KR102113218B1 - A Unified Deep Learning Model for Time Series Data Prediction - Google Patents

Abstract

An integrated deep learning analysis method is disclosed. An integrated deep learning analysis method according to an embodiment may include receiving time series data; Converting the time series data into a frequency domain to generate frequency data; Generating features corresponding to the time series data by using a plurality of first neural networks receiving time series data and a plurality of second neural networks receiving frequency data; Combining features generated corresponding to at least one output category; And based on the combined features, outputting result data of at least one output category.

Description

Integrated deep learning system for analysis and prediction of time series data {A Unified Deep Learning Model for Time Series Data Prediction}

The embodiments described below are for an integrated deep learning system for analyzing and predicting time series data. Disclosed is a system that enables deep learning of various types of time series data that is most suitable for each data.

Deep learning, deep learning, is defined as a set of machine learning algorithms that attempt high-level abstractions through a combination of several nonlinear transformation methods. It can be a field of machine learning that teaches people.

It may be a machine learning technology built on an artificial neural network (ANN) to allow computers to learn on their own like a person using multiple data. Deep learning can train a machine to discern objects by imitating the way that the human brain discovers patterns in a lot of data and then sorts things out. When deep learning technology is applied, a computer may be able to recognize, reason, and judge by itself even if a person does not set all judgment criteria. It can be widely used for voice / image recognition and photo analysis.

The embodiments described below are intended to more accurately predict data by providing an integrated deep learning technology that accurately predicts various time series data.

In addition, we want to learn a separate deep learning model to learn smooth time series data, time series data with rapid change patterns, and time series data with periodic changes. It is intended to provide a technique for accurately classifying and predicting existing time series data with less learning data even when new learning data is input by preceding learning.

An integrated deep learning analysis method according to one side includes receiving time series data; Generating frequency data by converting the time series data into a frequency domain; Generating features corresponding to the time series data using a plurality of first neural networks receiving the time series data and a plurality of second neural networks receiving the frequency data; Combining the generated features corresponding to at least one output category; And based on the combined feature, outputting result data of the at least one output category.

The combining may include combining the generated features using an attention layer in which weight information for combining the generated features corresponding to the at least one output category is previously learned. Can be.

The at least one output category includes a first output category for facility diagnosis; A second output category for facility control; And a third output category for predicting change.

The time series data may include multiple time series data.

The first neural networks and the second neural networks may each include at least one of a convolutional neural network (CNN), a group convolutional neural network, and a cyclic neural network (RNN).

The first neural networks and the second neural networks may include neural networks of the same type.

The time series data may include variable information that changes over time.

The time series data may include learning time series data. The integrated deep learning analysis method may further include learning at least some of the first neural networks and the second neural networks based on a difference between the output result data and data tagged to the learning time series data. have. The integrated deep learning analysis method may further include learning the concentration layer based on a difference between the output result data and the data tagged in the learning time series data.

An integrated deep learning analysis apparatus according to one side includes a receiver configured to receive time series data; And converting the time series data into a frequency domain to generate frequency data. Generating features corresponding to the time series data using a plurality of first neural networks receiving the time series data and a plurality of second neural networks receiving the frequency data; Combining the generated features corresponding to at least one output category; And a control unit outputting result data of the at least one output category based on the combined features.

The integrated deep learning analysis apparatus may further include a storage unit that stores first parameters determined through prior learning of the first neural networks and second parameters determined through prior learning of the second neural networks.

The control unit may combine the generated features using an attention layer in which weight information for combining the generated features corresponding to the at least one output category is previously learned.

The integrated analysis deep learning system according to the embodiments may be accurately predictable through one integrated deep learning system even when characteristics of time series data are different. Using group convolutional neural networks, fast and accurate predictions can be made even for multi-time series data.

In addition, by learning the integrated deep learning system in advance even when the learning data is insufficient, it is possible to accurately classify and predict with less learning data, thereby reducing the burden on demand.

1 is a diagram illustrating a deep learning model for comprehensively analyzing time series data according to an embodiment
2 is a diagram illustrating an integrated deep learning device according to an embodiment.
3 is a diagram illustrating an integrated deep learning method using backpropagation according to an embodiment.
4 is a flowchart illustrating an integrated deep learning analysis method according to an embodiment.
5 is a diagram illustrating an integrated deep learning analysis apparatus according to an embodiment.

Hereinafter, embodiments will be described in detail with reference to the accompanying drawings. However, various changes may be made to the embodiments, and the scope of the patent application right is not limited or limited by these embodiments. It should be understood that all modifications, equivalents, or substitutes for the embodiments are included in the scope of rights.

The terms used in the examples are used for illustrative purposes only and should not be construed as limiting. Singular expressions include plural expressions unless the context clearly indicates otherwise. In this specification, the terms "include" or "have" are intended to indicate the presence of features, numbers, steps, actions, components, parts or combinations thereof described herein, one or more other features. It should be understood that the existence or addition possibilities of fields or numbers, steps, operations, components, parts or combinations thereof are not excluded in advance.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by a person skilled in the art to which the embodiment belongs. Terms such as those defined in a commonly used dictionary should be interpreted as having meanings consistent with meanings in the context of related technologies, and should not be interpreted as ideal or excessively formal meanings unless explicitly defined in the present application. Does not.

In addition, in the description with reference to the accompanying drawings, the same reference numerals are assigned to the same components regardless of reference numerals, and redundant descriptions thereof will be omitted. In describing the embodiments, when it is determined that detailed descriptions of related known technologies may unnecessarily obscure the subject matter of the embodiments, detailed descriptions thereof will be omitted.

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

Deep learning, deep learning, is defined as a set of machine learning algorithms that attempt high-level abstractions through a combination of several nonlinear transformation methods. It can be a field of machine learning that teaches people.

It may be a machine learning technology built on an artificial neural network (ANN) to allow computers to learn on their own like a person using multiple data. Deep learning can train a machine to discern objects by imitating the way that the human brain discovers patterns in a lot of data and then sorts things out. When deep learning technology is applied, a computer may be able to recognize, reason, and judge by itself even if a person does not set all judgment criteria. It can be widely used for voice / image recognition and photo analysis.

Machine learning can learn by accumulating in the database after the machine analyzes a large amount of sample data, repeats the training data, extracts rules, regulations, expressions, conditions, and judgment criteria by itself. When data that requires recognition and analysis is entered, the machine can analyze and identify the data based on the accumulated database, think and predict relationships, and replace the 'thinking' work that humans normally do.

Neural networks are a type of machine learning that can be a technique that mimics the human brain. The neural network can extract the 'feature value' of an object and learn in a similar way to a human. For example, humans have learned and possessed 'features' to determine whether the human brain is an individual or a cat through experiences so far, and based on it, look at several images and classify them into 'dogs' and 'cats'. can do.

In a neural network, it is possible to calculate a feature value rather than a human-provided rule to identify a subject. If the feature value is taught as 'dog', it can be classified as a dog. By repeating, the machine itself can calculate the feature value and increase the information to be classified into dogs. The feature values handled by the computer may be numerical values in the form of vectors. The features are automatically learned by the machine through training, so the developer can be freed from the task of defining detailed rules as before. When an image is input, the computer analyzes the feature values of the image and determines that the dog is a dog if it matches the range of the dog's feature values. It is not a human-made feature value that determines a dog, but a machine can calculate it through learning.

The neural network may include an input layer, an output layer, and an intermediate layer. The middle layer can be said to be a hidden layer. The presence of an intermediate layer can increase the number of layers of neurons responsible for processing. Having a multilayer neuron layer in the middle layer can be referred to as a deep neural network. Machine learning in deep neural networks can be called deep learning. The middle layer can be further layered to train and process more data. If the middle layer is excessively multi-layered, the number of parameters becomes too large, the neurons are large and complicated, and the unrelated bonds may increase, resulting in over-fitting. Performance may degrade due to overfitting. A convolutional neural network (CNN) can solve the problem.

Convolutional neural networks are a type of multi-layer perceptrons designed to use minimal preprocessing. CNN consists of one or several convolutional layers and common artificial neural network layers on top of it, and additional weights and pooling layers can be utilized. CNN can fully utilize the input data of a two-dimensional structure. Compared to other deep learning structures, CNN can perform well in both video and audio fields. CNNs can also be trained through standard reverse delivery. CNN is more easily trained than other feedforward artificial neural network techniques and has the advantage of using fewer parameters.

The convolutional neural network according to an embodiment may analyze a partial region of an image, move the window of the region as if sliding, and then repeat to the next region and the next region. Instead of propagating the information of the identified image region to all neurons, it is possible to propagate only to neurons having a high relationship. This can solve the overfitting problem.

A Recurrent Neural Network (RNN), which has high performance for time series data analysis, may include an input layer, an intermediate layer, and an output layer, and the value of the current intermediate layer may influence the calculation of the next output layer. Recurrent neural networks may be a type of artificial neural network in which a node in the middle layer is connected to an edge with a direction to form a cyclic structure (directed cycle). The core of the recurrent neural network may be to find optimal parameters between nodes and layers. Recurrent neural networks can be extended to bidirectional recurrent neural networks. In bidirectional recurrent neural networks, when modeling sequence data, the current sequence data can be influenced by the previous data as well as the previous data. Unlike a unidirectional recurrent neural network that references only the previous data and the current data, a bidirectional recurrent neural network can refer to both before and after data and current data.

Time series data may be data in which observation values are recorded at regular intervals with the passage of time. Time series data analysis techniques can be used in many fields such as finance, military and manufacturing. Accurately analyzing time series data can be very important for financial forecasting, manufacturing diagnostics, and more. For example, accurately predicting a signal input from a large-scale system of a power plant can predict a failure and accurately calculate a maintenance cycle, thereby reducing the risk of an accident and significantly reducing the cost of maintenance. Multiple time series data may be records of multiple variables that change over time. For example, multiple time series data may include changes in multiple stock prices over time and multiple exchange rates over time. Time series data can be persistent. For example, people can understand each word based on previous words, and previous video frames may affect understanding the current frame. Unlike static data, time series data can have several characteristics. For example, there may be time series data that changes relatively smoothly over time, and time series data that changes rapidly (high frequency). In addition, there may be time series data repeating a specific pattern.

Due to the development of deep learning technology, not only static data analysis, but also time series data can be analyzed and predicted. There may be a limitation that a large amount of training data is required to obtain reliable analysis and prediction results through deep learning of time series data. In reality, it can be difficult or time consuming to obtain a lot of data. A suitable deep learning algorithm may be different for each time series data. If time-series data is input to an inappropriate deep learning algorithm, it may be difficult to obtain a desired result. A suitable deep learning algorithm is different for each characteristic of time series data, and there may be a problem that training data is often insufficient. Even if there is little training data, an integrated deep learning model that can output reliable results regardless of the characteristics of time series data may be required.

The integrated deep learning model according to an embodiment may perform prior learning using various training time series data (or learning time series data). In addition to time series data, frequency data obtained by converting time series data into a frequency domain can be trained. Significant results can also be obtained in terms of frequency of time series data. As time-series data for training, published time-series data can be used. If you put a lot of data and do prior learning, you can get a feature map with meaningful parameter values.

According to an embodiment, after a meaningful feature map is created, new time series data to be analyzed and predicted may be additionally trained. Since learning is done after a meaningful feature map has already been completed, reliable results can be obtained even with a small amount of data.

The integrated deep learning model includes multiple deep learning algorithms. Appropriate deep learning algorithms can be applied regardless of the type of time series data.

1 is a diagram illustrating a deep learning model for comprehensively analyzing time series data according to an embodiment.

Referring to FIG. 1, a deep learning model for comprehensively analyzing time series data according to an embodiment may include an integrated deep learning layer 130, a fusion layer 140, and an output layer 150. The time series data 110 and the frequency data 120 obtained by frequency-converting the time series data may be input to the integrated deep learning layer 130. The integrated deep learning layer 130 may include a plurality of neural networks to which a plurality of deep learning algorithms are applied.

When the input data passes through the integrated deep learning layer 130, a plurality of features corresponding to time series data may be generated. The generated features may be input to the fusion layer 140. The fusion layer 140 may combine features corresponding to at least one output category. The output category may include a first output category for facility diagnosis, a second output category for facility control, a third output category for change prediction, and various combinations thereof.

The integrated deep learning analysis model can output results for a specific service. For example, the quantitative output value of the deep learning algorithm can be output as a result of specific analysis and prediction of water level prediction, speech recognition, stock price prediction, facility diagnosis of failure, and failure prediction.

As an example, the fusion layer 140 learns so that a neural network with a high contribution has a higher weight than a network with a low contribution in deriving output information of a specific category among various neural networks included in the integrated deep learning layer 130. Can be. The fusion layer 140 may be referred to as an attention layer in terms of determining which of the plurality of neural networks included in the integrated deep learning layer 130 should pay higher attention to. The operation of the learning phase for learning the fusion layer 140 and the integrated deep learning layer 130 will be described later with reference to FIG. 3.

Neural networks included in the integrated deep learning layer may include a convolutional neural network (CNN), a group convolutional neural network, a cyclic neural network, and various combinations thereof.

Unlike static data, time series data can have several characteristics. For example, there may be time series data that changes relatively smoothly over time, and time series data that changes rapidly (high frequency). For example, smoothly changing time series data can be international oil prices or stock prices over time. Time-series data that changes rapidly may have a vibrating voice. In addition, there may be time series data repeating a specific pattern. For example, EEG data may repeat certain patterns. Depending on the above characteristics, a deep learning algorithm suitable for each time series data may be different.

For example, a cyclic neural network including a long and short-term memory circulating neural network may be suitably used as time series data that changes relatively smoothly over time. Time-series data that is rapidly changing can be expressed in an image format by converting it to the frequency domain, so a convolutional neural network or a group convolutional neural network may be suitable. Time series data repeating a positive pattern may be a convolutional neural network (CNN) or a group convolutional neural network.

In the case of a model composed of one type of deep learning algorithm, only time series data of one characteristic can be learned. When using time series data with different characteristics, reliable results may not be obtained. For example, in the case of a model in which only a convolutional neural network exists, a desired result may be obtained in the analysis of EEG data, but a desired result may not be obtained in an international oil price analysis and prediction over time.

When an integrated deep learning model is used in one embodiment, learning may be performed through a deep learning algorithm suitable for each time series data. In particular, by converting time series data into the frequency domain, it can be analyzed and predicted through learning in the frequency domain as well as the time domain.

2 is a diagram illustrating an integrated deep learning device according to an embodiment.

Referring to FIG. 2, an integrated deep learning apparatus according to an embodiment may include RNN1 231, RNN2 232, CNN1 233, and CNN2 234. The RNN1 231, RNN2 232, CNN1 233, and CNN2 234 may correspond to the integrated deep learning layer 130 of FIG. 1.

Time series data 210 may be input to RNN1 231 and CNN1 233. The frequency data 220 obtained by frequency-converting time series data may be input to the RNN2 232 and the CNN2 234. Through the structure according to the embodiment of FIG. 2, the integrated deep learning model can train frequency data 220 as well as time series data 210 to accurately analyze various data.

The integrated deep learning device includes FCL 240. The FCL 240 may correspond to the fusion layer 140 of FIG. 1. In addition, the integrated deep learning device may include a first output layer 251 for facility diagnosis, a second output layer 252 for facility control, and a third output layer 253 for predicting change. The first output layer 251, the second output layer 252 for facility control, and the third output layer 253 for change prediction may correspond to the output layer 150 of FIG. 1.

3 is a diagram illustrating an integrated deep learning method using backpropagation according to an embodiment.

Referring to FIG. 3, the integrated deep learning model includes an integrated lip learning layer 330, a fusion layer 340, and an output layer 350. The integrated lip learning layer 330, the fusion layer 340, and the output layer 350 correspond to the integrated lip learning layer 130, the fusion layer 140, and the output layer 150 of FIG. 1 as they are. Detailed description is omitted.

In the learning phase, the time series data 310 may be received with the label 370 tagged. The label 370 may be used as a ground truth directed by the result data output from the output layer 350. For example, result data may be output from the time series data 310 and the frequency data 320 based on the parameters of the integrated deep learning layer 330 and the fusion layer 340 in the current state. The parameter may include weights of synapses connecting nodes included in the neural network.

Loss 360 may be generated by comparing the output result data with the label 370. Based on the generated loss, the fusion layer 340 and the integrated deep learning layer 330 may be learned through a backpropagation learning technique. In some cases, the output layer 350 can also be learned together. Through learning, the weight of synapses connecting between nodes included in the neural network may be updated.

4 is a flowchart illustrating an integrated deep learning analysis method according to an embodiment.

Referring to FIG. 4, an integrated deep learning analysis method according to an embodiment includes pre-learning and analysis and prediction through time-series data learning to actually use using the pre-learned results. The time series data may include at least one of water level, voice, sensor, sales, stock, and interest rate according to the passage of time. The time series data may include multiple time series data.

In order to obtain reliable results through deep learning, it is necessary to learn using a large amount of data. Learning a large amount of data can take a long time. Everyone may not have enough data and may want to get results in a shorter time. If you can create a meaningful feature map by pre-training a deep learning algorithm, you can get reliable results in a shorter time with a small amount of data. Before learning the time series data to be actually used, prior learning can be performed using training data. The training data may be public data that has already been made publicly available.

The pre-training step includes receiving learning time series data (400), generating learning frequency data using learning time series data (410), and generating a feature map using learning time series data and learning frequency data. 420. The learning frequency data can be obtained by Fourier transforming the learning time series data.

Analysis and prediction through learning new time series data to be actually used by using the previously learned results include receiving time series data (430), and generating frequency data using the received time series data (440), And outputting features 450 using the learned feature map data, time series data, and frequency data as inputs of the integrated deep learning device.

The integrated deep learning analysis method according to an embodiment may further include a step 460 of receiving features as input and outputting combined features corresponding to the time series data characteristics based on a predetermined weight. Features that are a plurality of output types that have passed a plurality of deep learning algorithms may be output as combined features that are one result type based on a predetermined weight. The weight for generating the combined features may be determined to correspond to the input time series data characteristics. For example, when stock price data according to time is input, a plurality of features may be output through a plurality of deep learning algorithms including a convolutional neural network and a cyclic neural network. Among the plurality of features, a large weight is given to features that have passed through a cyclic neural network corresponding to a stock price data characteristic over time, and a small weight is applied to a convolutional neural network that is not, to generate a calculated combined feature. have.

5 is a diagram illustrating an integrated deep learning analysis apparatus according to an embodiment.

Referring to FIG. 5, the integrated deep learning analysis apparatus includes a reception unit 510, a frequency changer 520, a deep learning algorithm storage 530, and a feature map storage 540.

The receiver 510 may receive time series data and published time series data. First, the time series data that has been released can be received for prior learning, and after the prior learning, time series data that is actually used can be received.

The frequency converter 510 may convert time series data and published time series data into a frequency domain to generate frequency data and published frequency data. The frequency converter 520 may convert time series data into a frequency domain through Fourier transform.

The deep learning algorithm storage 530 may store an integrated deep learning device including a plurality of deep learning algorithms. Through deep learning algorithms, it is possible to pre-learn published time series data and published frequency data. Time series data and frequency data to be actually used can be input after the preceding nuclear attack.

The feature map storage 540 may store learned feature maps obtained through repetitive operations of inputting published time series data and published frequency data into an integrated deep learning device. The integrated deep learning device may output features using the learned feature map data, time series data, and frequency data as inputs.

The integrated deep learning analysis apparatus according to an embodiment may further include a fusion layer storage that receives features as input and outputs combined features corresponding to the time series data characteristics based on a predetermined weight. An output deep learning algorithm store that receives a combined feature as an input and outputs a result including at least one of diagnosis, facility control, and change prediction using a predetermined deep learning algorithm may be further included. The specific service may include at least one of diagnosis, facility control, and change prediction. For example, the quantitative output value of the deep learning algorithm can be output as a result of specific analysis and prediction of water level prediction, speech recognition, stock price prediction, facility diagnosis of failure, and failure prediction.

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, or the like alone or in combination. The program instructions recorded on the medium may be specially designed and configured for the embodiments or may be known and usable by 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, DVDs, and magnetic media such as floptical disks. -Hardware devices specifically 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 code that can be executed by a computer using an interpreter, etc., as well as machine language codes produced by a compiler. The hardware device described above may be configured to operate as one or more software modules to perform the operations of the embodiments, and vice versa.

The software may include a computer program, code, instruction, or a combination of one or more of these, and configure the processing device to operate as desired, or process independently or collectively You can command the device. Software and / or data may be interpreted by a processing device, or to provide instructions or data to a processing device, of any type of machine, component, physical device, virtual equipment, computer storage medium or device. , Or may be permanently or temporarily embodied in the transmitted signal wave. The software may be distributed over networked computer systems, and stored or executed in a distributed manner. Software and data may be stored in one or more computer-readable recording media.

As described above, although the embodiments have been described with limited drawings, those skilled in the art can apply various technical modifications and variations based on the above. For example, the described techniques are performed in a different order than the described method, and / or the components of the described system, structure, device, circuit, etc. are combined or combined in a different form from the described method, or other components Alternatively, even if substituted or substituted by equivalents, appropriate results can be achieved.

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

Claims (18)

In the integrated deep learning analysis method performed by the integrated deep learning analysis device implemented by at least one processor,
Receiving time series data;
Generating frequency data by converting the time series data into a frequency domain;
Generating a plurality of features corresponding to the time series data using a plurality of first neural networks receiving the time series data and a plurality of second neural networks receiving the frequency data;
Combining the generated features using an attention layer in which weight information for predicting the contribution of neural networks included in the plurality of first neural networks and the plurality of second neural networks is learned; And
Based on the combined feature, outputting result data of at least one output category
Including,
The first neural networks and the second neural networks
Includes a convolutional neural network (CNN) and a circulatory neural network (RNN),
The contribution of the neural networks
An integrated deep learning analysis method that varies according to the degree of change of the time series data over time.
delete According to claim 1,
The at least one output category
A first output category for facility diagnostics;
A second output category for facility control; And
Third output category for predicting change
Integrated deep learning analysis method comprising at least one of the.
According to claim 1,
The time series data includes multiple time series data,
Integrated deep learning analysis method.
delete delete According to claim 1,
The time series data,
Including variable information that changes over time,
Integrated deep learning analysis method.
According to claim 1,
The time series data includes learning time series data,
The integrated deep learning analysis method
Learning at least some of the first neural networks and the second neural networks based on the difference between the output result data and data tagged to the learning time series data.
Further comprising, integrated deep learning analysis method.
According to claim 1,
The time series data includes learning time series data,
The integrated deep learning analysis method
Based on a difference between the output result data and the data tagged in the learning time series data, learning the attention concentration layer
Further comprising, integrated deep learning analysis method.
A computer program stored in a medium for executing the method of any one of claims 1, 3 to 4 and 7 to 9 in combination with hardware.
A receiving unit for receiving time series data; And
Converting the time series data into a frequency domain to generate frequency data, and using a plurality of first neural networks receiving the time series data and a plurality of second neural networks receiving the frequency data, a plurality corresponding to the time series data Generating the features, and using the attention layer in which weight information for predicting the contribution of neural networks included in the plurality of first neural networks and the plurality of second neural networks is learned, the generated features A control unit that combines them and outputs result data of at least one output category based on the combined features
Including,
The first neural networks and the second neural networks
Includes a convolutional neural network (CNN) and a circulatory neural network (RNN),
The contribution of the neural networks
An integrated deep learning analysis device that varies according to the degree of change of the time series data over time.
The method of claim 11,
A storage unit that stores first parameters of the first neural networks previously learned and second parameters of the second neural networks previously learned
Further comprising, an integrated deep learning analysis device.
delete The method of claim 11,
The at least one output category
A first output category for facility diagnostics;
A second output category for facility control; And
Third output category for predicting change
Integrated deep learning analysis device comprising at least one of.
The method of claim 11,
The time series data includes multiple time series data,
Integrated deep learning analysis device.
delete delete The method of claim 11,
The time series data,
Including variable information that changes over time,
Integrated deep learning analysis device.

Images