KR102471201B1 - Outlier detection and automatic threshold system for unsupervised learning-based time series data - Google Patents

Abstract

The present invention is an outlier detection and automatic threshold point system for unsupervised learning-based time series data. The system comprises: a database unit that stores time series data collected from an open source and sensors in a factory; a data preprocessing unit that refines the data to extract and distinguish features for detecting outliers from the time series data; an outlier detection unit that trains the data preprocessed by the data preprocessing unit in an outlier prediction model by using an autoencoder structure in adversarial machine learning and detects outliers; a threshold calculation unit that derives a threshold for finding an outlier using a machine learning algorithm based on the outlier prediction model; and an outlier distinguishing unit that distinguishes an outlier of the time series data based on the threshold. The present invention aims to detect outliers of the time series data not labeled in advance by applying an outlier detection algorithm and deep learning technology.

Description

Outlier detection and automatic threshold system for unsupervised learning-based time series data}

The present application relates to an outlier detection and automatic threshold point system for time series data based on unsupervised learning.

An anomaly detection system using data collected through sensors utilizes supervised learning and unsupervised learning-based ai technology depending on whether or not the collected data is labeled. Supervised learning requires pre-labeled data as normal and abnormal, but since abnormalities occur very rarely in actual industrial sites, it takes time and money to acquire abnormal data and label the collected data.

In order to overcome the above problem, an anomaly detection system based on unsupervised learning is used, and learning is performed under the assumption that the collected data is normal. In general, abnormal data is detected by obtaining the difference between input and output through the process of reducing and restoring training data.

However, unsupervised learning-based performance is dependent on hyper-parameters, so the accuracy is unstable compared to supervised learning-based anomaly detection systems, and the critical point for detecting anomalies must be calculated or arbitrarily specified.

Therefore, industries that need to introduce anomaly detection systems must make choices such as investing time and money in collecting and labeling abnormal data in advance, or considering unstable accuracy.

Therefore, there is no need for abnormal data collection and pre-labeling work, and a technology that allows learners to detect outliers without separately calculating critical points is needed.

The background technology of the present application is disclosed in Korean Patent Registration No. 10-1916926.

The present invention is to solve the above-mentioned problems of the prior art, and an object of the present invention is to detect an outlier by calculating a critical point of unlabeled time series data by applying an anomaly detection algorithm and deep learning technology.

The present invention is to solve the above-mentioned problems of the prior art, to learn time series data that have not been labeled as normal or abnormal in advance with artificial intelligence, and to calculate critical points using machine learning techniques to discriminate outliers that are not included in the normal category. The purpose.

However, the technical problem to be achieved by the embodiments of the present application is not limited to the technical problems described above, and other technical problems may exist.

As a technical means for achieving the above technical problem, an outlier detection and automatic threshold point system of unsupervised learning-based time series data according to an embodiment of the present application includes a database unit for storing time series data collected from open source and factory sensors, A data pre-processing unit that extracts features for detecting outliers from the time series data and refines the data to make them discriminable. It may include an outlier detection unit that detects an outlier, a threshold calculation unit that derives a critical point for finding an outlier using a machine learning algorithm based on the outlier prediction model, and an outlier discriminator that determines an outlier of the time series data based on the critical point. can

According to an embodiment of the present application, the data pre-processing unit may extract statistical characteristic variables after pre-processing the time-series data stored in the database unit for learning.

Also, the autoencoder may be constructed with a first autoencoder and a second autoencoder sharing one encoder for adversarial machine learning.

In addition, the first autoencoder restores the input data, learns a model that deceives the second autoencoder, the second autoencoder restores the input data, and reconstructs the reconstruction data and the input data restored by the first autoencoder. It may be to learn a model that discriminates.

In addition, the second autoencoder distinguishes between real data and reconstruction data of the first autoencoder, and the first autoencoder performs adversarial machine learning to deceive the second autoencoder, so that the autoencoder generates outliers similar to normal data. can detect

In addition, for stable learning, the autoencoder may assign a weight to the data reconstruction error of the first autoencoder at the beginning of learning, and weight an outlier similar to normal data in the second half of learning.

In addition, the threshold point calculation unit may calculate a threshold point for classifying the normal data learned by the outlier detection unit and the outlier based on a K-means algorithm, and may define the outlier based on the calculated threshold point.

In addition, in the outlier detection and automatic threshold point system of unsupervised learning-based time series data, storing time series data collected from open source and factory sensors, extracting characteristics for outlier detection from the time series data and data to be able to discriminate A data pre-processing step of refining , learning the data pre-processed in the pre-processing unit in an outlier prediction model using an autoencoder structure in adversarial machine learning, and detecting an anomaly, using a machine learning algorithm based on the outlier prediction model. The method may include a threshold point calculation step of deriving a threshold point at which an outlier can be found, and a step of determining an outlier value of the time-series data based on the threshold point.

The above-described problem solving means are merely exemplary and should not be construed as intended to limit the present disclosure. In addition to the exemplary embodiments described above, additional embodiments may exist in the drawings and detailed description of the invention.

According to the above-described problem solving means of the present application, it is possible to determine normal or abnormal time series data only with the previously collected normal data without defining an abnormal state in advance.

According to the above-described problem solving means of the present invention, it is possible to automatically find an optimal threshold point without defining a threshold point for detecting an outlier in advance.

The present application can save the time and cost of collecting and labeling abnormal data in advance through outlier detection through unsupervised learning.

However, the effects obtainable herein are not limited to the effects described above, and other effects may exist.

1 is a diagram showing a schematic configuration of an outlier detection and automatic threshold point system of unsupervised learning-based time series data according to an embodiment of the present invention.
Figure 2a is a diagram schematically showing the structure of an autoencoder in adversarial machine learning according to an embodiment of the present invention.
Figure 2b is a diagram schematically showing two decoders sharing one encoder for adversarial machine learning according to an embodiment of the present invention.
3 is a diagram illustrating an example of a method of detecting outliers and automatically calculating threshold points of unsupervised learning-based time series data according to an embodiment of the present invention.

Hereinafter, embodiments of the present application will be described in detail so that those skilled in the art can easily practice with reference to the accompanying drawings. However, the present disclosure may be implemented in many different forms and is not limited to the embodiments described herein. And in order to clearly describe the present application in the drawings, parts irrelevant to the description are omitted.

Throughout the present specification, when a part is said to be “connected” to another part, it is not only “directly connected”, but also “electrically connected” or “indirectly connected” with another element in between. "Including cases where

Throughout the present specification, when a member is referred to as being “on,” “above,” “on top of,” “below,” “below,” or “below” another member, this means that a member is located in relation to another member. This includes not only the case of contact but also the case of another member between the two members.

Throughout the present specification, when a part "includes" a certain component, it means that it may further include other components without excluding other components unless otherwise stated.

1 is a diagram showing a schematic configuration of an outlier detection and automatic threshold point system 100 of unsupervised learning-based time series data according to an embodiment of the present invention.

Referring to FIG. 1, the outlier detection and automatic threshold point system 100 of unsupervised learning-based time series data includes a database unit 110, a data pre-processing unit 120, an outlier detection unit 130, a threshold calculation unit 140, An outlier determination unit 150 may be included.

Unsupervised learning-based time series data outlier detection and automatic threshold point system 100 does not define an abnormal state in advance, determines normality and abnormality of time series data only with the normal data collected, and defines a threshold for detecting outliers. By automatically finding the optimal threshold point without labeling, the time and cost of labeling in advance can be saved.

The database unit 110 is a component in which various types of data are stored, and time-series data collected from open sources and factory sensors are stored, and input data for factory facilities having predetermined parameter values and output data corresponding thereto are stored. Saved.

In addition, open source data and sensor data collected from factories in actual operation include only structured data that includes time-series information.

For example, the database unit 110 may be implemented to enable WIP Transaction Log collection, analysis, and processing, and may be implemented to enable facility status change, alarm, log, sensor data collection, analysis, and processing, and production It can also be implemented to enable collection, analysis, and processing of flow history and indicators.

The data pre-processing unit 120 may extract features for detecting outliers from time-series data and refine the data to enable discrimination. According to one embodiment of the present application, the characteristic variable is the shape (Shape), size (Amplitude), phase (Phase) of the waveform extracted from the waveform representing the change in vibration acceleration over time for the rotating equipment, and the ampere of the electric motor. . In other words, the data pre-processing unit 120 may extract characteristic variables for at least one of shape, size, phase, temperature of the engine body and loop, circulating flow rate, and current and voltage of the electric motor from the vibration data.

Key variable data identified through the data pre-processing unit 120 may change depending on whether or not a sensor is attached and how it is collected.

In addition, the data pre-processing unit 120 may extract statistical characteristic variables after pre-processing the time-series data stored in the database unit 110 for learning.

Statistical characteristic (characteristic) variables derived from the database unit 110 may include variables that may have numerical values related to vibration data. For example, the database unit 110 derives (extracts) at least one feature (characteristic) variable from among RMS, Variance, Skewness, Kurtosis, Shape factor, Crest factor, Impulse factor, Margin factor, and Peak to Peak in the time domain. )can do. In addition, the pre-processing unit may derive (extract) at least one characteristic (feature) variable among FC, RMSF, RVF, SS, and SK in the frequency domain. The RMS feature variable included in the time domain may include a feature that gradually increases as a defect occurs. In addition, the variance characteristic variable included in the time domain may include a feature measuring the degree to which a signal is far from the average. In addition, the skewness feature variable included in the time domain may include a feature representing information whose distribution is out of symmetry and skewed to one side. In addition, the Kurtosis feature variable included in the time domain may include a feature representing sharp information of the probability density function. In addition, the shape factor characteristic variable included in the time domain may include a feature representing a change according to the shape of the equipment. In addition, the crest factor characteristic variable included in the time domain may include a feature representing the degree of sharpness of various waveforms. In addition, the impulse factor characteristic variable included in the time domain may include a feature of measuring a signal of a shock wave. In addition, the margin factor characteristic variable included in the time domain may include a feature of measuring a shock wave signal between the rotating element and the track. In addition, the Peak to Peak (PTP) characteristic variable included in the time domain may include a feature representing amplitude representing severe information of vibration and noise.

In addition, the FC characteristic variable included in the frequency domain may include a feature indicating a positional change of the main frequency. In addition, the RMSF feature variable included in the frequency domain may include a feature indicating a positional change of the main frequency. In addition, the RCF feature variable included in the frequency domain may include a feature indicating the matching of the spectrum. In addition, the SS feature variable included in the frequency domain may include a feature indicating symmetry of the amplitude spectrum distribution around the mean. In addition, the SK feature variable included in the frequency domain may include a feature used to measure a distribution value and compare it with a normal distribution.

By extracting feature variables including each feature in the time domain and the frequency domain from the database unit 110, the outlier detection unit 130 may generate a highly accurate outlier prediction model. In addition, by applying a PCA (multivariate data analysis) technique in the database unit 110, it is possible to easily understand the data by reducing the dimension of the correlation between various variables to a small number of principal components.

In addition, according to an embodiment of the present application, the data pre-processing unit 120 removes noise by applying a preset low-pass filter or band-pass filter to the received data can do.

Here, the noise may include Gaussian noise, salt and pepper noise, and the like.

In addition, the data pre-processing unit 120 performs RMS analysis, center frequency (CF) analysis, variance (Var) analysis, intermediate frequency (IF) analysis, skewness analysis, medium wave (MF) analysis, and kurtosis for vibration data. Variables (characteristic variables) required for one or more predictions may be extracted from the vibration data based on at least one of a Kurtosis analysis and a PTP analysis.

Exemplarily, the data pre-processor 120 applies Fast Fourier Transform (FFT) or Short Time Fourier Transform (STFT) to noise-removed vibration data to convert noise-removed vibration data in the frequency domain. It can be converted into data or data in the time-frequency domain. In other words, the data pre-processing unit 120 performs pre-processing on the collected data, but transforms the collected time waveform data (vibration data in the time domain) into the frequency domain through fast Fourier transform (FFT) or spectroscopy. Transformation into the time-frequency domain through the gram is performed, and variables (characteristic variables) required for prediction in each transformed domain can be extracted. The data preprocessed by the data preprocessor 120 may be used as learning data of the outlier detection unit 130 .

According to an embodiment of the present application, the outlier detection unit 130 may learn the data preprocessed by the preprocessor in an outlier prediction model using the structure of the autoencoder 200 in adversarial machine learning and detect the anomaly.

According to an embodiment of the present application, encoding data extraction may extract encoding data by applying basic information data to a neural network model. In other words, encoding data extraction may extract encoding data through an Encoder-Decoder LSTM model (neural network model) based on data stored in the database unit 110 . Encoding data extraction may extract encoding data by applying time-sequentially collected information to a neural network model (eg, LSTM) for each of the collected maintenance history data and state data.

The outlier prediction model (neural network model) described below refers to an encoder and a decoder that are responsible for a certain function within the LSTM model. It is a generative model that reconstructs time series signals. The LSTM model is trained with an objective function that minimizes the error by comparing the reconstructed value with the actual value.

The objective function can use MSE (Mean Squared Error). MSE is a measure of the concept of distance and means a value that can quantitatively confirm the difference between two objects.

Figure 2a is a diagram schematically showing the structure of an autoencoder 200 in adversarial machine learning according to an embodiment of the present invention.

Encoding data extraction may generate encoding data of time t data by applying basic information data at time t to an outlier prediction model for time series data. The encoding data extraction unit may generate encoding data by applying each of the t-n th basic information data from time t to before time t to the neural network model with respect to the basic information data collected in time series.

For example, in encoding data extraction, if the time series data is text data, the encoder sequentially receives all the words of the input sentence and finally compresses all the word information to create a vector, which is a contest vector. (context vector).

According to an embodiment of the present application, an LSTM cell receiving input data may be referred to as an encoder, and an LSTM cell outputting output data may be referred to as a decoder. Exemplarily, when the basic information data is text (TEXT) data or sentence data, it is divided into word units through word tokenization, and each word token may be input at each time point of the LSTM cell. The hidden state of the last time of the encoder LSTM cell can be passed to the decoder LSTM cell, which can be referred to as a vector. A vector can be used as the first hidden state of a decoder LSTM cell. However, the above-described information is only an example and is not limited thereto.

Illustratively, the outlier prediction model (neural network model) may include two modules, an encoder module and a decoder module. The Encoder module can encode (encode) input data and the Decoder can decode (decode) the encoded data.

The outlier prediction model disclosed herein may be characterized by using the structure of an LSTM-based model, which is a time series analysis algorithm, to extract characteristics of encoded data from time series data. In particular, the LSTM (Long Short Term Memory networks) algorithm is one of the artificial recursive neural network (RNN) architectures used in the field of deep learning, and unlike feed-forward neural networks, there is a feedback connection. Therefore, according to the LSTM algorithm, there is an advantage that learning and processing can be performed not only for a single data point but also for an entire data sequence.

These LSTM algorithms are suitable for classifying, processing, and predicting predictions based on time series data, and LSTM has the advantage of resolving the vanishing gradient problem that can occur in training through traditional RNNs.

Meanwhile, in the outlier prediction model, the input and output data do not necessarily have the same shape. That is, the window size of input data and the window size of output data can be set differently. For example, the anomaly detection system can perform learning by configuring data for the past 8 weeks as input data and configuring data for the future 4 weeks as output data, with the goal of predicting whether maintenance will occur within 4 weeks. In the outlier prediction model, each encoder and decoder part is not composed of one hidden layer, but is composed of multiple hidden layers. Therefore, in order to extract encoded data, it is necessary to select the output value of the last layer of the encoder part composed of a plurality of hidden layers, and this value can be stored and used.

Figure 2b is a diagram schematically showing two decoders sharing one encoder for adversarial machine learning according to an embodiment of the present invention.

Referring to FIGS. 2A and 2B , an autoencoder 200 may be composed of a first autoencoder 201 and a second autoencoder 202 sharing one encoder.

2A and 2B, after the encoder compresses the input data, the decoder restores the compressed data to data following a normal distribution of the input data. A first autoencoder 201 and a second autoencoder 202 can learn each.

An autoencoder in the field of anomaly detection learns using only data defined as “normal”, and then determines the presence or absence of anomaly through reconstruction loss generated by the difference between input and output at the time of anomaly detection.

That is, when abnormal data is input, a large loss value is generated by a decoder trained to generate only normal data, and normal and abnormal data can be distinguished based on a threshold of a loss value of a certain size or more.

Referring to FIG. 2B, the reconstruction data restored by the first autoencoder 201 is received and input to the second autoencoder 202, the second autoencoder 202 distinguishes between real data and reconstruction data, and The autoencoder 201 may perform adversarial machine learning to trick the second autoencoder 202 well.

Adversarial machine learning can be a network composed of generators and discriminators.

Through this, the second autoencoder 202 can detect an outlier similar to normal data.

According to an embodiment of the present application, the first autoencoder 201 may serve as a generator, and the second autoencoder 202 may serve as a discriminator.

The first autoencoder 201 maps the distribution of real data to the latent space and generates sophisticated fake data by receiving variables derived from the distribution, and the second autoencoder 202 maps the generated fake data and real data. data can be distinguished.

Adversarial machine learning can be learned by using a game-theoretic type of objective function to find a balance point (Nash equilibrium) while competing with two players (the first autoencoder and the second autoencoder).

Anomaly detection using adversarial machine learning optimizes latent variables according to the detection data using the first autoencoder 201 and the second autoencoder 202 that have been trained, and then each generated reconstruction loss value and discriminant loss An ideal score can be derived by weighted summing the values.

That is, adversarial machine learning can generate fake data that simulates normality by receiving latent variables that assume a normal distribution as input values, and derive anomaly scores using reconstruction loss values between real data and fake data.

Therefore, the first autoencoder 201 of the outlier detection unit 130 restores the input data, learns a model that deceives the second autoencoder 202, and the second autoencoder 202 restores the input data and , A model for distinguishing the reconstruction data restored by the first autoencoder 201 and the input data can be learned.

That is, in the anomaly detection unit 130, the second autoencoder 202 distinguishes between the real data and the reconstruction data of the first autoencoder 201, and the first decoder deceives the second autoencoder 202. By learning, the autoencoder can detect outliers similar to normal data. Through this adversarial machine learning, it is possible to create a reconstruction group with a similar distribution by learning data that is considered normal. However, since the conventional adversarial autoencoder model can make the restoration error small even for abnormal data, the anomaly detection unit 130 according to the present invention divides the autoencoders into the first autoencoder 201 and the second autoencoder 202. ), and for stable learning, weight can be given to the data reconstruction error of the first decoder at the beginning of learning, and weighted to outliers similar to normal data in the second half of learning.

The autoencoder 200 may be a neural network structure mainly used in an unsupervised learning methodology. The autoencoder 200 is an unsupervised machine learning model in the form of reducing the dimension of an input value and restoring it again, and may have a function of learning characteristics of values used for learning. More specifically, the autoencoder 200 can learn a function that approximates an output value to an input value, extracts a feature for an input value through an encoder, and reconstructs an input value through a decoder.

A neural network (neural network) may be composed of a set of interconnected computational units, which may be generally referred to as nodes, and these nodes may be referred to as neurons. A neural network is generally composed of a plurality of nodes. Nodes constituting a neural network may be interconnected by one or more links.

Some of the nodes constituting the neural network may configure one layer based on distances from the first input node. For example, a set of nodes having a distance of n from the first input node may constitute n layers.

The neural network described in this specification may include a deep neural network (DNN) including a plurality of hidden layers in addition to an input layer and an output layer. Deep neural networks may include convolutional neural networks (CNNs), recurrent neural networks (RNNs), generative adversarial networks (GANs), and the like.

The neural network structure of the autoencoder 200 can be defined by the number of nodes, the connection relationship of links interconnecting the nodes, and the weights assigned to each link.

Also, in the autoencoder 200, the number of nodes constituting the hidden layer may be less than the number of nodes constituting the input layer or the output layer.

A set of input layers and hidden layers, which are subsets of layers constituting the autoencoder 200, may be defined as an encoder network as one neural network. In addition, a set of hidden layers and output layers may be defined as a decoder network as another neural network. An output layer of an encoder network can be an input layer of a decoder network.

The autoencoder 200 compresses the data input to the input layer of the encoder network through the hidden layer, outputs the result as a latent vector, and outputs the output latent variable through the hidden layer of the decoder network to the output layer. You can perform the function of restoring through

Unsupervised learning is used to discover hidden features or structures of data. Illustratively, unsupervised learning may include k-means divided into clustering, hierarchical cluster analysis (HCA), and expectation maximization. In addition, Principal Component Analysis (PCA), Kernel PCA, Locally-Linear Embedding (LLE, Locally-Linear Embedding), t-SNE ( t-distributed StochasticNeighbor Embedding). In addition, Apriori and Eclat, which are classified as association rule learning, may be included.

For example, a logistic regression algorithm, a random forest algorithm, a support vector machine (SVM) algorithm, a decision-making algorithm, and a clustering algorithm may be used as the unsupervised learning-based artificial intelligence algorithm, but are not limited thereto. The Random Forest algorithm is an algorithm that forms a forest with numerous decision trees and averages each prediction result into one outcome variable. It is a non-probabilistic algorithm. The Extra Tree algorithm is similar to the Random forest, but is faster than the Random forest. The XGBoost algorithm is a boost algorithm that applies the result of the XGBoost tree to the next tree if the Random forest tree is independent. The deep learning algorithm is an algorithm that learns by adjusting the effect of variable patterns on results based on a multi-layered neural network with weights. In addition, the K-means clustering algorithm is a traditional classification technique that repeatedly subdivides a target group into K groups based on the average value of distance (similarity). It is a learning and clustering technique. In addition, the EM & Canopy algorithm refers to a technique of clustering by updating parameter values through an iterative process, starting from the one with the maximum probability with a given initial value.

Hierarchical clustering algorithms allow each group to be subdivided into smaller groups. Visualization algorithms take large, unlabeled, high-dimensional data and create a 2D or 3D representation that can be plotted. Dimensionality reduction can be used to simplify data without losing too much information. For example, since a car's mileage is highly correlated with its age, a dimensionality reduction algorithm can combine the two features into a single feature representing the wear and tear of the car, which is called feature extraction. Outlier detection is an operation of automatically removing strange values from the database unit 110 before being injected into the learning algorithm, and can be trained with normal samples and determine whether new samples are normal or not. Association rule learning finds interesting relationships between features in a large amount of data, and is used, for example, to find that people who purchase a certain product tend to purchase other products. However, the artificial intelligence algorithm based on unsupervised learning is not limited thereto, and various algorithms may be included according to the results of research and development.

According to an embodiment of the present disclosure, the critical point calculation unit 140 may derive a critical point capable of finding an outlier by using a machine learning algorithm.

In addition, the threshold calculation unit 140 may derive an optimal threshold point for finding an outlier by clustering patterns defined by the outlier detection unit 130 .

Exemplarily, the threshold calculation unit 140 calculates a threshold for classifying normal data and abnormal data learned by the outlier detection unit 130 based on a K-means algorithm, and defines the outlier based on the calculated threshold. can do.

The K-means algorithm arbitrarily divides normal data and abnormal data learned in the outlier detection unit 130 into k clusters and divides the average of the clusters into representative values, and data can be rearranged based on similarity. .

In addition, an anomaly score may be calculated based on the output data output by the anomaly detection unit 130 . Here, the ideal score may be scored based on the distribution of latent variables, or may be scored based on the difference between data reconstructed through an autoencoder and original data.

Thereafter, it is possible to determine whether the input data input to the autoencoder is abnormal by thresholding the calculated anomaly score. Here, as a threshold value for distinguishing between normal and abnormal input data through the calculated abnormality score, the user can find an optimal threshold and detect abnormal data based on this threshold. For example, the average and standard deviation of restoration errors for normal data may be obtained, and a threshold may be set using the values.

The outlier determination unit 150 may determine an outlier value of the time series data based on the threshold point derived by the threshold point calculation unit 140 .

The outlier determination unit may include an algorithm for determining whether newly input data is normal or abnormal.

3 is a diagram illustrating an example of a method of detecting outliers and automatically calculating threshold points of unsupervised learning-based time series data according to an embodiment of the present invention.

The outlier detection and automatic threshold point calculation of the unsupervised learning-based time series data shown in FIG. 3 may be performed by an outlier detection and automatic threshold point calculation device. Therefore, even if the contents are omitted below, the description of the outlier detection and automatic threshold point system 100 of unsupervised learning-based time series data is the same as the description of the method of detecting outliers and automatically calculating threshold points of unsupervised learning-based time series data. can be applied

In step S301, time-series data collected from open source and factory sensors may be stored.

In step S302, data preprocessing may be performed to refine the data so that features for detecting outliers are extracted from the time series data and discriminable.

In step S303, using the structure of the autoencoder 200 in adversarial machine learning, the preprocessed data in the preprocessor can be learned from an outlier prediction model, and anomalies can be detected.

In step S304, a critical point for deriving a critical point for finding an outlier may be calculated using a machine learning algorithm.

In step S305, an outlier of the time series data may be determined based on a critical point.

The system 100 for automatic threshold detection and automatic threshold point detection of unsupervised learning-based time series data according to an embodiment of the present disclosure may be implemented in the form of program instructions that can be executed through various computer means and may be recorded on a computer readable medium. The computer readable medium may include program instructions, data files, data structures, etc. alone or in combination. Program instructions recorded on the medium may be specially designed and configured for the present invention, 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, etc. 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. The hardware devices described above may be configured to act as one or more software modules to perform the operations of the present invention, and vice versa.

In addition, the above-described method for monitoring vibration conditions of rotating equipment using time series analysis based on deep learning may be implemented in the form of a computer program or application stored in a recording medium and executed by a computer.

The above description of the present application is for illustrative purposes, and those skilled in the art will understand that it can be easily modified into other specific forms without changing the technical spirit or essential features of the present application. Therefore, the embodiments described above should be understood as illustrative in all respects and not limiting. For example, each component described as a single type may be implemented in a distributed manner, and similarly, components described as distributed may be implemented in a combined form.

The scope of the present application is indicated by the following claims rather than the detailed description above, and all changes or modifications derived from the meaning and scope of the claims and equivalent concepts thereof should be construed as being included in the scope of the present application.

100: Outlier detection and automatic thresholding system
110: database unit
120: data pre-processing unit
130: outlier detection unit
140: critical point calculation unit
150: outlier determination unit

Claims (9)

In the outlier detection and automatic threshold point system of unsupervised learning-based time series data,
A database unit for storing time-series data collected from open sources and factory sensors;
a data pre-processing unit extracting features for detecting outliers from the time series data and refining the data to enable discrimination;
an outlier detection unit learning the data preprocessed in the preprocessing unit in an outlier prediction model using an autoencoder structure in adversarial machine learning and detecting anomaly;
a threshold point calculation unit for deriving a critical point for finding an outlier using a machine learning algorithm based on the outlier prediction model; and
an outlier determining unit determining an outlier of the time-series data based on the critical point;
including,
The autoencoder is built with one encoder and a first autoencoder and a second autoencoder sharing the encoder for adversarial machine learning,
The first autoencoder restores input data, learns a model that deceives the second autoencoder, the second autoencoder restores the input data, and sets reconstruction data and input data restored by the first autoencoder in advance. Learning a model for distinguishing normal or abnormal state of the input data based on a loss value divergence point of one or more magnitudes,
The first autoencoder and the second autoencoder are given different weights of data reconstruction errors and outliers to be learned according to a preset ratio.
Outlier detection and automatic threshold system. According to claim 1,
The data pre-processing unit extracts statistical characteristic variables after pre-processing the time series data stored in the database unit for learning,
Outlier detection and automatic threshold system. delete delete According to claim 1,
The second autoencoder distinguishes between real data and reconstruction data of the first autoencoder, and the first autoencoder performs adversarial machine learning to deceive the second autoencoder so that the autoencoder detects an outlier similar to normal data. sign,
Outlier detection and automatic threshold system. delete According to claim 1,
The threshold point calculation unit calculates a threshold point for classifying normal data and abnormal data learned in the outlier detection unit based on a K-means algorithm, and defines the outlier based on the calculated threshold point.
Outlier detection and automatic threshold system. In the outlier detection and automatic threshold point system of unsupervised learning-based time series data,
Storing time-series data collected from sensors in open sources and factories;
a data pre-processing step of extracting features for detecting outliers from the time-series data and refining the data to enable discrimination;
learning the data preprocessed in the preprocessing step in an outlier prediction model using an autoencoder structure in adversarial machine learning and detecting the anomaly;
a threshold point calculation step of deriving a threshold point capable of finding an outlier by using a machine learning algorithm based on the outlier prediction model; and
determining an outlier of the time series data based on the critical point;
including,
The autoencoder is built with one encoder and a first autoencoder and a second autoencoder sharing the encoder for adversarial machine learning,
The first autoencoder restores input data, learns a model that deceives the second autoencoder, the second autoencoder restores the input data, and sets reconstruction data and input data restored by the first autoencoder in advance. Learning a model for distinguishing normal or abnormal state of the input data based on a loss value divergence point of one or more magnitudes,
The first autoencoder and the second autoencoder are given different weights of data reconstruction errors and outliers to be learned according to a preset ratio.
Outlier detection method. A computer-readable recording medium recording a program for executing the method of claim 8 on a computer.

Images