than a preset ratio, learning is stopped and a final machine learning-based state classification model is derived. A domain adaptation and semantic clustering algorithm is applied diagnostic method.
{Formula 7}

(remind

is the modified said weight, said

is the weight before fertilization,

is a constant set by the user,

is the differential value of the classification loss with respect to the weight)
18. The method of claim 17,
The differential value of the classification loss with respect to the weight is,
A diagnostic method to which a domain adaptation and semantic clustering algorithm is applied, characterized in that it is calculated according to a back-propagation algorithm and a chain rule.
17. The method of claim 16,
and an output step of outputting the state of the second system diagnosed using the final diagnosis algorithm by an output unit.
17. The method of claim 16,
The feature extractor step is
It uses a convolution neural network (CNN),
The label classifier step is,
A diagnosis method to which a domain adaptation and semantic clustering algorithm is applied, characterized in that it is provided as a fully connected layer.

Images