KR20230083570A - Health monitoring method of power driving system using sound signal based on deep learning - Google Patents

Abstract

A soundness monitoring method of a power driving system using a noise signal based on deep learning is disclosed. A soundness detection method of a power driving system performed by a soundness monitoring system according to an embodiment includes generating an image using a noise signal measured in the power driving system; receiving the generated image as an input to a learning model for diagnosing a defect; and classifying a defect pattern of the power driving system from the generated image using the learning model.

Description

Health monitoring method of power driving system using noise signal based on deep learning {HEALTH MONITORING METHOD OF POWER DRIVING SYSTEM USING SOUND SIGNAL BASED ON DEEP LEARNING}

The description below relates to techniques for monitoring the health of power driving systems.

The power driving system (PDS) has been applied as an important mechanical system in various industrial fields such as machine tools and automobile engine timing systems that require large amounts of power. Various techniques have been studied to prevent economic loss and human damage by early detection of defects occurring in the power driving system.

Defects in power driving systems are caused by defects in bearings that support shafts and gears that transmit power, or by breakage of chains. In order to detect defects at an early stage, acoustic vibration signals are generally measured, and defect characteristics are extracted from the measured signals to determine the presence or absence of defects. Various methods have been used to extract defect features from measured acoustic vibration signals. For example, feature extraction methods such as time averaging method, cepstrum analysis, pseudo-Wigner-Ville distribution, wavelet transform, higher order method, adaptive linear enhancement method, empirical mode decomposition, and periodic analysis have been used. A nearest neighbor algorithm, a Bayesian classifier, a support vector machine, an artificial neural network, and the like are used to classify the types of defects by utilizing the characteristics of these defects.

A technology for more accurately diagnosing and classifying a defect in a power driving system is required.

It is possible to provide a fault diagnosis and pattern classification method and system of a power driving system using continuous wavelet transform (CWT) and deep learning technology.

A soundness detection method of a power driving system performed by a soundness monitoring system includes generating an image using a noise signal measured in the power driving system; receiving the generated image as an input to a learning model for diagnosing a defect; and classifying a defect pattern of the power driving system from the generated image using the learning model for defect diagnosis, wherein the learning model for defect diagnosis includes a plurality of defect conditions and normal conditions from the image. It may have been learned to classify a defect pattern that does.

The generating may include generating image data in a time-frequency domain by applying a continuous wavelet (CWT) to the measured noise signal.

The learning model includes an input layer into which the generated image is input, a convolutional layer having 16 feature maps with a filter size of 3 × 3, a maximum pooling layer with a size of 2 × 2, and a filter size of 3 × 3. It can be composed of a convolutional layer with feature maps, a convolutional layer with a maximum pooling layer of size 2×2, and a final convolutional layer to which a softmax function is applied.

The learning model may include extracting defect information from an image; and extracting a feature map including the extracted defect information.

The power driving system includes parts including a motor, chain, pinion, sprocket, gear, bearing, or rotating shaft, and a combination of the parts, and classifying the defect pattern includes learning for diagnosing the defect. The method may include extracting defect feature information related to a defect condition from the generated image using a model, and classifying a defect pattern of the power driving system based on the extracted defect feature information.

Classifying the defect pattern may include diagnosing a defective part of the power driving system from the generated image, and classifying a single defect pattern or multiple defect patterns for the diagnosed defective component.

The defect patterns including the plurality of defect conditions and normal conditions include gear cracks in drive gears, gear breakage in drive gears, eccentricity of drive shafts in motors, outer race defects due to holes in bearings, inner race defects due to foreign substances in bearings, and motors. eccentricity of the drive shaft + gear breakage of the drive gear, eccentricity of the drive shaft of the motor + gear breakage of the drive gear + outer ring defect of the bearing, eccentricity of the drive shaft of the motor + gear breakage of the drive gear + inner ring defect of the bearing, or normal condition. can

The soundness monitoring system includes an image generating unit generating an image using a noise signal measured in a power driving system; an image input unit that receives the generated image into a learning model for diagnosing defects; and a defect pattern classification unit classifying a defect pattern of the power driving system from the generated image using the learning model for defect diagnosis, wherein the learning model for defect diagnosis includes a plurality of defect conditions and normal conditions from the image. It may be learned to classify a defect pattern including .

Continuous wavelet transform (CWT) and deep learning techniques can be used to improve the accuracy of fault diagnosis and pattern classification of power driving systems.

Complete damage to the power driving system due to defective parts can be prevented.

1 is a diagram for explaining an operation of classifying a defect pattern of a power driving system in a health monitoring system according to an embodiment.
2 is a block diagram illustrating a configuration of a health monitoring system according to an exemplary embodiment.
3 is a flowchart illustrating a method for monitoring health of a power driving system in a health monitoring system according to an exemplary embodiment.
4 is a diagram for explaining an operation of measuring noise data of a power driving system according to an exemplary embodiment.
5 is a diagram for explaining the configuration of a power driving system according to one embodiment.
6 is an example for explaining a defect pattern of a power driving system according to one embodiment.
7 is a graph showing time-domain noise data according to an embodiment.
8 is a graph showing the power spectral density of noise according to an embodiment.
9 is a graph illustrating a continuous wavelet transform operation for noise data, according to an exemplary embodiment.
10 is an example for explaining the structure of a learning model for diagnosing a defect according to an embodiment.
11 and 12 are graphs illustrating the accuracy of a learning model for diagnosing defects, according to an embodiment.
13 is an example of a feature map of a learning model for diagnosing a defect according to an embodiment.
14 is an example illustrating data visualization using a t-probabilistic embedding method of a learning model for diagnosing defects, according to an embodiment.

Hereinafter, an embodiment will be described in detail with reference to the accompanying drawings.

1 is a diagram for explaining an operation of classifying a defect pattern of a power driving system in a health monitoring system according to an embodiment.

The soundness monitoring system 100 may diagnose a component defect early by using the learning model 101 for defect diagnosis in order to prevent complete damage due to a component defect of the power driving system. At this time, for example, the power driving system may be composed of parts such as chains, sprockets, gears, bearings, rotating shafts, and/or combinations of parts.

As an example, the learning model 101 for fault diagnosis may be configured based on a convolutional neural network (CNN) for fault diagnosis of a power driving system, and fault diagnosis and fault patterns are learned using a data set for fault diagnosis. It may have been built to be.

An image generated by using a signal measured in a power driving system may be input to the learning model 101 for fault diagnosis as input data 110 . In the embodiment, image data in the time-frequency domain may be input as input data 110 to the learning model 101 for diagnosing defects.

Output data 120 may be output through the learning model 101 for diagnosing a defect input to the input data 110 . A defect pattern of the power driving system may be classified from an image generated using the learning model 101 for defect diagnosis. In other words, defect patterns can be classified as output data 120 .

In detail, an image may be generated using a noise signal measured in a power driving system. Image data in the time-frequency domain may be generated by applying continuous wavelet transform (CWT) to the measured noise signal. The continuous wavelet transform analysis method can be defined as follows.

Equation 1:

는 모 웨이블렛으로 시간과 주파수 영역에서 국부화된 고정함수이다. 함수

는 시간 도메인에서 이동(b-tanslation)과 주파수 도메인에서 모웨이블렛으로 스케일링(a-dilation)을 적용한다. 실시예에서 사용된 모 웨이블렛은 모아렛(Morlet) 웨이블렛이며 다음과 같이 정의될 수 있다. here,

is a parent wavelet, which is a localized fixed function in the time and frequency domains. function

applies translation (b-translation) in the time domain and scaling (a-dilation) to the mo wavelet in the frequency domain. The parent wavelet used in the embodiment is a Morlet wavelet and can be defined as follows.

Equation 2:

는 모 웨이블렛이 주파수 도메인으로 변환될 때 모 웨이블렛의 중심 주파수이다. 실시예에서는 측정된 소음 신호에 연속 웨이블렛 변환을 적용하여 시간-주파수 영역의 이미지 데이터가 생성되고, 상기 생성된 시간-주파수 영역의 이미지 데이터가 결함 진단을 위한 학습 모델(101)의 입력 데이터로 사용될 수 있다.

is the center frequency of the parent wavelet when the parent wavelet is transformed into the frequency domain. In the embodiment, continuous wavelet transform is applied to the measured noise signal to generate time-frequency domain image data, and the generated time-frequency domain image data is used as input data of the learning model 101 for defect diagnosis. can

은 다음과 같이 수학식3을 통하여 획득할 수 있다.The learning model 101 for diagnosing such a defect may be configured based on a convolutional neural network (CNN). In the convolutional neural network, defect information about the generated image may be extracted through a plurality of convolutional layers and a pooling layer of an image generated through continuous wavelet transform on a measured noise signal. In the convolutional neural network, the neural network is composed of a plurality of convolutional layers and pooling layers, and a feature map having defect information can be extracted from each convolutional layer. Feature maps in the first convolutional layer

can be obtained through Equation 3 as follows.

Equation 3:

는 비선형 활성화 함수이고

은 l번째 층의 스칼라(scalar) 바이어스(bias) 이다.

는 (l-1)번째 층에서 선택된 특징 지도이다.

는 이전 층의

활성화 함수를 합성곱하는 합성곱 연산자를 나타낸다.

는 2차원 필터 (filter)이며 필터의 가중치는 특정한 특징을 감지하도록 학습될 수 있다. 따라서 새로운 입력 이미지의 정확한 분류를 위해 서로 다른 카테고리를 구별할 수 있는 연속적인 단계에서 효과적인 특징 선택이 필요하다. 이러한 이유로 풀링층이 사용되며 l번째 풀링층

은 다음과 같이 나타낼 수 있다.here,

is a nonlinear activation function and

is the scalar bias of the lth layer.

is a feature map selected in the ( l -1)th layer.

of the previous layer

Represents a convolution operator that convolutions an activation function.

is a two-dimensional filter, and the weights of the filter can be learned to detect specific features. Therefore, for accurate classification of new input images, effective feature selection is required in successive steps capable of distinguishing different categories. For this reason, a pooling layer is used and the l- th pooling layer is used.

can be expressed as:

Equation 4:

는

에 의해 축소된 평균 또는 최대값 함수와 같은 축소 함수이다. 그리고

는 축소되기 위한 합성곱된 특징 지도이다. 합성곱 신경망의 입력값이 연속적으로 합성곱 및 풀링 과정을 지나면서 네트워크는 모든 이미지를 효율적으로 특징을 추출하는 학습을 진행할 수 있다. 마지막 신경망 층의 출력값

는 수학식5 와 같이 주어진다.

Is

A reduction function, such as a mean or maximum function reduced by and

is the convolutional feature map to be reduced. As the input values of the convolutional neural network continuously pass through convolution and pooling processes, the network can proceed with learning to efficiently extract features from all images. Output of the last neural network layer

is given as in Equation 5.

Equation 5:

는 출력 층의 바이어스이고, W는 완전히 연결된 층의 입력층과 출력층 사이의 가중치 행렬이며,

는 완전히 연결된 층의 특징을 나타내는 특징 벡터 (feature vector)를 나타내고,

는 결함을 분류하는 소프트맥스(softmax) 함수이다. 학습 매개변수,

,

,

및

는 학습 과정에서 최적화된 값이 결정될 수 있다. 실제 출력값과 원하는 목표값 간의 오차를 최소화하기 위해 확률적 경사하강법(SGD; stochastic gradient descent)을 통해 학습 매개변수의 최적화가 이루어질 수 있다. 여기서 경사값은 역전파(back propagation) 방법을 통하여 계산될 수 있다.

is the bias of the output layer, W is the weight matrix between the input and output layers of the fully connected layer,

Represents a feature vector representing the features of a fully connected layer,

is the softmax function for classifying defects. learning parameters,

,

,

and

An optimized value may be determined in the learning process. Optimization of learning parameters may be performed through stochastic gradient descent (SGD) in order to minimize an error between an actual output value and a desired target value. Here, the gradient value may be calculated through a back propagation method.

2 is a block diagram illustrating a configuration of a health monitoring system according to an exemplary embodiment, and FIG. 3 is a flowchart illustrating a method for monitoring health of a power driving system in the health monitoring system according to an exemplary embodiment.

The processor of the soundness monitoring system 100 may include an image generator 210 , an image input unit 220 and a defect pattern classification unit 230 . These components of the processor may represent different functions performed by the processor according to control instructions provided by program code stored in the health monitoring system. The processor and components of the processor may control the health monitoring system to perform steps 310 to 330 included in the health monitoring method of FIG. 3 . In this case, the processor and components of the processor may be implemented to execute instructions according to the code of an operating system included in the memory and the code of at least one program.

The processor may load the program code stored in the file of the program for the health monitoring method into the memory. For example, when a program is executed in the health monitoring system, the processor may control the health monitoring system to load a program code from a file of the program into a memory under the control of an operating system. At this time, each of the image generator 210, image input unit 220, and defect pattern classification unit 230 executes a command of a corresponding part of the program code loaded into the memory to execute subsequent steps 310 to 330. There may be different functional representations of the processor for

In step 310, the image generator 210 may generate an image using the noise signal measured in the power driving system. The image generator 210 may generate image data in the time-frequency domain by applying a continuous wavelet (CWT) to the measured noise signal.

In step 320, the image input unit 220 may receive the generated image as an input to a learning model for diagnosing defects. In this case, the learning model may be one learned to classify a defect pattern including a plurality of defect conditions and a normal condition from the image. In detail, the learning model consists of an input layer with generated images, a convolutional layer with 16 feature maps with a filter size of 3 × 3, a maximum pooling layer with a size of 2 × 2, and a filter size of 3 × 3. It can be composed of a convolution layer with 32 feature maps, a convolution layer with a maximum pooling layer of size 2 × 2, and a final convolution layer with a softmax function applied. The learning model configured as described above may be learned through a process of extracting defect information from image data and extracting a feature map including the extracted defect information.

In step 330, the defect pattern classification unit 230 may classify the defect pattern of the power driving system from the generated image using the learning model for defect diagnosis. The defect pattern classification unit 230 extracts defect feature information related to the defect condition from the generated image using the learning model for defect diagnosis, and classifies the defect pattern of the power driving system based on the extracted defect feature information. can The defect pattern classification unit 230 may diagnose defective parts of the power driving system from the generated image and classify a single defect pattern or multiple defect patterns for the diagnosed defective parts. For example, the defect pattern classification unit 230 may diagnose a single defect when only one defect is detected in the power driving system, and diagnose multiple defects when several defects are detected in the power driving system. there is.

4 is a diagram for explaining an operation of measuring noise data of a power driving system according to an exemplary embodiment.

4 is an example of a configuration diagram and a control device of a power driving system test apparatus used in an experiment. The power driving system may be composed of mechanical parts such as motors, gearboxes, bearings, chain systems, rotating shafts, sprockets, and pinions, or/and a combination of mechanical parts, in order to drive power transmission. In the embodiment, artificially manufacturing and combining mechanical parts by making defects will be described as an example. A microphone for measuring noise may be installed in order to measure abnormal noise generated from defects in mechanical parts during operation of the power driving system and to classify patterns of the defects. For example, noise data may be measured by placing a microphone at a distance of 1M from the power driving system. Noise data can be measured using a 1/2-inch free-field microphone (B & K 4192, Denmark) and transmitted to the health monitoring system via the Data Acquisition System (NI 9233, USA). At this time, the soundness monitoring system may be implemented as a computer and capable of signal analysis. The measured noise data may be converted into digital data through measurement equipment and transmitted to the health monitoring system. The health monitoring system diagnoses defective parts and classifies patterns of defects.

5 is a diagram for explaining the configuration of a power driving system according to one embodiment.

5 shows the main components for the drive system of the power driving system, and shows the detailed positions of bearings and gears applied to the power driving system. For example, an experimental power driving system may consist of 4 bearings, 4 gears, 1 chain, 2 sprockets and 1 motor. The rotation speed of the driving shaft for input is 1800 rpm (30 Hz), the rotation speed of the drive shaft for output is 59.4 rpm (0.99 Hz), and the reduction ratio is 30. In the embodiment, artificial defects are applied to motor shafts, bearings, and gears, and complete damage may occur when defects occur in sprockets and chains, so that artificial defects are not applied will be described as an example.

6 is an example for explaining a defect pattern of a power driving system according to one embodiment.

6 is an example showing three gear cracks in drive gear 2, gear breakage in drive gear 2, and outer ring defects due to a hole in bearing 1 in the power driving system.

The defect conditions (types) are: gear crack of drive gear 2 (fault 1), gear breakage of drive gear 2 (fault 2), eccentricity of drive shaft in motor (fault 3), outer ring defect due to hole in bearing ① (fault 4) , inner ring defect due to foreign matter in bearing ① (fault 5), motor drive shaft eccentricity + gear breakage of drive gear 2 (fault 6), motor drive shaft eccentricity + gear breakage of drive gear 2 + outer ring defect of bearing ① (defect 7) ), the eccentricity of the driving shaft of the motor + the gear breakage of the driving gear 2 + the defect of the inner ring of the bearing ① (defect 8), and the normal state (normal).

In other words, the defect patterns can be tabulated as follows.

Table 1:

A microphone may be disposed at a location separated by a specific distance from the power driving system in order to measure noise data of the power driving system. Noise data of the power driving system may be measured through the disposed microphone, and normal and defective data may be obtained based on the measured noise data. Experiments can be performed for each fault condition. For example, a total of 700 experiments can be performed, each test run time is 5 minutes, and noise data of the last 30 seconds obtained using a microphone can be used for fault diagnosis.

7 is a graph showing time-domain noise data according to an embodiment, and FIG. 8 is a graph showing power spectral density of noise according to an embodiment.

Noise data measured in the time domain can be subjected to frequency and continuous wavelet transform analysis using MATLAB (MathWorks, USA) signal processing toolbox. 7 shows time domain data of 8 defective and 1 normal signals measured with a microphone.

8 shows a comparison of power spectra of a normal noise signal and eight malfunction noise signals. Depending on each defect, a difference occurs in the form of time domain noise data and spectrum. In the spectrum, the blue curve represents the faulty state and the orange curve represents the steady state.

Specifically, the power spectra of the 8 faulty signals are different from those of the normal signals. According to the vibration theory of rotating machinery, in the case of a rotating machine such as a power driving system, there are several excitation forces that cause vibration and noise of the machine. and impact excitation force due to the impact of the sprocket and chain. The excitation force is noise vibration that is basically present because it exists even under normal conditions of the power driving system. Eight defects of the power driving system can change the excitation force, and the changed excitation force causes a change in the frequency spectrum of the noise signal. Accordingly, the power spectrum of the 8 faulty signals differs from that of the normal signal.

9 is a graph illustrating a continuous wavelet transform operation for noise data according to an exemplary embodiment.

In an embodiment, an image may be generated by applying continuous wavelet transform to a noise signal measured by a microphone. As shown in FIG. 9 , the image analyzed through continuous wavelet transform provides information on the driving state of the power driving system. 9 shows continuous wavelet transform results for noise signals measured under normal operating conditions of a power driving system. According to this result, it can be seen that noise signals related to excitation forces of various parts are generated under normal conditions. When a defect occurs, the continuous wavelet transform of the noise signal changes. As such, as a result of the continuous wavelet transformation, it can be confirmed that the frequency of the sprocket, the meshing frequency of gear 2, and the defect frequency of bearing 1 appear.

In the embodiment, continuous wavelet transform is applied to noise data measured every 30 seconds in 8 defect conditions to obtain an image, and then it can be used as input data for a CNN-based learning model. For example, 100 consecutive wavelet transform images for each defect may be obtained, and a total of 900 image data including normal conditions may be used as input image data for a CNN-based learning model. In this case, the data size of each image may be 244 Х 244 Х 3.

10 is an example for explaining the structure of a learning model for diagnosing a defect according to an embodiment.

For the continuous wavelet transform algorithm, CNN toolbox provided by MATLAB can be used. In order to successfully apply the learning model for fault diagnosis to fault pattern classification and fault feature extraction, the architecture of the network is important. There is no special skill required to configure the architecture of the network, and it is to find the optimal architecture configuration through various configurations.

Referring to FIG. 10 , an input layer into which an image subjected to continuous wavelet transform is input, and a first layer located after the input layer is a convolutional layer having 16 feature maps with a filter size of 3Х3. Then there is an average pooling layer of size 2Х2. The next layer is a convolution layer with 32 feature maps with a filter size of 3Х3 and an average pooling layer of 2Х2. The output layer has 9 neurons corresponding to 8 defects and 1 normal condition. A softmax function may be applied as a classification function to the last fully connected layer. Optimized by the gradient descent learning method, the learning rate of the learning process can be used to train the network with an initial learning rate of 0.001. The batch size is taken as 128, and learning can be performed by repeating 1300 times (30 epochs). The change in accuracy during learning is shown in FIG. 11 .

11 and 12 are graphs showing the accuracy of a learning model for diagnosing defects, according to an embodiment.

For example, an optimal weight value may be obtained after learning 400 times with a minimum error. For training, 700 samples out of 900 data are used for training, 100 samples are used for verification, and 100 samples can be used for testing.

The accuracy of defect classification is the ratio of the number of correctly classified test samples to the total number of test samples and is expressed in Equation 6.

Equation 6:

Through experiments, it can be confirmed that 90 samples out of 100 test samples were classified with high accuracy as shown in FIG. 12 .

13 is an example of a feature map of a learning model for diagnosing a defect according to an embodiment. As a learning result of the learning model for fault diagnosis, feature maps of 8 faults and 1 steady state can be extracted from the final convolutional layer. According to these results, it can be confirmed that different images are displayed for each defect.

14 is an example illustrating data visualization using a t-probabilistic embedding method of a learning model for diagnosing defects, according to an embodiment.

The feature map can be visualized as shown in FIG. 14 by converting high-dimensional data into low-dimensional data using a t-stochastic neighbor embedding (t-SNE) method. “High-dimensional” means that the number of features is large, and “two-dimensional” means that the number of features is two. 14(a) shows 2D features of 100 input images whose dimensions have been reduced by the learning model. 14(b) shows the 2-dimensional features of the feature maps extracted from 10 different final network convolution layers learned by the learning model. The features of the original image are randomly scattered, making it difficult to classify defects and normal states. However, the features obtained by the learned network can be grouped according to defect or normal state. Therefore, it is possible to classify the faulty state and the normal state.

The devices described above may be implemented as hardware components, software components, and/or a combination of hardware components and software components. For example, devices and components described in the embodiments may include, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA) , a programmable logic unit (PLU), microprocessor, or any other device capable of executing and responding to instructions. A processing device may run an operating system (OS) and one or more software applications running on the operating system. A processing device may also access, store, manipulate, process, and generate data in response to execution of software. For convenience of understanding, there are cases in which one processing device is used, but those skilled in the art will understand that the processing device includes a plurality of processing elements and/or a plurality of types of processing elements. It can be seen that it can include. For example, a processing device may include a plurality of processors or a processor and a controller. Other processing configurations are also possible, such as parallel processors.

Software may include a computer program, code, instructions, or a combination of one or more of the foregoing, which configures a processing device to operate as desired or processes independently or collectively. The device can be commanded. Software and/or data may be any tangible machine, component, physical device, virtual equipment, computer storage medium or device, intended to be interpreted by or provide instructions or data to a processing device. can be embodied in Software may be distributed on networked computer systems and stored or executed in a distributed manner. Software and data may be stored on one or more computer readable media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer readable medium. The computer readable medium may include program instructions, data files, data structures, etc. alone or in combination. Program commands recorded on the medium may be specially designed and configured for the embodiment or may be known and usable to those skilled in computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floptical disks. - includes hardware devices specially configured to store and execute program instructions, such as magneto-optical media, and ROM, RAM, flash memory, and the like. Examples of program instructions include high-level language codes that can be executed by a computer using an interpreter, as well as machine language codes such as those produced by a compiler.

As described above, although the embodiments have been described with limited examples and drawings, those skilled in the art can make various modifications and variations from the above description. For example, the described techniques may be performed in an order different from the method described, and/or components of the described system, structure, device, circuit, etc. may be combined or combined in a different form than the method described, or other components may be used. Or even if it is replaced or substituted by equivalents, appropriate results can be achieved.

Therefore, other implementations, other embodiments, and equivalents of the claims are within the scope of the following claims.

Claims (8)

In the soundness detection method of the power driving system performed by the soundness monitoring system,
generating an image using a noise signal measured in a power driving system;
receiving the generated image as an input to a learning model for diagnosing a defect; and
Classifying a defect pattern of the power driving system from the generated image using the learning model for diagnosing the defect
including,
The learning model for diagnosing the defect,
A method for detecting healthiness of a power driving system, characterized in that a defect pattern including a plurality of defect conditions and a normal condition is learned to be classified from an image. According to claim 1,
The generating step is
generating image data in the time-frequency domain by applying a continuous wavelet (CWT) to the measured noise signal;
Health detection method of a power driving system comprising a. According to claim 1,
The learning model,
An input layer into which the generated image is input, a convolution layer with 16 feature maps with a filter size of 3 × 3, a maximum pooling layer with a size of 2 × 2, and a filter size of 3 × 3 with 32 feature maps A method for detecting healthiness of a power driving system, characterized in that it consists of a convolution layer, a convolution layer having a maximum pooling layer having a size of 2 × 2, and a final convolution layer to which a softmax function is applied. According to claim 3,
The learning model,
extracting defect information from the image; and
Learning through the step of extracting a feature map including the extracted defect information
Health detection method of a power driving system, characterized in that. According to claim 1,
The power driving system includes parts including motors, chains, pinions, sprockets, gears, bearings, or rotating shafts, and a combination of the parts,
Classifying the defect pattern,
Extracting defect feature information related to a defect condition from the generated image using the learning model for defect diagnosis, and classifying a defect pattern of the power driving system based on the extracted defect feature information.
Health detection method of a power driving system comprising a. According to claim 1,
Classifying the defect pattern,
Diagnosing a defective part of the power driving system from the generated image, and classifying a single defect pattern or multiple defect patterns for the diagnosed defective part.
Health detection method of a power driving system comprising a. According to claim 1,
The defect pattern including the plurality of defect conditions and normal conditions,
Drive gear gear crack, drive gear gear breakage, drive shaft eccentricity in motor, outer ring defect due to hole in bearing, inner ring defect due to foreign matter in bearing, drive shaft eccentricity of motor + tooth breakage of drive gear, drive shaft eccentricity of motor + A method for detecting soundness of a power driving system comprising any one of gear breakage of drive gear + outer race defect of bearing, eccentricity of drive shaft of motor + gear breakage of drive gear + inner race defect of bearing, or normal condition. In the health monitoring system,
An image generator for generating an image using the noise signal measured in the power driving system;
an image input unit that receives the generated image into a learning model for diagnosing defects; and
A defect pattern classification unit classifying a defect pattern of the power driving system from the generated image using the learning model for diagnosing the defect.
including,
The learning model for diagnosing the defect,
A health detection system, characterized in that it is learned to classify a defect pattern including a plurality of defect conditions and normal conditions from the image.

Images