KR102416474B1 - Fault diagnosis apparatus and method based on machine-learning - Google Patents

Abstract

The present invention relates to a failure diagnosis technology, and more particularly, to a machine learning-based failure diagnosis apparatus and method that can effectively perform a failure diagnosis for equipment whose characteristics change over time even when learning only using steady state data.
To this end, the machine learning-based failure diagnosis apparatus according to the present invention includes a first analyzer that receives a sensing signal for a facility and outputs normal or abnormal for each time step within a predetermined time period of the sensing signal, and the sensing signal. A signal conversion unit that performs frequency conversion for each time step and outputs the frequency of each time step, and a second analysis that receives the frequency of each time step from the signal conversion unit and outputs normal or abnormal for the frequency of each time step and a classification unit configured to output whether or not the equipment has failed based on the output value of the first analysis unit and the output value of the second analysis unit.

Description

Fault diagnosis apparatus and method based on machine-learning

The present invention relates to a failure diagnosis technology, and more particularly, to a machine learning-based failure diagnosis apparatus and method that can effectively perform a failure diagnosis for equipment whose characteristics change over time even when learning only using steady state data.

Deep learning improves the performance of many known problems of artificial neural networks (i.e., vanishing problem, overfitting, etc.) by developing an activation function (ReLU) and improving algorithms such as drop-out. Also, thanks to the development of hardware such as GPU (Graphic Processing Units) and the power of big data that can learn complex structures, it is showing excellent performance in various fields recently.

Such deep learning technology is being developed rapidly by many overseas companies (Google, Facebook, Apple, Microsoft, Alibaba, Baidu) and is being applied to fields such as face recognition, voice recognition, natural language processing, search service, and medical care. It is urgent to secure the latest technology of deep learning, which is developing rapidly, and further preempt the application fields and rapidly commercialize it.

A defect classification method used in a conventional defect inspection equipment is shown in FIG. 1 .

An algorithm developer extracts features that are likely to be classified well from an image with an image processing algorithm, and then learns these features with a classifier (SVM, Decision Tree).

An optical device is constructed using lighting and a camera, and the S/N ratio of the defective area is increased by imaging the change in the amount of light entering the camera as the light path changes in the defective area. In such an image, a defect detection algorithm detects a defect candidate, and the defect is detected and classified using a feature extraction algorithm and a classification algorithm for the defect candidate image.

The performance limit of this method is how well a person can design a feature extraction algorithm to extract features. In addition, there is a problem in that classification performance is limited according to selected features, and classification performance varies according to image rotation, brightness change, size change, and the like. Since each product has different characteristics of the image, it is analyzed and developed, but there is a disadvantage that it takes a lot of time.

Among the latest technologies, deep learning, a classification algorithm applying the CNN method is a method in which artificial intelligence extracts and learns features from an image by itself. The above problem can be solved by constructing a defect classification algorithm by applying deep learning technology. Among the various deep learning structures, deep learning in the image field uses a structure called Convolutional Neural Network (CNN).

Image classification using CNN has the feature that CNN itself extracts and learns features that can improve classification performance. If this feature is applied to the image-based defect detection field, a dramatic performance improvement can be expected.

The configuration of a general classifier using deep learning is shown in FIG. 2 .

After applying a convolution layer and an activation function to the input image, the feature map is reduced in size through a pooling layer and transferred to the next convolution layer. do. By repeatedly and deeply stacking these basic structures, it is possible to effectively extract features for fault classification or fault diagnosis from images.

On the other hand, the smart factory includes technologies that are an evolved form of factory automation, that is, facility management using IoT, real-time diagnosis of the current state of the facility, and technology that can predict failures to take preemptive measures.

In particular, equipment status and fault diagnosis are essential technologies to prevent mass defects, ensure safety, stable operating conditions, and product quality. This technology is one of the large areas of failure prediction and health management (PHM: Prognostics and Health management).

In general, the configuration of an automated system for equipment failure diagnosis is as follows.

It collects signals that can indicate the status of equipment from various sensors such as vibration, displacement, temperature, and ultrasonic waves. These signals are transmitted in real time to the signal analysis PC. Signal analysis PC uses signal processing and deep learning technology to extract various statuses of equipment and diagnose failures. The detected fault and status information is sent to the data server, and this information is recorded in the data server together with product information produced by the equipment.

The conventional fault diagnosis method determines whether it is normal or abnormal based on the physical model of the facility. However, as the complexity of the facility increases and the operating state of the facility changes according to various environmental conditions, it is difficult to find a physical model accordingly. Recently, many approaches according to data analysis methods such as machine learning have been studied based on collected data rather than a physical model.

Specifically, it senses signals such as vibration, displacement, temperature, and ultrasound generated from an object, and based on these sensed values, time series analysis through linear prediction coefficients, frequency analysis through fast Fourier transform, and discrete analysis for each frequency band are performed. After analyzing the values and variances, the method of extracting the failure characteristics was adopted by classifying the results of testing and validation of these data through a multi-layer perceptron network.

In general, for facility diagnosis using deep learning, both normal and abnormal data must be secured. However, while it is easy to obtain steady state data, it is difficult to secure abnormal state data.

The method for fault diagnosis using only steady-state data is to use a model that receives steady-state data and predicts the next time step. In other words, it is a method of determining the difference between the predicted data and the actual input data, and if there is more than a certain difference, it is determined that the data is abnormal. In this method, there is a difficulty in optimizing a threshold, which is a criterion for determining that a specific difference or more occurs, and thus it is difficult to secure excellent diagnostic performance. In addition, in order to set the threshold, there must be abnormal data, so it is difficult to learn only from the actual normal data.

US Patent Publication No. 2020-0064822

The present invention was devised to solve the above problems, and an object of the present invention is to effectively diagnose the condition of a complex facility whose characteristics are variously changed over time even when learning using only steady state data.

To this end, the machine learning-based failure diagnosis apparatus according to the present invention includes a first analyzer that receives a sensing signal for a facility and outputs normal or abnormal for each time step within a predetermined time section of the sensing signal, and the sensing signal. A signal conversion unit that performs frequency conversion for each time step and outputs the frequency of each time step, and a second analysis that receives the frequency of each time step from the signal conversion unit and outputs normal or abnormal for the frequency of each time step and a classification unit configured to output whether or not the equipment has failed based on the output value of the first analysis unit and the output value of the second analysis unit.

In addition, the machine learning-based failure diagnosis method according to the present invention is a machine learning-based failure diagnosis method in which each step is performed in a machine learning-based failure diagnosis apparatus implemented with a computer, A first analysis step of outputting a normal or abnormal value, and a second analysis of receiving the frequency-converted frequency of each time step within a predetermined time section of the facility sensing signal and outputting a normal or abnormal value for the frequency of each time step and a classification step of outputting whether the equipment has failed based on the normal or abnormal values output in the first and second analysis steps.

In addition, the machine learning-based failure diagnosis method according to the present invention is a machine learning-based failure diagnosis method in which each step is performed in a machine learning-based failure diagnosis apparatus implemented with a computer. Determining whether the time step value is in a normal state or an abnormal state through a Recurrent Neural Networks (RNN) or Long Short-Term Memory (LSTM)-based model, and frequency conversion of each time step value within a predetermined time period of the sensing signal A step of receiving the input frequency and determining whether the frequency of each time step is in a normal state or an abnormal state through a CNN (Convolutional Neural Network)-based model, and whether the equipment is malfunctioning based on the normal or abnormal state determined through the two models It includes the step of judging.

In addition, the computer-readable recording medium according to the present invention includes a first analysis step of receiving each time step value within a predetermined time interval of time series data and outputting 0 or 1 for each time step, and within a predetermined time interval of the time series data A second analysis step of receiving the frequency-converted frequency of each time step value as input and outputting 0 or 1 for the frequency of each time step, and calculating an average of the values output in the first and second analysis steps Stores the program that performs the steps.

As described above, since the machine learning-based failure diagnosis technology according to the present invention uses a deep learning-based model with a dual structure that can be learned only with steady-state data, it is possible to solve the difficulty of securing abnormal-state data or optimizing the threshold. there is

In addition, since the present invention analyzes the sensing signal not only in the time domain but also in the frequency domain, there is an effect of further enhancing the failure diagnosis performance of the equipment.

The failure diagnosis technology using deep learning according to the present invention can be used not only as a core technology for diagnosing failures and conditions of various industrial facilities, but also in diagnostic fields such as non-destructive testing.

In addition, it will be applicable to technology that optimizes production conditions by judging the state of equipment according to production operating conditions and predicting product quality according to equipment conditions as an essential technology for implementing a smart factory.

In addition, the failure diagnosis apparatus according to the present invention can be used in the field of product defect classification as well as equipment failure diagnosis. That is, as an essential technology for implementing a smart factory, it can also be used to determine product quality according to production operating conditions.

1 is a view for explaining a conventional defect classification method.
2 is a view for explaining a general deep learning-based classification method.
3 is a diagram showing a schematic configuration of a machine learning-based failure diagnosis apparatus according to the present invention.
4 is a view showing a configuration for learning the first analysis unit according to the present invention.
5 is a diagram showing a configuration for learning a second analysis unit according to the present invention.
6 is a flowchart illustrating a processing process of a machine learning-based failure diagnosis method according to the present invention.

Hereinafter, with reference to the accompanying drawings, the embodiments of the present invention will be described in detail so that those of ordinary skill in the art to which the present invention pertains can easily implement them.

However, the present invention may be embodied in several different forms and is not limited to the embodiments described herein.

And in order to clearly explain the present invention in the drawings, parts irrelevant to the description are omitted, and similar reference numerals are attached to similar parts throughout the specification.

Throughout the specification, when a part "includes" a certain component, it means that other components may be further included, rather than excluding other components, unless otherwise stated.

In addition, the terms "... unit" and "... module" described in the specification mean a unit that processes at least one function or operation, which may be implemented as hardware or software or a combination of hardware and software.

Hereinafter, a machine learning-based failure diagnosis apparatus and method according to an embodiment of the present invention will be described in detail with reference to the drawings.

3 shows the configuration of a machine learning-based failure diagnosis apparatus according to the present invention.

Referring to FIG. 3 , the machine learning-based failure diagnosis apparatus includes a first analysis unit 10 , a signal conversion unit 20 , a second analysis unit 30 , a classification unit 40 , and the like.

The first analyzer 10 receives the sensing signal for the facility and outputs normal or abnormal for each time step within a predetermined time section of the sensing signal. The first analyzer 10 outputs 0 when the value of the time step is normal, and outputs 1 when it is abnormal.

A model based on Recurrent Neural Networks (RNN) or Long Short-Term Memory (LSTM) is applied to the first analyzer 10 according to an embodiment of the present invention.

Although not shown in FIG. 3, the sensing signal is cropped to a predetermined window at regular intervals by the preprocessor, and then the signal magnitude is extracted at every sampling period (ie, time step) and converted into a one-dimensional signal vector. It is input to the first analysis unit 10 .

Here, the sensing signal refers to a signal indicating the state of the facility or object by installing various sensors in a facility (factory equipment, equipment) or an object (composite structure, product). Sensors attached to equipment include vibration, displacement, temperature, and ultrasonic sensors.

The signal converter 20 frequency-converts the sensing signal to generate a frequency. The signal converter 20 frequency-converts a sensing signal that changes according to time using a Fourier transform (FFT) for each time step, and outputs a frequency of each time step.

The second analyzer 30 receives the frequency of each time step and outputs normal or abnormal with respect to the frequency of each time step. The second analyzer 10 outputs 0 when the frequency of the time step is normal, and outputs 1 when it is abnormal.

A Convolutional Neural Network (CNN)-based model is applied to the second analysis unit 10 according to an embodiment of the present invention.

Although not shown in FIG. 3 , the frequency signal of each time step is converted into a one-dimensional frequency signal vector by the preprocessor and input to the second analysis unit 30 .

The classification unit 40 outputs whether the equipment has failed based on the output value of the first analysis unit 10 and the output value of the second analysis unit 20 . Since the output values of the first analysis unit 10 and the second analysis unit 20 are 0 and 1, the classification unit 40 calculates the average of the output values and classifies it as normal if the average value is close to 0, and the average value is If it is close to 1, it is classified as abnormal.

4 shows a process of learning the model applied to the first analysis unit 10 of the present invention.

Referring to FIG. 4 , the first model 12 is a module for predicting data of the next time step with respect to inputted time series data (actual data). The first model 12 receives X(n-1), which is data when t=n-1 of time series data, and outputs X'(n), which is prediction data when t=n, which is the next time step. . The time series data input to the first model 10 is steady state data. The first model 12 is trained so that the value output from the second model 20 becomes 0 (normal value).

The second model 14 is a module for discriminating between real data and predicted data. The second model 20 receives X(n), which is actual data when t=n, and X'(n), which is predicted data, and divides actual data and predicted data.

The second model 14 receives the real data and the predicted data output from the first model 12 and learns to output 0 (normal value) for the real data and 1 (abnormal value) for the predicted data do. If the predicted data is the same as the actual data, the second model 20 will output 0 for all input data.

In this way, the first model 12 is trained so that the second model 14 outputs 0 for all input data, that is, the predicted data becomes real data, and the second model 14 is 0 for the real data. It is learned to output , and output 1 for the prediction data, that is, to accurately classify the real data and the predicted data.

When the learning of the first model 12 and the second model 14 is completed, the first model 12 is used to learn a model to be applied to the second analysis unit 30 , and the second model 14 . is used as a model to be applied to the first analysis unit 30 .

The first model 12 and the second model 14 may be trained using a model such as Recurrent Neural Networks (RNN) or Long Short-Term Memory (LSTM).

5 shows a process of learning a model applied to the second analysis unit 10 of the present invention.

Referring to FIG. 5 , the third model 30 is a module for discriminating frequency-converted data of real data (hereinafter referred to as an actual frequency) and frequency-converted data of prediction data (hereinafter, referred to as a predicted frequency).

The first model 12 receives X(n-1), which is data when t=n-1 of time series data, and outputs X'(n), which is prediction data when t=n, which is the next time step. .

The signal conversion unit 20 receives X'(n) from the first model 12 and receives X(n), which is data when t=n of time series data, as input, X'(n) and X(n) ) is frequency converted.

The third model 32 receives the frequency-converted value from the signal conversion unit 20 and calculates the frequency-converted value (predicted frequency) of X'(n) and the frequency-converted value (actual frequency) of X(n). separate The third model 30 is trained to output 0 (normal value) for the actual frequency and 1 (abnormal value) for the predicted frequency.

When the learning of the first model 12 and the third model 32 is completed, the third model 32 is used as a model to be applied to the third analysis unit 30 . The third model 32 may be trained using a model such as a Convolutional Neural Network (CNN).

When configuring the third model 32 using a CNN model, the third model 32 includes a plurality of convolutions, batch normalization, activation function (ReLU), and maximization pooling. It is composed of layers and the signal feature vector is extracted from the frequency input vector of each time step. The element values of the signal feature vectors extracted in this way are classified into actual frequencies or predicted frequencies through a neural network of a fully connected layer.

6 shows a processing process of a machine learning-based failure diagnosis method according to the present invention.

Referring to FIG. 6 , first, as described above with reference to FIGS. 4 and 5 , a learning step S10 for the first model 12 , the second model 14 , and the third model 32 is performed.

In the learning step (S10), the first model 12 is trained so that the data of the next time step of the time series data is predicted so that the second model 14 can output 0 for all input data, and the second model ( 14) is trained to correctly classify the real data and the predicted data so that 0 is output for the real data and 1 is output for the predicted data, and the third model 32 outputs 0 for the actual frequency and outputs 0 for the predicted data It is trained to correctly classify the actual frequency and the predicted frequency so that 1 can be output for .

When the learning process is completed in this way, the second model 14 is applied to the first analysis unit 10 , and the third model 32 is applied to the second analysis unit 30 , and the first analysis unit 10 . and the second analysis unit 30 analyzes the time series data (sensing signal), respectively (S20 and S30).

That is, the first analysis unit 10 performs a first analysis step (S20) of outputting a normal value (0) or an abnormal value (1) for each time step within a predetermined time interval of the sensing signal.

In addition, the second analyzer 20 receives the frequency-converted frequency of each time step within a predetermined time section of the sensing signal and outputs a normal value (0) or an abnormal value (1) for the frequency of each time step. 2 The analysis step (S30) is performed.

Next, the average of the normal or abnormal values output in the first analysis step (S20) and the second analysis step (S30) is calculated (S40), and a fault diagnosis result of the equipment is output based on the average of the output values (S50) .

The above description is merely illustrative of the present invention, and various modifications may be made by those of ordinary skill in the art to which the present invention pertains without departing from the technical spirit of the present invention.

Therefore, the embodiments disclosed in the specification of the present invention do not limit the present invention. The scope of the present invention should be construed by the following claims, and all technologies within the scope equivalent thereto should be construed as being included in the scope of the present invention.

10: first analysis unit 12: first model
14: second model 20: first analysis unit
20: signal conversion unit 30: second analysis unit
32: third model 40: classification unit

Claims (13)

To predict the data of the next time step of the time series data, which is the equipment sensing signal in a steady state, receive the predicted data and actual data output from the first model trained to output 0 for the actual data and output 1 for the predicted data A first analysis unit that receives a sensing signal for a facility using the learned second model and outputs normal or abnormal for each time step within a predetermined time section of the sensing signal;
a signal converter for frequency-converting the sensing signal for each time step and outputting a frequency for each time step;
Using the third model trained to receive the frequency-converted prediction frequency of the prediction data output from the first model and the frequency-converted actual frequency of the actual data, and output 0 for the actual frequency and 1 for the predicted frequency a second analysis unit that receives the frequency of each time step from the signal conversion unit and outputs normal or abnormal for the frequency of each time step;
and a classification unit configured to output whether the equipment has failed based on the output value of the first analysis unit and the output value of the second analysis unit. delete The method of claim 1,
The second model applied to the first analysis unit is a machine learning-based failure diagnosis apparatus, characterized in that it is a Recurrent Neural Networks (RNN) or a Long Short-Term Memory (LSTM). delete The method of claim 1,
The third model applied to the second analysis unit is a machine learning-based failure diagnosis apparatus, characterized in that it is a Convolutional Neural Network (CNN). In a machine learning-based failure diagnosis method in which each step is performed in a computer-implemented machine learning-based failure diagnosis device,
Using the second model trained to output 1 for the data predicted by the first model that predicts the data of the next time step of the time series data, which is the equipment sensing signal in a steady state, and 0 for the actual data, A first analysis step of outputting a normal or abnormal value for each time step within a predetermined time interval;
Using the third model learned to output 1 for the predicted frequency obtained by frequency-converting the data predicted by the first model and output 0 for the actual frequency obtained by frequency-converting the actual data within a predetermined time interval of the facility sensing signal a second analysis step of receiving the frequency-converted frequency of each time step and outputting a normal or abnormal value for the frequency of each time step;
and a classification step of outputting whether the equipment has failed based on the normal or abnormal values output in the first and second analysis steps. 7. The method of claim 6,
The frequency-converted frequency is a machine learning-based failure diagnosis method, characterized in that it is a frequency by a Fourier transform (FFT). delete delete delete 7. The method of claim 6,
The classification step calculates an average value of the output normal or abnormal values and outputs a normal or abnormal state of the equipment based on the average value. delete delete

Images