KR102377474B1 - Lost data recovery method and apparatus using parameter transfer lstm - Google Patents

Abstract

Disclosed are a method and apparatus for recovering lost data using a parameter transfer long short-term memory (LSTM) layer. The method for recovering loss data can comprise: a step of receiving N units of normal-directional data sets and N units of reverse-directional data sets corresponding to a previous space and a subsequent space for the loss data of a specific space; a step of learning the individual LSTM layers placed in a space order by using the N units of normal-directional data sets and the N units of reverse-directional data sets, and generating a first training LSTM layer for the N units of normal-directional data sets and a second training LSTM layer for the N units of reverse-directional data sets; a step of respectively extracting normal-directional prediction data and reverse-directional prediction data for the specific space by using the generated first training LSTM layer and the second training LSTM layer; and a step of recovering the lost data on the specific space based on the correlation level of the normal-directional prediction data and the reverse-directional prediction data. The present invention aims to provide a method and apparatus for recovering loss data using a parameter transfer LSTM to reduce the calculation volume required for learning the whole data.

Description

LOST DATA RECOVERY METHOD AND APPARATUS USING PARAMETER TRANSFER LSTM

The present invention relates to a method and apparatus for recovering lost data using parameter transfer LSTM (Long Short-Term Memory), and more particularly, to a cloud server, a local server, a peripheral base station (Road Side Unit, RSU) and a vehicle terminal (On- When data collection is impossible due to a mechanical failure of an RSU in a V2I (Vehicle to Infrastructure) network environment composed of a Board Unit, OBU), it relates to a method and apparatus for recovering spatiotemporal data for the corresponding RSU section.

Recently, intelligent transport systems (ITS) are attracting attention because of the innovative development of smart cities and effective management of transport systems. ITS technology can actively collect a variety of traffic information from a variety of sources, allowing service providers to provide data-driven services such as vehicle and road maintenance, long-term and short-term traffic management, road safety, and convenience services. became the basis For service providers, the most important goal is to provide smooth service to their vehicles even in the event of unexpected network malfunctions. To this end, replacement of missing data due to such network malfunction or network failure can play an essential role in providing smooth data-based services.

However, it may be difficult to recover missing data from a vehicle to an infrastructure (V2I) network due to traffic with dynamic, complex, and stochastic characteristics. Since the traffic data collected from the RSU changes with time and location (ie, space-time domain), data recovery performance is highly dependent on changes in the environment, so it is difficult to use a predetermined model or algorithm.

Moreover, considering the efficiency of data recovery, the problem can be made more difficult. For example, if large amounts of data need to be processed, the time required to recover missing data may be long and the recovered data may not be available when needed. Therefore, there is a need for an efficient algorithm capable of reducing the time required for data recovery without degrading data recovery performance.

The present invention provides a method and apparatus for recovering lost data that reduces the amount of computation required to learn the entire data by transferring the parameters of the learned LSTM layer to the next layer and using the initial parameter values as parameters of the previously learned data.

A method for recovering lost data according to an embodiment of the present invention includes: receiving N forward data sets and N backward data sets corresponding to a space before and after space for lost data in a specific space; A first training LSTM layer for the N forward data sets and the N by learning individual Long Short-Term Memory (LSTM) layers arranged in spatial order using the N forward data sets and the N backward data sets. generating a second training LSTM layer for the backward data sets; extracting forward prediction data and backward prediction data for the specific space using the generated first training LSTM layer and the second training LSTM layer, respectively; and recovering the lost data for the specific space based on the correlation between the forward prediction data and the backward prediction data.

The generating may include: learning a first LSTM layer over time using a first forward data set among the N forward data sets composed of time series data; and sequentially learning the second LSTM layers corresponding to each of the forward data sets by applying the parameters learned through the previous LSTM layer to each of the forward data sets after the first forward data set. generating a training LSTM layer.

The generating may include: learning a third LSTM layer over time using a first backward data set among the N backward data sets composed of time series data; and by sequentially learning the fourth LSTM layers corresponding to each of the reverse data sets by applying the parameters learned through the previous LSTM layer to each of the backward data sets after the first backward data set, the second for the reverse data sets generating a training LSTM layer.

The extracting step is performed by inputting a forward data set corresponding to the immediately preceding space of the specific space and a backward data set corresponding to the immediately immediate space to the generated first training LSTM layer and the second training LSTM layer, respectively, for the specific space. Forward prediction data and backward prediction data may be extracted.

The restoring may include: calculating a weight using a first correlation coefficient for the N forward data sets and a second correlation coefficient for the N backward data sets; and recovering lost data for the specific space by applying the calculated weight to the forward prediction data and the backward prediction data.

The weight may be calculated through a ratio of the first correlation coefficient and the second correlation coefficient.

In the generating, each parameter of the individual LSTM layers may be updated using a loss function corresponding to a mean squared error (MSE).

Loss data recovery apparatus according to an embodiment of the present invention includes a processor, wherein the processor receives N forward data sets and N backward data sets corresponding to spaces before and after loss data in a specific space, and , a first training LSTM layer for the N forward data sets by learning individual Long Short-Term Memory (LSTM) layers arranged in spatial order using the N forward data sets and the N backward data sets; and Generate a second training LSTM layer for N backward data sets, and extract forward prediction data and backward prediction data for the specific space using the generated first training LSTM layer and second training LSTM layer, respectively, Loss data for the specific space may be recovered from a missing value determined based on the correlation between the forward prediction data and the backward prediction data.

The processor learns the first LSTM layer over time by using the first forward data set among the N forward data sets composed of time series data, and for each of the forward data sets after the first forward data set through the previous LSTM layer The first training LSTM layer for the forward data sets may be generated by sequentially learning the second LSTM layers corresponding to each of the corresponding forward data sets by applying the learned parameter.

The processor learns a third LSTM layer over time using a first backward data set among the N backward data sets composed of time series data, and for each of the backward data sets after the first backward data set, through the previous LSTM layer A second training LSTM layer for the backward data sets may be generated by sequentially learning the fourth LSTM layers corresponding to each of the corresponding backward data sets by applying the learned parameter.

The processor is configured to input a forward data set corresponding to the immediately preceding space of the specific space and a backward data set corresponding to the immediately preceding space to the generated first training LSTM layer and the second training LSTM layer, respectively, to make forward prediction for the specific space. Data and backward prediction data can be extracted.

The processor calculates a weight by using a first correlation coefficient with respect to the N forward data sets and a second correlation coefficient with respect to the N backward data sets, and applies the calculated weight to the forward prediction data and the backward prediction data. By applying to , it is possible to recover lost data for the specific space.

The weight may be calculated through a ratio of the first correlation coefficient and the second correlation coefficient.

The processor may update a parameter of each of the individual LSTM layers using a loss function corresponding to a mean squared error (MSE).

According to the present invention, the amount of computation required to learn the entire data can be reduced by transferring the parameters of the learned LSTM layer to the next layer and using the initial parameter values as parameters of the previously learned data.

1 is a diagram illustrating the structure of a V2I network according to an embodiment of the present invention.
2 is a diagram illustrating an example of an RSU failure scenario and data recovery according to an embodiment of the present invention.
3 is a diagram illustrating an overall diagram of a lost data recovery algorithm using a parameter transition LSTM network according to an embodiment of the present invention.
4 is a diagram illustrating a method of calculating forward prediction data according to an embodiment of the present invention.
5 is a diagram illustrating a linear interpolation method according to an embodiment of the present invention.

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

1 is a diagram illustrating the structure of a V2I network according to an embodiment of the present invention.

Referring to FIG. 1 , the V2I network 100 includes a cloud server 110 , a local server 120 , a peripheral base station (Road Side Unit, RSU) 130 , and an On-Board Unit (OBU) 140 ). can be composed of First, the RSU 130 may upload the traffic data collected from the OBU 140 to the cloud server 110 through the local server 120 . Similarly, the cloud server 110 may provide a V2I service to the OBU 140 using the traffic data uploaded from the local server 120 and the RSU 130 . It is assumed that the V2I network 100 of the present invention has a limited storage space of the local server 120 to store some of the recently observed traffic data by the OBU 140 .

로 표시된 RSU(K는 RSU의 인덱스 세트)(131)에는 인접한 RSU들(

)이 존재할 수 있다. RSU(131)는 특정 방향으로 이동하는 OBU(140)들의 교통 데이터를 수집할 수 있는데, 이때 수집된 교통 데이터는

로 표시될 수 있으며, OBU(140)들의 평균 교통 속도를 나타낼 수 있다. 이때,

는 시간 t에서의 속도 데이터이고, L은 데이터 길이를 나타낸다. 인접한 RSU들은 고정된 거리에 배치되기 때문에 RSU(131)를 포함하여 인접한 RSU들을 통해 수집된 교통 데이터들(

)은 시간 및 공간적으로 상관 관계가 있을 수 있다.Referring to FIG. 1 , each OBU 140 may be dynamically connected to an RSU 130 when moving, and each RSU 130 may be connected to a local server 120 .

RSU (K is the index set of the RSU) 131 indicated by the adjacent RSUs (

) may exist. The RSU 131 may collect traffic data of the OBUs 140 moving in a specific direction. At this time, the collected traffic data is

It may be expressed as , and may represent the average traffic speed of the OBUs 140 . At this time,

is the velocity data at time t, and L denotes the data length. Since the adjacent RSUs are arranged at a fixed distance, traffic data collected through the adjacent RSUs including the RSU 131 (

) can be temporally and spatially correlated.

2 is a diagram illustrating an example of an RSU failure scenario and data recovery according to an embodiment of the present invention.

Referring to (a) of FIG. 2 , the RSU 131 temporarily fails, and the local server 120 provides the traffic data of the OBUs 140 collected by the failed RSU 131 to the cloud server 110 . You can consider scenarios where you can't. Traffic data missing from the RSU 131 may be recovered through the lost data recovery method provided by the present invention. This lost data recovery method may be performed by the local server 120 .

를 장애가 발생한 RSU(131)의 손실된 교통 데이터라고 가정하자. RSU(130)들 각각에서 수집된 교통 데이터는 시간적, 공간적으로 상관 관계가 있으므로, RSU(131)에서 손실된 교통 데이터

는 인접한 RSU들의 교통 데이터를 사용하여 복구될 수 있다.first,

Assume that is lost traffic data of the faulty RSU 131 . Since the traffic data collected from each of the RSUs 130 are temporally and spatially correlated, the traffic data lost from the RSU 131 are

can be recovered using traffic data of adjacent RSUs.

를 이용하여 RSU(131)의 교통 데이터

를 복구할 수 있다. To this end, the local server 120 collects traffic data from 2N adjacent RSUs located before and after the failed RSU 131 as shown in FIG. 2(b).

traffic data of RSU (131) using

can be restored

및

각각은 일반적인 LSTM 네트워크를 이용한 손실 데이터 복구 알고리즘 및 본 발명에서 제공하는 파라미터 전이 LSTM 네트워크를 이용한 손실 데이터 복구 알고리즘을 나타낸다. 해당 손실 데이터 복구 알고리즘들을 통해 복구된 교통 데이터

및

는 아래의 식 1과 같이 표현될 수 있다. function

and

Each represents a lost data recovery algorithm using a general LSTM network and a lost data recovery algorithm using a parameter transition LSTM network provided in the present invention. Traffic data recovered through the corresponding lost data recovery algorithms

and

can be expressed as Equation 1 below.

<Equation 1>

와 최대 성능 편차 파라미터

가 주어지면 본 발명의 파라미터 전이 LSTM 네트워크를 이용한 손실 데이터 복구 알고리즘은 아래의 식 2와 같은 제약 조건을 이용함으로써 복구된 교통 데이터

를 계산하는 데 필요한 훈련 및 복구 시간 T를 줄일 수 있다.The loss data recovery algorithm using the parameter transfer LSTM network provided in the present invention can provide a method of reducing the learning and replacement time for the loss data recovery with less performance deviation compared to the loss data recovery algorithm using the general LSTM network. At this time, the performance measurement parameter

and maximum performance deviation parameters

Given that , the lost data recovery algorithm using the parameter transfer LSTM network of the present invention recovers traffic data by using the constraint as in Equation 2 below.

can reduce the training and recovery time T required to compute

<Equation 2>

로 아래의 식 3과 같이 평균 제곱근 편자(Root Mean Square Error, RMSE)를 사용할 수 있다. In the present invention, the performance measurement parameter

As shown in Equation 3 below, Root Mean Square Error (RMSE) can be used.

<Equation 3>

Therefore, the local server 120 of the present invention can provide the traffic data of the RSU 131 section in which the failure occurs to the cloud server 110 more stably and efficiently by using a loss data recovery algorithm using a parameter transfer LSTM network.

3 is a diagram illustrating an overall diagram of a lost data recovery algorithm using a parameter transition LSTM network according to an embodiment of the present invention.

를 복구하는 방법을 제공할 수 있다. Referring to FIG. 3 , the lost data recovery algorithm using the parameter transfer LSTM network of the present invention uses 2N data sets received from RSUs adjacent to the failed RSU 131 to correspond to the RSU 131 and the missing data. three

can provide a way to restore it.

First, the local server 120 transmits traffic data collected from N adjacent RSUs positioned before the RSU 131 among traffic data collected from 2N adjacent RSUs positioned before and after the failed RSU 131, that is, the forward direction. Forward prediction data corresponding to the RSU 131 may be calculated by learning the parameter transfer LSTM network using the data sets.

Then, the local server 120 transmits traffic data collected from N adjacent RSUs positioned after the RSU 131 among traffic data collected from 2N adjacent RSUs positioned before and after the failed RSU 131 , that is, in the reverse direction. Backward prediction data corresponding to the RSU 131 may be calculated by learning the parameter shift LSTM network using the data sets.

를 복구할 수 있다.Thereafter, the local server 120 performs linear interpolation using the calculated forward prediction data and backward prediction data to thereby perform traffic data corresponding to the RSU 131 .

can be restored

A more detailed prediction data calculation method and a lost data recovery method will be described in detail with reference to the following drawings.

4 is a diagram illustrating a method of calculating forward prediction data according to an embodiment of the present invention.

The loss data recovery algorithm using the parameter transition LSTM network provided in the present invention is performed based on N consecutive LSTM layers, and parameters learned in each LSTM layer may be transferred to the next LSTM layer. In addition, each LSTM layer may predict traffic data for a subsequent RSU in the spatial domain.

를 계산하기 위하여 입력 데이터로

를 사용할 수 있다. 이때,

,

,

,

및

를 각각 시간 t에서 잊혀진 게이트, 입력 게이트, 셀 상태, 출력 게이트, 순방향 예측 데이터를 계산하기 위한 l 번째 LSTM 레이어의 은닉 상태의 변수로 가정하자. 이때, 도 6을 참고하면, 잊혀진 게이트(

)는 이전 단계의 셀 상태(Sell state) 값을 얼마나 이용할지 또는 버릴지를 결정하는 파라미터이고, 입력 게이트(

)는 입력 값

에서 tanh를 태워 나온 새로운 특징(Feature)를 얼마나 반영할지를 결정하는 파라미터이다. 그리고, 셀 상태(

)는 기억을 오랫동안 유지하기 위하여 사용되는 파라미터로 기존 특징에 새로운 특징을 계속 추가하는 방식으로 사용될 수 있다. 출력 게이트(

)는 현재 단계의 출력을 이후 단계로 보낼 때 얼마나 반영할지를 결정하는 파라미터이고, 은닉 상태(

)는 현재 단계의 출력이고, 다음 시간 단계로 넘기는 정보로서 다음 단계로 보낼 때 얼마나 반영할지를 결정하는 파라미터이다.As an example, the local server 120 provides forward prediction data in the l -th LSTM layer.

as input data to calculate

can be used At this time,

,

,

,

and

Let be the variables of the hidden state of the l -th LSTM layer for calculating the forgotten gate, input gate, cell state, output gate, and forward prediction data at time t, respectively. At this time, referring to FIG. 6, the forgotten gate (

) is a parameter that determines how much to use or discard the cell state value of the previous step, and the input gate (

) is the input value

It is a parameter that determines how much of a new feature from burning tanh will be reflected. And, the cell state (

) is a parameter used to keep the memory for a long time, and it can be used as a way to continuously add new features to existing features. output gate (

) is a parameter that determines how much of the output of the current stage is reflected when sending it to the next stage, and the hidden state (

) is the output of the current step, and as information passed to the next time step, it is a parameter that determines how much to reflect when sending to the next step.

This can be expressed as Equation 4 below.

<Equation 4>

는 아다마르 곱(Hadamard product),

,

,

및

는 게이트의 가중치 행렬,

,

및

는 바이어스 벡터 파라미터,

는 시그모이드(Sigmoid) 함수를 나타낸다. 이때, 각 LSTM 레이어에서 평균 제곱 오차(MSE)는 손실 함수로 사용되며 다음의 식 5와 같이 표현될 수 있다.here

is the Hadamard product,

,

,

and

is the weight matrix of the gate,

,

and

is the bias vector parameter,

denotes a sigmoid function. In this case, the mean square error (MSE) in each LSTM layer is used as a loss function and can be expressed as Equation 5 below.

<Equation 5>

는 시간 t에서 l 번째 LSTM 레이어의 순방향 예측 데이터를 나타낸다. 이와 같은 LSTM 레이어의 학습 절차는 l 번째 역방향 LSTM 레이어에서 입력 데이터

를 이용하여 역방향 예측 데이터

을 계산하는 방법과 동일할 수 있다.here

denotes forward prediction data of the l -th LSTM layer at time t. The training procedure of such an LSTM layer is based on the input data in the l -th reverse LSTM layer.

Reverse prediction data using

may be the same as the method of calculating .

And, as for the parameters of the LSTM network learned through the l -th LSTM layer, the weight matrix, cell state, and hidden state may be transferred to the next LSTM layer as shown in Equation 6 below.

<Equation 6>

In this way, the parameters learned in the l -th LSTM layer may be transferred to the l +1 LSTM layer and used as initial parameters for learning. This process may be repeated up to the N-1 th LSTM layer as shown in FIG. 4 , and forward prediction data of the RSU 131 in which the failure occurs may be calculated using the last learned N th LSTM layer.

As described above, since the loss data recovery algorithm using the parameter transfer LSTM network of the present invention can efficiently initialize the parameters of the LSTM layer, it is possible to greatly reduce the computational complexity by omitting repetition of characterizing general patterns of spatiotemporal traffic data.

5 is a diagram illustrating a linear interpolation method according to an embodiment of the present invention.

및 역방향 예측 데이터

가 주어지면, 로컬 서버(120)는 효율적인 손실 데이터 복구를 위한 공간 상관을 이용하여 두 예측 데이터

및

간의 선형 보간을 수행할 수 있다. 그리고, 로컬 서버(120)는 이와 같은 선형 보간을 통해 장애가 발생한 RSU(131)의 손실된 교통 데이터

를

로 복구할 수 있다.Referring to FIG. 5 , forward prediction data respectively calculated through a loss data recovery algorithm using a parameter transfer LSTM network of the present invention

and backward prediction data

Given , the local server 120 uses spatial correlation for efficient loss data recovery to perform two prediction data

and

Linear interpolation between In addition, the local server 120 loses traffic data of the RSU 131 that has failed through such linear interpolation.

cast

can be restored with

와

를 아래의 식 7과 같이 순방향 예측 데이터 셋 및 역방향 예측 데이터 셋의 평균 공간 상관 계수로 지정할 수 있다. To this end, the local server 120

Wow

can be designated as the average spatial correlation coefficient of the forward prediction data set and the backward prediction data set as shown in Equation 7 below.

<Equation 7>

이고,

이다. here

ego,

am.

및

의 선형 보간을 통해 복구된 교통 데이터

를 획득할 수 있다.Afterwards, the local server 120 through Equation 8 below

and

traffic data recovered through linear interpolation of

can be obtained.

<Equation 8>

및

에 의해 계산된 가중치

을 사용하여 교통 데이터

를 복구할 수 있다.At this time, the local server 120 is the average spatial correlation coefficient

and

weight calculated by

use traffic data

can be restored

Meanwhile, the method according to the present invention is written as a program that can be executed on a computer and can be implemented in various recording media such as magnetic storage media, optical reading media, and digital storage media.

Implementations of the various techniques described herein may be implemented in digital electronic circuitry, or in computer hardware, firmware, software, or combinations thereof. Implementations may be implemented for processing by, or controlling the operation of, a data processing device, eg, a programmable processor, computer, or number of computers, a computer program product, ie an information carrier, eg, a machine readable storage It may be embodied as a computer program tangibly embodied in an apparatus (computer readable medium) or a radio signal. A computer program, such as the computer program(s) described above, may be written in any form of programming language, including compiled or interpreted languages, as a standalone program or in a module, component, subroutine, or computing environment. It can be deployed in any form including as other units suitable for use. A computer program may be deployed to be processed on one computer or multiple computers at one site or to be distributed across multiple sites and interconnected by a communications network.

Processors suitable for processing a computer program include, by way of example, both general and special purpose microprocessors, and any one or more processors of any kind of digital computer. In general, a processor will receive instructions and data from either read-only memory or random access memory or both. Elements of a computer may include at least one processor that executes instructions and one or more memory devices that store instructions and data. In general, a computer may include, receive data from, transmit data to, or both, one or more mass storage devices for storing data, for example magnetic, magneto-optical disks, or optical disks. may be combined to become Information carriers suitable for embodying computer program instructions and data are, for example, semiconductor memory devices, for example, magnetic media such as hard disks, floppy disks and magnetic tapes, Compact Disk Read Only Memory (CD-ROM). ), optical recording media such as DVD (Digital Video Disk), magneto-optical media such as optical disk, ROM (Read Only Memory), RAM (RAM) , Random Access Memory), flash memory, EPROM (Erasable Programmable ROM), EEPROM (Electrically Erasable Programmable ROM), and the like. Processors and memories may be supplemented by, or included in, special purpose logic circuitry.

In addition, the computer-readable medium may be any available medium that can be accessed by a computer, and may include both computer storage media and transmission media.

While this specification contains numerous specific implementation details, they should not be construed as limitations on the scope of any invention or claim, but rather as descriptions of features that may be specific to particular embodiments of particular inventions. should be understood Certain features that are described herein in the context of separate embodiments may be implemented in combination in a single embodiment. Conversely, various features that are described in the context of a single embodiment may also be implemented in multiple embodiments, either individually or in any suitable subcombination. Furthermore, although features operate in a particular combination and may be initially depicted as claimed as such, one or more features from a claimed combination may in some cases be excluded from the combination, the claimed combination being a sub-combination. or a variant of a sub-combination.

Likewise, although acts are depicted in the drawings in a particular order, it should not be construed that all acts shown must be performed or that such acts must be performed in the specific order or sequential order shown to obtain desirable results. In certain cases, multitasking and parallel processing may be advantageous. Further, the separation of the various device components of the above-described embodiments should not be construed as requiring such separation in all embodiments, and the program components and devices described may generally be integrated together into a single software product or packaged into multiple software products. You have to understand that you can.

On the other hand, the embodiments of the present invention disclosed in the present specification and drawings are merely presented as specific examples to aid understanding, and are not intended to limit the scope of the present invention. It will be apparent to those of ordinary skill in the art to which the present invention pertains that other modifications based on the technical spirit of the present invention can be implemented in addition to the embodiments disclosed herein.

100: V2I network
110: cloud server
120 : local server
130: peripheral base station (RSU)
131: Failed RSU
140: vehicle terminal (OBU)

Claims (14)

A method for recovering lost data, comprising:
receiving N forward data sets and N backward data sets corresponding to a space before and after space for loss data in a specific space;
A first training LSTM layer for the N forward data sets and the N by learning individual Long Short-Term Memory (LSTM) layers arranged in spatial order using the N forward data sets and the N backward data sets. generating a second training LSTM layer for the backward data sets;
extracting forward prediction data and backward prediction data for the specific space using the generated first training LSTM layer and the second training LSTM layer, respectively; and
Recovering lost data for the specific space based on the correlation between the forward prediction data and the backward prediction data
How to recover lost data that includes. According to claim 1,
The generating step is
learning a first LSTM layer over time using a first forward data set among the N forward data sets composed of time series data; and
First training on the forward data sets by sequentially learning the second LSTM layers corresponding to each of the forward data sets by applying the parameters learned through the previous LSTM layer to each of the forward data sets after the first forward data set Steps to create an LSTM layer
A method of recovering lost data that includes. According to claim 1,
The generating step is
learning a third LSTM layer over time using a first backward data set among the N backward data sets composed of time series data; and
Second training on the reverse data sets by sequentially learning the fourth LSTM layers corresponding to each of the reverse data sets by applying the parameters learned through the previous LSTM layer to each of the reverse data sets after the first reverse data set Steps to create an LSTM layer
A method of recovering lost data that includes. According to claim 1,
The extracting step is
Forward prediction data and reverse direction for the specific space by respectively inputting a forward data set corresponding to the immediately preceding space of the specific space and a backward data set corresponding to the immediately following space to the generated first training LSTM layer and the second training LSTM layer Loss data recovery method that extracts predictive data. According to claim 1,
The recovery step is
calculating a weight using a first correlation coefficient for the N forward data sets and a second correlation coefficient for the N backward data sets; and
Recovering lost data for the specific space by applying the calculated weight to the forward prediction data and the backward prediction data
A method of recovering lost data that includes. 6. The method of claim 5,
The weight is
Loss data recovery method calculated through the ratio of the first correlation coefficient and the second correlation coefficient. According to claim 1,
The generating step is
A loss data recovery method for updating each parameter of the individual LSTM layers using a loss function corresponding to a mean squared error (MSE). A lost data recovery device comprising:
including a processor;
The processor is
Receive N forward data sets and N backward data sets corresponding to a space before and after space for loss data in a specific space,
A first training LSTM layer for the N forward data sets and the N by learning individual Long Short-Term Memory (LSTM) layers arranged in spatial order using the N forward data sets and the N backward data sets. Create a second training LSTM layer for the backward data sets of
Extracting forward prediction data and backward prediction data for the specific space using the generated first training LSTM layer and second training LSTM layer, respectively,
Loss data recovery apparatus for recovering lost data for the specific space with a missing value determined based on the correlation between the forward prediction data and the backward prediction data. 9. The method of claim 8,
The processor is
The first LSTM layer is learned over time using the first forward data set among the N forward data sets composed of time series data, and the parameters learned through the previous LSTM layer for each of the forward data sets after the first forward data set Loss data recovery apparatus for generating a first training LSTM layer for the forward data sets by sequentially learning the second LSTM layers corresponding to each of the forward data sets by applying . 9. The method of claim 8,
The processor is
The third LSTM layer is learned over time by using the first backward data set among the N backward data sets composed of time series data, and the parameters learned through the previous LSTM layer for each of the backward data sets after the first backward data set A loss data recovery apparatus for generating a second training LSTM layer for the reverse data sets by sequentially learning the fourth LSTM layers corresponding to each of the reverse data sets by applying . 9. The method of claim 8,
The processor is
Forward prediction data and reverse direction for the specific space by respectively inputting a forward data set corresponding to the immediately preceding space of the specific space and a backward data set corresponding to the immediately following space to the generated first training LSTM layer and the second training LSTM layer Loss data recovery device that extracts predictive data. 9. The method of claim 8,
The processor is
By calculating a weight using the first correlation coefficient for the N forward data sets and the second correlation coefficient for the N backward data sets, and applying the calculated weight to the forward prediction data and the backward prediction data, the Lost data recovery device to recover lost data for a specific space. 13. The method of claim 12,
The weight is
Loss data recovery apparatus calculated through the ratio of the first correlation coefficient and the second correlation coefficient. 9. The method of claim 8,
The processor is
A loss data recovery apparatus for updating each parameter of the individual LSTM layers by using a loss function corresponding to a mean squared error (MSE).

Images