KR101463425B1 - Abnormality observation data detection method using time series prediction model and abnormality observation data of ground water level - Google Patents

Abstract

The present invention relates to a method of detecting anomaly data and a method of detecting anomaly data using a time series prediction model and more particularly to a method of detecting anomaly data based on past results using predicted data .
In particular, the present invention uses a support vector machine (SVM) algorithm to compare predicted results from previously measured data with predicted data from past predicted data to identify abnormal observation data in real time can do.
Specifically, by using weather data such as rainfall and groundwater observation data that affect natural changes in groundwater, and by using a time series model of groundwater level forecasting, it is possible to greatly improve the detection ability of anomaly data on groundwater .
Thus, it is possible to improve reliability and competitiveness in the field of water resources management, particularly in the field of groundwater management, as well as in the field of management of natural resources associated with various climate change and related or similar fields.

Description

[0001] The present invention relates to a method for detecting anomalous observation data using a time series prediction model and a method for detecting anomalous observation data,

The present invention relates to a method of detecting anomaly data and a method of detecting anomaly data using a time series prediction model, and more particularly, to a method of detecting anomaly data based on past data .

Particularly, the present invention can discriminate anomalous observation data by comparing a prediction result by past observation data and a past prediction data by using a support vector machine (SVM) algorithm The present invention relates to a method of detecting anomaly data using a time series prediction model and a method of detecting anomaly data of groundwater.

Climate change is one of the most serious environmental problems in the world today, and its scale is widespread globally, and there are signs in Korea as well as its impacts are spreading throughout society.

In addition, climate change is associated with changes in weather and ecosystems such as floods, droughts, typhoons, etc. In order to respond to these climate changes, evaluation based on the impact of climate change is required for each sector. Is the most important.

In particular, the water resources most closely related to our lives are not only limited by climate change but also due to rapid changes in the industrial structure, increased demand for water due to population growth, and changes in awareness of ecological environment. The role of groundwater as an alternative water resource is increasing.

However, if systematic management of groundwater development and utilization is not carried out, it is difficult to grasp water resources including groundwater, and it causes management problems such as depletion of groundwater due to indiscreet groundwater development.

On the other hand, as the remote management system and automatic observation (measurement) technology using wired / wireless communication network are developed, groundwater real time observation systems such as national groundwater observation network, auxiliary groundwater observation network, and seawater penetration observation network are being constructed.

However, in an apparatus and a system using an electronic device and a program, an error of hardware or an operation error of the software may occur, which may cause an error in observation (measurement) data.

Therefore, in order to secure the reliability of observation data and to make efficient use of observation network and sustainable utilization of groundwater resources, it is required to detect real-time abnormal observation data or error detection method on observation data.

One of these methods would be to detect abnormal data by comparing the actual observed data with the historical data, but this method is limited in that it is difficult to consider the natural changes in the groundwater level in the region can do.

The following conventional patent documents do not disclose a method of providing observation data capable of ensuring sufficient reliability by detecting abnormal data in real time in consideration of the natural change pattern of the groundwater level in the relevant region.

Korean Patent Registration No. 10-1013182 discloses a computer-readable recording medium on which a program for executing calculation of a groundwater discharge amount using groundwater level change, a water level measurement sensor,

In order to solve the above problems, the present invention makes use of a weather-time forecasting time series model using weather data such as rainfall and groundwater observation data that affect the natural change of the groundwater level, This paper presents a method for detecting anomaly data using a time series prediction model and a method for detecting anomaly data.

In other words, it provides a method of detecting anomalous observation data and a method of detecting anomalous groundwater observation data using a time series prediction model in which anomalous observation data can be discriminated based on predicted results using past observation data for predicted objects There is a purpose.

In particular, the present invention uses a support vector machine (SVM) algorithm to compare predicted results from past measured data with past predicted data to determine anomaly data And to provide a method for detecting anomaly data using a time series prediction model capable of improving reliability and a method for detecting anomaly data of groundwater.

In order to achieve the above object, an abnormal observation data detection method using a time series prediction model according to the present invention includes: an observation data group selection step of extracting data for setting a time series prediction model among past observation data to be predicted; A model function generation step of generating a SVM (Support Vector Machine) based model function based on the observation data group; A predicted data calculation step of predicting a change of the predicted object according to current observation data using the model function; And an abnormal data discrimination step of discriminating the abnormal observation data based on the predicted data.

The model function generating step may include: a structure setting step of setting a structure of a model for SVM-based time series prediction; And an optimization prediction function generation step of setting and optimizing the SVM based prediction function, and the model function can be generated by the optimization prediction function.

In addition, the structure setting step may include: an input vector according to the current observation data; A support vector (SV) extracted from the observation data group; A dimension transform function for transforming the input vector and the support vector into a high dimensional vector; And an output function (Output) for calculating predictive data using the kernel function, thereby setting the structure of the model.

The optimization prediction function generation step may include: setting the SVM-based prediction function; Setting an objective function for calculating the prediction function; And a step of setting an optimization condition to the objective function and converting it into the optimization prediction function.

The prediction data calculation step may include: actual basis prediction data predicted by applying past observation data and current observation data to the model function; And at least one of prediction-based prediction data predicted by applying past prediction data and current observation data to the model function.

Also, the abnormal data discrimination step may compare the actual-based prediction data and the prediction-based prediction data, and discriminate the abnormal observation data based on the difference value between the two prediction data.

In addition, the method for detecting anomalous data on groundwater according to the present invention includes extracting a data group in which no anomalous observations have occurred in the past observation data of influential factors including the groundwater level to be predicted and the rainfall affecting the groundwater Selection of observational data group; Based on the observation data set, a production based model function for performing prediction using past observation data based on SVM (Support Vector Machine) and a prediction based model function for performing prediction using past prediction data are generated Model function creation step; A predicted data calculation step for predicting the change of the groundwater level according to the current observation data for each of the actual-based model function and the prediction-based model function; And an abnormal data discriminating step of comparing the predictive data of the actual modeling function and the predictive data of the prediction based model function and determining the corresponding data as abnormal observation data when the difference value between the two predictive data is larger than the abnormal judgment reference value .

The model function generating step may include: a structure setting step of setting a structure of a model for SVM-based time series prediction; And an optimization prediction function generation step of setting and optimizing a prediction function using an SMO (Sequential Minimal Optimization) algorithm, and the model function can be generated by the optimization prediction function.

In addition, the structure setting step may include: an input vector according to the current observation data; A support vector (SV) extracted from the observation data group; A kernel function for high-level transforming the input vector and the support vector into a kernel trick; And an output function (Output) for calculating predictive data using the kernel function, thereby setting the structure of the model.

The kernel function may also include a " radial basis kernel "function of the Gaussian type.

The optimization prediction function generation step may include: setting the SVM-based prediction function to the SMO algorithm; Setting an objective function for calculating the prediction function; And a step of setting an optimization condition to the objective function and converting it into the optimization prediction function.

In addition, the optimization prediction function generating step may employ a saddle point problem using a Lagrangian multiplier to calculate the objective function.

According to the above solution, the present invention is advantageous in that it is possible to accurately determine anomaly observation data in real time based on predicted results using past observation data with respect to a predicted object.

In particular, the reliability of the discrimination of the real-time abnormal observation data can be greatly improved by processing the prediction factors and the influential factors influencing the prediction objects in parallel.

Specifically, by using weather data such as rainfall and groundwater observation data that affect natural changes in groundwater, and by using a time series model of groundwater level forecasting, it is possible to greatly improve the detection ability of anomaly data on groundwater There is an effect that can be made.

Thus, it is possible to improve reliability and competitiveness in the field of water resources management, particularly in the field of groundwater management, as well as in the field of management of natural resources associated with various climate change and related or similar fields.

1 is a flowchart for explaining an abnormal observation data detection method using a time series prediction model according to the present invention.
FIG. 2 is a flowchart for explaining a method for detecting anomalous observation data on groundwater according to the present invention.
FIG. 3 is a view for explaining the concept of the method of detecting anomaly data according to the present invention.
4 is a conceptual diagram for explaining the structure of a prediction model according to the present invention.
FIG. 5 is a graph showing the amount of rainfall and the groundwater level data for applying the process of FIG.
FIG. 6 is a graph showing the results predicted by the production-based model function shown in FIG.
7 is a graph showing the results predicted by the prediction-based model function shown in FIG.
FIG. 8 is a graph showing the difference between the actual-based model function and the predictive-based model function shown in FIG.

The method of detecting anomaly data using a time series prediction model according to the present invention and the method of detecting anomaly data of an underground water can be applied in various ways. In the following, a preferred embodiment will be described with reference to the accompanying drawings. .

1 is a flowchart for explaining an abnormal observation data detection method using a time series prediction model according to the present invention.

Referring to FIG. 1, an abnormal observation data detection method using a time series prediction model selects an observation data group for extracting data for setting a time series prediction model among past observation data to be predicted (step S100). Here, it is preferable that the extracted observation data group is selected as a data group in which the observation value of the past is not considered to have occurred.

When the observation data set is selected, a SVM (Support Vector Machine) based model function is generated based on the observation data set (Step S200).

In the process of generating the model function, it is possible to generate the optimization prediction function by setting and optimizing the SVM-based prediction function after setting the basic structure of the model for the SVM-based time series prediction. The model function can be generated.

In addition, the basic structure of the model for time series prediction is composed of an input vector according to the current observation data, a support vector (SV) extracted from the observation data group, a dimension for high- A transform function, and an output function (Output) for calculating the predicted data using the dimension transform function.

Also, the optimization prediction function can be generated by setting an SVM-based prediction function, setting an objective function for calculating a prediction function, and then setting an optimization condition to an objective function and performing conversion.

When the model function is generated, the generated model function is used to predict the change of the prediction object according to the current observation data and to calculate the prediction data (step S300). For example, the prediction data may include actual-based prediction data predicted by applying past observation data and current observation data, past prediction data, and prediction-based prediction data predicted by applying current observation data.

When the predictive data including the actual-based predictive data and the predictive-based predictive data are calculated, the anomalous observation data is identified (discriminated) based on the calculated predictive data (step S400). For example, the determination of the abnormal observation data can be made by comparing the actual-based prediction data and the prediction-based prediction data, and determining the case where the difference between the two prediction data is large, as the abnormal observation data.

1 described above can be applied to various natural resources. Hereinafter, the case of applying the present invention to water resources, particularly, groundwater will be described.

FIG. 2 is a flowchart for explaining a method for detecting anomalous observation data on groundwater according to the present invention. Hereinafter, a specific embodiment of FIG. 2 will be described with reference to FIGS. 3 to 8. FIG.

2 and 3, among the past observation data (data prior to the time point t-1) of the influencing factors including the amount of rainfall affecting the groundwater level and the groundwater to be predicted, a data group tk Tk-tau) are extracted and selected as observation data groups (step S110). For example, as shown in FIG. 5, the observation data of 2000 to 2001 can be extracted as a data group.

Based on the selected observational data group, a prediction based on a support vector machine (SVM) is performed using a model based on a model based on past observation data (model A) and past prediction data To generate a prediction-based model function (model B) (step S120).

At this time, the prediction model can be expressed by Equations (1) and (2).

(1)

(2)

Where f is the model A function, fB is the model B function, a is the delay time of the explanatory variable, b is the delay time of the dependent variable, to be.

5, the present observation data (t + 1 in FIG. 3) is applied to the data of 2002 to 2004 for each of the production-based model function of Equation 1 and the prediction-based model function of Equation 2, (Step S130).

Based on the prediction results, the prediction data of the actual-based model function and the prediction-based model function shown in FIG. 7 are compared as shown in FIG. 6. When the difference value between the two prediction data shown in FIG. 8 is larger than the error judgment reference value (Dotted line portion in Fig. 8), and determines the corresponding data as anomalous observation data (step S140).

Hereinafter, a method of generating (generating) a time series model will be described in more detail.

The present invention can generate a time series model by a support vector machine (SVM) algorithm.

First, the SVM algorithm is based on Structural Risk Minimization (SRM) rather than Empirical Risk Minimization (ERM) of general learning models such as ANN (Artificial Neural Network) And it can show higher generalization performance than general learning method.

In terms of classification of data, existing methods including ANN are designed to minimize classification error rate by learning data, but SVM is designed to maximize the margins existing among data groups to be classified , The model generalization ability can be maximized.

The overall structure of the SVM is as shown in FIG.

Referring to FIG. 4, SVM looks similar to ANN, but its principle is different. The structure of the ANN is a structure in which the input and output vector components are set as nodes and the data given to the nodes are input. On the other hand, the SVM is suitable for supporting the model among input / output vectors to be used for training The vectors are selected to construct the model. The vectors thus selected are called a support vector (SV).

When a new input vector is input to the SVM composed of the selected SV, a new output value is predicted through calculation with the support vector (SV).

In this process, SV and input vector move to higher dimension by the transition of kernel function, and nonlinear part can be calculated more easily in this elevated dimension, and this method is called kernel trick .

In the present invention, a "radial basis kernel" function of Gaussian form can be used as a kernel function as shown in Equation (3).

(3)

On the other hand, although SVM has excellent learning and generalization ability, since the SVM solution is obtained by solving the second-order quadratic programming, the time required for learning increases exponentially as the number of learning data increases.

In order to solve this problem, the present invention minimizes the second-order scale parameter to be calculated by using the sequential minimal optimization (SMO) algorithm, Can be improved. This makes real-time processing possible.

The prediction function by SVM can be expressed by Equation (4), and the objective function for obtaining the prediction function can be expressed by Equation (5).

(4)

(5)

Here, w is the connection strength vector, b is the degree of bias, φ is the nonlinear transfer function, and the nonlinear transfer function φ can be expressed as the kernel function K below.

ξ is a parameter giving a penalty for errors in the training process through a specific loss function, ε is an error tolerance, and C is an antagonistic parameter that provides the degree of experience error in the optimization problem.

In Equation (5), the problem is solved through a saddle point problem using a Lagrangian multiplier.

Equation (5) can be transformed as shown in Equation (6) by using the Lagrange multiplier (?,?).

(6)

Further, Equation (6) can be solved as Equation (8) by satisfying the optimization condition of Equation (7).

(7)

(8)

라 하고 수학식 8를 β에 대한 2차 함수 문제로 만들어서, β의 다양한 구간에 대해 최소값의 해석해를 반복적으로 구함으로써, SV와 그 연결강도를 결정하여 SVM 예측함수를 제공할 수 있다.In SMO

(8) as a quadratic function problem for β, and by repeatedly finding the minimum value analysis solution for various intervals of β, it is possible to provide the SVM prediction function by determining the SV and its connection strength.

Therefore, it is possible to accurately identify the anomaly data in real time based on the predicted results using past observation data on the groundwater to be predicted.

The method of detecting anomaly data using the time series prediction model according to the present invention and the method of detecting anomaly data of groundwater are described above. It will be understood by those skilled in the art that the technical features of the present invention may be embodied in other specific forms without departing from the spirit or essential characteristics thereof.

Therefore, it is to be understood that the embodiments described above are intended to be illustrative, and not restrictive, in all respects, and that the scope of the invention is indicated by the appended claims rather than the foregoing description, And all equivalents and modifications that come within the meaning and range of equivalency of the claims are to be embraced within their scope.

Claims (12)

delete delete delete delete delete delete An observation data group selection step of extracting a data group in which anomalous observation has not occurred in the past observation data of the influence factors including the predicted groundwater level and the rainfall affecting the groundwater;
Based on the observation data set, a production based model function for performing prediction using past observation data based on SVM (Support Vector Machine) and a prediction based model function for performing prediction using past prediction data are generated Model function creation step;
A predicted data calculation step for predicting the change of the groundwater level according to the current observation data for each of the actual-based model function and the prediction-based model function; And
Based model function and predictive data of the prediction-based model function, and if the difference value between the two predictive data is greater than the anomaly determination reference value, determining the corresponding data as anomalous observation data; Including,
The model function generating step may include:
A structure setting step of setting a structure of a model for SVM-based time series prediction; And
Generating a model function by the optimization prediction function including an optimization prediction function generation step of setting and optimizing a prediction function using an SMO (Sequential Minimal Optimization) algorithm,
Wherein the structure setting step comprises:
An input vector according to the current observation data;
A support vector (SV) extracted from the observation data group;
A kernel function for high-level transforming the input vector and the support vector into a kernel trick; And
And an output function (Output) for calculating predicted data using the kernel function to set a structure of the model,
The optimization prediction function generation step includes:
Setting the SVM-based prediction function to the SMO algorithm;
Setting an objective function for calculating the prediction function; And
And setting an optimization condition to the objective function to convert the optimization function into the optimization prediction function.
delete delete 8. The method of claim 7,
The kernel function may include:
And a "radial basis kernel" function of a Gaussian type.
delete 8. The method of claim 7,
The optimization prediction function generation step includes:
Wherein a saddle point problem using a Lagrangian multiplier is applied to calculate the objective function.

Images