KR102666070B1 - Apparatus for providing data-driven parametric insurance and method thereof - Google Patents

Abstract

The present invention discloses a data-based parametric insurance providing device and method. In other words, the present invention can provide a framework to insurance providers who want to introduce index insurance by accurately predicting the degree of risk due to a specific disaster using artificial intelligence based on accumulated data, and can provide a framework for the specific disaster. The trigger selection process according to the degree of risk can be efficiently performed.

Description

Apparatus for providing data-driven parametric insurance and method thereof}

The present invention relates to a device and method for providing data-based parametric insurance, and in particular, to a device and method for providing data-based parametric insurance that accurately predicts the degree of risk due to a specific disaster using artificial intelligence based on accumulated data. It's about.

Insurance is a system that requires a person to pay a certain amount of money in advance in preparation for something that may occur in the future, such as an accidental accident or illness, and pays a certain amount accordingly when the agreed conditions are met.

In order to pay benefits under such insurance, consultation between the insured and the insurance provider is necessary, and sometimes the insurance provider must inspect the insured's damages in detail.

This insurance payment method requires a lot of time and money in developing countries.

Additionally, as climate change poses new and unpredictable challenges to society, insurance is an essential means of protecting users from losses due to extreme events.

These traditional insurance risk models use statistical analysis that is inaccurate and increasingly flawed as climate change makes weather more erratic and extreme.

In addition, these traditional insurance risk models have difficulty predicting the degree of risk due to a specific disaster, making it difficult to apply the traditional insurance risk models to index insurance.

Korean Patent Publication No. 10-2021-0034631 [Title: Business methods, devices and systems for monitoring quantitative and qualitative environmental analysis and managing related financial transactions]

The purpose of the present invention is to provide a data-based parametric insurance providing device and method that accurately predicts the degree of risk due to a specific disaster using artificial intelligence based on accumulated data.

A device for providing data-based parametric insurance according to an embodiment of the present invention includes a collection unit that collects damage data from an information provision server; And performing preprocessing on the collected damage data, performing artificial intelligence-based machine learning based on the preprocessed damage data, and determining the connectivity dropout probability for the network in relation to the damage data based on the machine learning results. Generate, determine whether the generated connectivity dropout probability can be used as a main trigger related to a specific parametric insurance, and as a result of the determination, determine whether the generated connectivity dropout probability can be used as a main trigger related to a specific parametric insurance. When available, it may include a control unit that adds conditions related to the damage data to a preset trigger in response to the specific parametric insurance.

As an example related to the present invention, the damage data is set according to the type of parametric insurance, and in the case of cyber insurance, includes personal information leakage, hacking, weather information, location of occurrence and date of occurrence, and storm and flood disasters. For insurance, it includes weather information, location and date of occurrence, and for earthquake insurance, it includes seismic intensity, location and date of occurrence, and the weather information includes wind speed, precipitation, temperature, cloud level, visibility/visibility, altimeter settings, It may include at least one of sea level, station pressure, dew point, wind direction, wind gust, and snow depth.

As an example related to the present invention, the control unit may be controlled to calculate an event occurrence probability density function distribution related to the damage data and display the calculated event occurrence probability density function distribution related to the damage data on the display unit.

A method of providing data-based parametric insurance according to an embodiment of the present invention includes collecting damage data from an information provision server by a collection unit; Preprocessing the collected damage data by a control unit; performing artificial intelligence-based machine learning based on the preprocessed damage data, by the control unit, and generating a connectivity dropout probability for a network with respect to the damage data based on the machine learning results; determining, by the control unit, whether the generated connectivity dropout probability can be utilized as a main trigger related to a specific parametric insurance; And, as a result of the determination, when the generated connectivity dropout probability can be used as a main trigger related to a specific parametric insurance, the control unit sets conditions related to the damage data to a preset trigger corresponding to the specific parametric insurance. It may include the step of adding .

As an example related to the present invention, the step of performing preprocessing on the collected damage data includes performing binning included in a preset Monte Carlo sampling method on the collected damage data, A process of removing noise included in the damage data; And it may include a process of extracting preset reference damage data from the damage data from which the noise has been removed.

As an example related to the present invention, generating a connectivity dropout probability for a network in relation to the damage data includes extracting the preprocessed damage data into individual damage data; and performing the artificial intelligence-based machine learning for each extracted individual damage data, respectively, and generating a connectivity dropout probability for the network for each individual damage data based on the machine learning results.

As an example related to the present invention, the step of determining whether the generated connectivity dropout probability can be used as a main trigger related to a specific parametric insurance involves determining that the connectivity dropout probability for the network generated in relation to the damage data is It is possible to determine whether a preset threshold is exceeded in response to a specific parametric insurance.

As an example related to the present invention, as a result of the determination, when the generated connectivity dropout probability cannot be used as a main trigger related to a specific parametric insurance, the control unit determines the specific parametric insurance with respect to the damage data. It may further include displaying information indicating that it cannot be used as a main trigger through a display unit.

As an example related to the present invention, the trigger is composed of different parameters depending on the type of parametric insurance, and different thresholds may be set for each parameter.

An example related to the present invention includes calculating, by the control unit, an event occurrence probability density function distribution related to the damage data; And it may further include displaying, by the control unit, the event occurrence probability density function distribution related to the calculated damage data on a display unit.

The present invention can provide a framework to insurance providers who want to introduce index insurance by accurately predicting the degree of risk due to a specific disaster using artificial intelligence based on accumulated data, and can provide a framework for the specific disaster. It has the effect of efficiently carrying out the trigger selection process according to the level of risk.

Figure 1 is a block diagram showing the configuration of a data-based parametric insurance provision device according to an embodiment of the present invention.
2 and 3 show correlation matrices between dropout probability and various meteorological parameters according to an embodiment of the present invention.
Figure 4 is a diagram illustrating examples of unique weather parameters most effective for dropout rates for each state according to an embodiment of the present invention.
Figure 5 is a diagram showing an example of accuracy in estimating the dropout probability between DSPP and Bayesian regression according to an embodiment of the present invention.
Figure 6 is a diagram showing a two-dimensional view of the dropout probability estimated from DSPP and Bayesian regression analysis in California and Texas according to an embodiment of the present invention.
Figure 7 is a diagram showing a three-dimensional view with uncertainty estimation in DSPP and Bayesian regression analysis of California and Texas according to an embodiment of the present invention.
Figure 8 is a diagram comparing estimates of DSPP, DSPP1, and DSPP2 in California and Texas according to an embodiment of the present invention.
Figures 9 and 10 are diagrams showing a three-dimensional view along with uncertainty estimation in DSPP and Bayesian regression analysis in California and Texas according to an embodiment of the present invention.
Figure 11 is a diagram showing an example of accuracy for estimating dropout probability using DSPP1 and DSPP2 according to an embodiment of the present invention.
Figure 12 is a flowchart showing a method for providing data-based parametric insurance according to an embodiment of the present invention.

It should be noted that the technical terms used in the present invention are only used to describe specific embodiments and are not intended to limit the present invention. In addition, the technical terms used in the present invention, unless specifically defined in a different sense in the present invention, should be interpreted as meanings generally understood by those skilled in the art in the technical field to which the present invention pertains, and are not overly comprehensive. It should not be interpreted in a literal or excessively reduced sense. Additionally, if the technical terms used in the present invention are incorrect technical terms that do not accurately express the idea of the present invention, they should be replaced with technical terms that can be correctly understood by those skilled in the art. In addition, general terms used in the present invention should be interpreted according to the definition in the dictionary or according to the context, and should not be interpreted in an excessively reduced sense.

Additionally, as used in the present invention, singular expressions include plural expressions unless the context clearly dictates otherwise. In the present invention, terms such as “consists of” or “comprises” should not be construed as necessarily including all of the various components or steps described in the invention, and some of the components or steps may not be included. It may be possible, or it should be interpreted as being able to further include additional components or steps.

Additionally, terms containing ordinal numbers, such as first, second, etc., used in the present invention may be used to describe constituent elements, but the constituent elements should not be limited by the terms. Terms are used only to distinguish one component from another. For example, a first component may be named a second component without departing from the scope of the present invention, and similarly, the second component may also be named a first component.

Hereinafter, preferred embodiments according to the present invention will be described in detail with reference to the attached drawings. However, identical or similar components will be assigned the same reference numbers regardless of the reference numerals, and duplicate descriptions thereof will be omitted.

Additionally, when describing the present invention, if it is determined that a detailed description of related known technologies may obscure the gist of the present invention, the detailed description will be omitted. In addition, it should be noted that the attached drawings are only intended to facilitate easy understanding of the spirit of the present invention, and should not be construed as limiting the spirit of the present invention by the attached drawings.

Figure 1 is a block diagram showing the configuration of a data-based parametric insurance providing device 100 according to an embodiment of the present invention.

As shown in FIG. 1, the data-based parametric insurance providing device 100 consists of a collection unit 110, a storage unit 120, a display unit 130, a voice output unit 140, and a control unit 150. . Not all of the components of the data-based parametric insurance providing device 100 shown in FIG. 1 are essential components, and the data-based parametric insurance providing device 100 is made up of more components than those shown in FIG. 1. may be implemented, or the data-based parametric insurance providing device 100 may be implemented with fewer components.

The data-based parametric insurance providing device 100 can be used in a smart phone, a portable terminal, a mobile terminal, a foldable terminal, or a personal digital assistant (PDA). ), PMP (Portable Multimedia Player) terminal, telematics terminal, navigation terminal, personal computer, laptop computer, Slate PC, Tablet PC, ultrabook ), wearable devices (e.g., including watch type terminal (Smartwatch), glass type terminal (Smart Glass), HMD (Head Mounted Display), etc.), Wibro terminal, IPTV (Internet Protocol Television) terminal , can be applied to various terminals such as smart TVs, digital broadcasting terminals, AVN (Audio Video Navigation) terminals, A/V (Audio/Video) systems, flexible terminals, and digital signage devices.

Additionally, the data-based parametric insurance providing device 100 may be implemented in the form of a web server, database server, proxy server, etc. In addition, the data-based parametric insurance providing device 100 may be installed with a network load balancing mechanism or at least one of various software that allows the data-based parametric insurance providing device 100 to operate on the Internet or another network. This can be implemented as a computerized system. Additionally, the network may be an http network, a private line, an intranet, or any other network. Furthermore, the connection between the data-based parametric insurance providing device 100, another terminal (not shown), another server (not shown), a database (not shown), and a bank server (not shown) allows the data to be stored by a random hacker or other It can be connected to a secure network to avoid attacks by third parties. In addition, the data-based parametric insurance providing device 100 may include a plurality of database servers, and these database servers may provide the data-based parametric insurance through any type of network connection, including a distributed database server architecture. It can be implemented in a way that it is connected separately from (100).

Additionally, the data-based parametric insurance providing device 100 may be an insurance company server (not shown) operated by an insurance company.

The collection unit 110 includes a communication unit (not shown). Here, the communication unit communicates with any internal component or at least one external terminal through a wired/wireless communication network. At this time, the external arbitrary terminal may include the RF generator 100, a server (not shown), a terminal (not shown), etc. Here, wireless Internet technologies include Wireless LAN (WLAN), DLNA (Digital Living Network Alliance), Wibro (Wireless Broadband: Wibro), Wimax (World Interoperability for Microwave Access: Wimax), and HSDPA (High Speed Downlink Packet Access). ), HSUPA (High Speed Uplink Packet Access), IEEE 802.16, Long Term Evolution (LTE), LTE-A (Long Term Evolution-Advanced), Wireless Mobile Broadband Service (WMBS), etc. In this case, the communication unit transmits and receives data according to at least one wireless Internet technology, including Internet technologies not listed above. In addition, short-range communication technologies include Bluetooth, RFID (Radio Frequency Identification), Infrared Data Association (IrDA), UWB (Ultra Wideband), ZigBee, and Near Field Communication (NFC). , Ultrasound Communication (USC), Visible Light Communication (VLC), Wi-Fi, Wi-Fi Direct, etc. may be included. Additionally, wired communication technologies may include Power Line Communication (PLC), USB communication, Ethernet, serial communication, optical/coaxial cables, etc.

Additionally, the communication unit can transmit information to and from any terminal through a universal serial bus (USB).

In addition, the communication unit sets technical standards or communication methods for mobile communication (e.g., Global System for Mobile communication (GSM), Code Division Multi Access (CDMA), Code Division Multi Access 2000 (CDMA2000), and EV-DO ( Enhanced Voice-Data Optimized or Enhanced Voice-Data Only), Wideband CDMA (WCDMA), High Speed Downlink Packet Access (HSDPA), High Speed Uplink Packet Access (HSUPA), Long Term Evolution (LTE), Long Term (LTE-A) Wireless signals are transmitted and received with a base station, the RF generator 100, the server, the terminal, etc. on a mobile communication network built according to (Evolution-Advanced), etc.).

In addition, the communication unit transmits data (or information) to the RS generator 100, etc. through serial communication such as RS-232, RS-485, UART, etc., and Ethernet communication such as TCP/IP, EtherCAT, Modbus-TCP, etc. You can also send/receive.

In addition, the collection unit 110 collects damage data from an information providing server (not shown) under the control of the control unit 150, at preset time intervals or at the request of the control unit 150. Here, the information provision server may be a server that collects various information and shares (or provides) the collected information for a fee or free of charge. In addition, the damage data (or raw data) is set according to the type of parametric insurance, and in the case of cyber insurance, it includes personal information leakage, hacking, weather information, location of occurrence, date of occurrence, etc., and In the case of insurance, it includes weather information, location of occurrence, date of occurrence, etc., and in the case of earthquake insurance, it includes seismic intensity, location of occurrence, date of occurrence, etc. Here, the weather information (or weather information) includes wind speed, precipitation, temperature, cloud level, visibility/visibility, altimeter settings, sea level, station pressure, dew point, wind direction, gusts, snowfall, etc.

The storage unit (or memory) 120 stores various user interfaces (UIs), graphic user interfaces (GUIs), etc.

Additionally, the storage unit 120 stores data and programs necessary for the data-based parametric insurance providing device 100 to operate.

That is, the storage unit 120 controls a plurality of applications (application programs or applications) running on the data-based parametric insurance providing device 100 and the operation of the data-based parametric insurance providing device 100. You can store data and commands for At least some of these applications may be downloaded from an external server via wireless communication. Additionally, at least some of these applications may exist on the data-based parametric insurance providing device 100 from the time of shipment for the basic functions of the data-based parametric insurance providing device 100. Meanwhile, the application program is stored in the storage unit 120, installed on the data-based parametric insurance providing device 100, and controls the operation (or function).

In addition, the storage unit 120 is a flash memory type, hard disk type, multimedia card micro type, and card type memory (for example, SD or XD). memory, etc.), magnetic memory, magnetic disk, optical disk, RAM (Random Access Memory), SRAM (Static Random Access Memory), ROM (Read-Only Memory: ROM), EEPROM (Electrically Erasable Programmable Read-Only Memory), It may include at least one storage medium of PROM (Programmable Read-Only Memory). In addition, the data-based parametric insurance providing device 100 may operate web storage that performs the storage function of the storage unit 120 on the Internet, or may operate in relation to the web storage. .

In addition, the storage unit 120 stores various information (including, for example, damage data, etc.) collected through the collection unit 110 under the control of the control unit 150.

The display unit (or display unit) 130 can display various contents, such as various menu screens, using the user interface and/or graphic user interface stored in the storage unit 120 under the control of the control unit 150. there is. Here, the content displayed on the display unit 130 includes various text or image data (including various information data) and a menu screen including data such as icons, list menus, and combo boxes. Additionally, the display unit 130 may be a touch screen.

In addition, the display unit 130 may be a liquid crystal display (LCD), a thin film transistor-liquid crystal display (TFT LCD), an organic light-emitting diode (OLED), or a flexible display. It may include at least one of a flexible display, 3D display, e-ink display, LED (Light Emitting Diode), beam projector, goggle-type VR, and hologram.

In addition, the display unit 130 displays various information (including, for example, damage data, etc.) collected through the collection unit 110 under the control of the control unit 150.

The audio output unit 140 outputs audio information included in a signal processed by the control unit 150. Here, the audio output unit 140 may include a receiver, speaker, buzzer, etc.

Additionally, the voice output unit 140 outputs a guidance voice generated by the control unit 150.

In addition, the audio output unit 140 provides audio information (or sound effects) corresponding to various information (including, for example, damage data, etc.) collected through the collection unit 110 under the control of the control unit 150. outputs.

The controller (or processor) (controller, or microcontroller unit (MCU) 150) executes the overall control function of the data-based parametric insurance providing device 100.

In addition, the control unit 150 executes the overall control function of the data-based parametric insurance providing device 100 using the program and data stored in the storage unit 120. The control unit 150 may include RAM, ROM, CPU, GPU, and a bus, and the RAM, ROM, CPU, GPU, etc. may be connected to each other through a bus. The CPU can access the storage unit 120, perform booting using the O/S stored in the storage unit 120, and use various programs, content, data, etc. stored in the storage unit 120. This allows you to perform various operations.

Additionally, the control unit 150 uses a plurality of damage data (or raw data) collected in advance as data for continuous machine learning (or deep learning). Here, the input dataset for machine learning is a training set (or training data) and a test set (or training data) of the plurality of damage data in a preset ratio (including, for example, 7:3, 8:2, etc.) By dividing it into a test set (or test/exam/verification data), training and testing functions can be performed.

In addition, the input dataset for machine learning includes multiple damage data collected later. In addition, the output dataset for machine learning is the part to be predicted, and is related to parametric insurance managed by the data-based parametric insurance providing device 100 by learning the collected damage data and predicting it at a later date. Includes connectivity dropout probability, etc.

That is, the control unit 150 sets a specific parameter for specific damage data, etc. in relation to specific raw data for the deep sigma point process (or the Bayesian neural network) included in the artificial neural network (or machine learning model) through preset learning data. It performs a learning function to generate (or calculate/calculate) connectivity dropout probability related to metric insurance. At this time, the control unit 150 stores raw data (or damaged data) in parallel and distributed, and unstructured data, structured data, and semi-structured data included in the stored raw data (or damaged data, etc.) Clean the data (semi-structured), perform preprocessing including classification into metadata, perform analysis including data mining on the preprocessed data, and learn based on at least one type of machine learning. You can build big data by conducting training and testing. At this time, at least one type of machine learning is Supervised Learning, Semi-Supervised Learning, Unsupervised Learning, Reinforcement Learning, and Deep Reinforcement Learning. It may be any one or a combination of at least one of them.

In this way, the control unit 150 performs a learning function on the deep sigma point process (or the Bayesian neural network) in the form of neural networks, etc. through the learning data.

In addition, the control unit 150 uses the Deep Sigma Point Process (DGP), a variant of the Deep Gaussian Process (DGP), to minimize basic risk and better estimate the consequences of extreme weather. Process: DSPP) is used (or applied).

Here, the deep Gaussian process is classified as a technique in the Bayesian Neural Network (BNN) framework. Here, the Bayesian neural network is different from existing neural networks because it can estimate uncertainty about the estimation. For example, a traditional neural network can always tell you how hazardous weather conditions are, but those decisions may be unreliable due to insufficient examples or unstable training processes. In contrast, the Bayesian neural network can be useful in the insurance business because it can estimate model reliability and apply more conservative policies for weather conditions with high uncertainty. In addition, the Bayesian neural network is prone to overfitting because the deterministic and fixed estimation of parameters in the existing neural network is replaced with a probability distribution representing uncertainty due to lack of data, optimal model architecture, etc. Robust against overfitting.

The deep sigma point process (DSPP) used in embodiments of the present invention allows the Gaussian process to directly represent uncertainty in function space, which can lead to better estimates than other Bayesian neural networks with weight space probability distributions.

Additionally, in an embodiment of the invention, weather parameters for dropouts on residential links are determined using a public data set collected over 8 years by ThunderPing systems (not shown) across the United States. , develop (or create/use) a framework that includes the deep sigma point process to create a model that can predict the impact of weather parameters).

Additionally, the control unit 150 collects information about the occurrence of dropout in a residential network link from the ThunderPing system through the collection unit 110. Here, the collected information includes the time when the dropout occurred, temperature, wind speed, precipitation, snow depth, cloud ceiling, visibility, Includes sea level, station pressure, etc. At this time, the control unit 150 excludes information whose records are incomplete or invalid among the collected information (or parameters).

The dropouts are extreme events and are very rare in the dataset. Therefore, the dropout is greatly affected (or suffers) from noise, which greatly interferes with the training of the neural network.

Accordingly, instead of estimating the occurrence of dropout from raw data, the control unit 150 uses the two methods described below.

That is, first, the occurrence of dropout is expressed as a dropout probability, and the dropout probability is calculated as the time to confirm the response divided by the time when the dropout occurs.

, ... ,

등 포함)의 특정 조건들에 해당한다.Second, to suppress undesirable effects from extreme cases, the control unit 150 uses an estimated mean of the dropout probability, taking into account the bins of measurements of the ThunderPing system. do. Here, each bin contains meteorological parameters (e.g.

, ... ,

(including, etc.) falls under certain conditions.

에 대해서, 본 발명의 실시예에서는 구간(interval)

을 제안한다. 여기서, 상기

는 상태에서 최솟값(

)과 최댓값(

) 사이에서 랜덤하게 선택된 실수(real number)이다. 또한, 상기 간격폭(interval width)(

)은

로 설정된다. 그 후, 빈(bin)은 파라미터들이 다음의 [수학식 1]을 만족하는 측정의 집합으로 정의한다.Specifically, the parameters

Regarding, in the embodiment of the present invention, the interval

suggests. Here, the above

The minimum value (

) and the maximum value (

) is a real number randomly selected among the ranges. In addition, the interval width (interval width) (

)silver

is set to Afterwards, a bin is defined as a set of measurements whose parameters satisfy the following [Equation 1].

)을 계산한다.Additionally, for each measurement (j) in a bin, the control unit 150 determines the dropout probability (

) is calculated.

)에 대한 상기 드롭아웃 확률을 근사화한다.In addition, the measured number of bins is represented by N, and the control unit 150 uses the median of means of the dropout probability to bin (

) approximates the dropout probability for .

)에 대한 상기 드롭아웃 확률을 근사화한다.That is, the control unit 150 determines the bin (in Equation 2) below:

) approximates the dropout probability for .

Here, q is a natural number smaller than N.

을 만족하는 빈을 고려(또는 사용)한다.The control unit 150 is

Consider (or use) a bean that satisfies .

In addition, q is selected so that the maximum value of k in the previous [Equation 2] is greater than 20.

)을 수집한다.Accordingly, the control unit 150 selects pairs from multiple bins as input to various models in order to estimate the dropout probability in meteorological parameters.

) is collected.

In this way, the control unit 150 performs a pre-processing process on the collected data (or damage data).

Additionally, the control unit 150 performs Bayesian regression using a conventional risk estimation scheme to compare the dropout probability estimate using deep sigma point processes (DSPPs).

Additionally, the control unit 150 applies (or considers/uses) the following model of [Equation 3] to reflect the fact that the weather parameters and the dropout probability may be nonlinear.

,

및

는 자연수이고, 상기

은 상기 평균(

)과 상기 표준 편차(

)를 갖는 정규 분포(normal distribution)이다. 또한, 상기 표준 편차(

)는 반정규 분포(half-normal distribution)에서 샘플링되고, 가중치(

)는 정규 분포에서 샘플링된다.Here, the above

,

and

is a natural number, and

is the above average (

) and the above standard deviation (

It is a normal distribution with ). Additionally, the standard deviation (

) is sampled from a half-normal distribution, and the weight (

) is sampled from a normal distribution.

Additionally, the control unit 150 fixes the number of weather parameters to 3 based on the initial selection of parameters.

개의 샘플들과 6개의 체인들(chains)로 몬테카를로 시뮬레이션(Monte Carlo simulation)을 수행한다.In addition, the control unit 150 is configured to optimize all model parameters.

Monte Carlo simulation is performed with 6 samples and 6 chains.

Additionally, the control unit 150 confirms that the R-hat value for all model parameters is 1 to confirm appropriate convergence.

In this way, the control unit 150 estimates the dropout probability using the deep sigma point process (DSPP) in order to apply Bayesian regression.

Additionally, the control unit 150 fits the deep sigma point process model to the preprocessed data set. Here, the deep sigma point process (DSPP) is a modification of the original deep Gaussian process (DGP), as previously introduced.

Additionally, in the training of the deep Gaussian process (DGP), the loss function is derived using the evidence lower bound due to the intractability of the model.

Additionally, latent function values in each layer are randomly sampled, which makes inference difficult and may deteriorate the predictive distribution. Here, using the deep sigma point process, such problems can be alleviated.

Additionally, the control unit 150 approximates a variational distribution having eight quadrature points and weights with a quadrature rule called QR3.

Additionally, the control unit 150 uses (or configures) a network composed of four layers whose output vector dimensions are 5, 3, 3, and 1.

Additionally, the covariance matrix of all layers is calculated using the Matern covariance function and a scalar scaler.

The mean vectors of the first three layers are modeled as fully-connected layers.

Additionally, the last layer uses a constant mean.

Additionally, as in the training of the original DGP, the control unit 150 uses 300 inducing points to calculate Kullback-Leibler divergence in all layers.

Additionally, the control unit 150 trains for 500 epochs for each state starting from an initial rate of 0.1, and sets the learning rate to 0.1 at the 100th, 250th, 350th, and 450th epochs. Decay until. At this time, the control unit 150 uses a mini-batch of size 1000.

In this way, the control unit 150 performs a training process for the deep sigma point process (DSPP).

Additionally, the meteorological parameters suffer from noise and unexpected fluctuations.

Accordingly, the control unit 150 predicts that if too many parameters (for example, n >> 1) are used, different noises and undesirable effects will be combined during the training process, and in [Equation 1] above, Because the intersection of the intervals is very small, it may be difficult to find a bin with enough measurements.

Therefore, in the initial stage, the control unit 150 roughly selects only some of the available parameters.

That is, the control unit 150 roughly selects only a portion of the available parameters by calculating the correlation value between the dropout probability and each parameter per measurement.

The control unit 150 selects three parameters (eg, wind speed, precipitation, and temperature) from among the weather parameters based on the correlation value.

That is, because each parameter is subjective to its own noise, using all available meteorological parameters can be problematic.

Accordingly, the control unit 150 selects some meteorological parameters that can well explain the dropout.

However, it is difficult to know the effect of meteorological parameters on the dropout until measurements are actually collected in multiple bins and a train model.

Accordingly, the control unit 150 constructs correlation matrices and selects some representative parameters in order to obtain an initial guess for a set of appropriate meteorological parameters.

For example, Figures 2 and 3 show correlation matrices between the dropout probability and various meteorological parameters.

Additionally, FIGS. 2 and 3 include hierarchical clustering results calculated using the UPGMA (unweighted pair group method with arithmetic mean) algorithm.

Additionally, in order to ensure a sufficient amount of information about the dropout, it is better to select parameters that are orthogonal to each other.

That is, as shown in Figures 2 and 3 above, in both states, (cloud level, visibility), (temperature, dew point), (station pressure, sea level, altitude setting), (wind speed, gusts), precipitation, There are several clusters, such as wind direction and snow depth.

Additionally, the control unit 150 considers general applicability among these parameters.

For example, wind direction may be random, and snow depth may not apply to a particular state.

Accordingly, the control unit 150 selects (or determines) precipitation, wind speed, and temperature as initial choices of meteorological parameters from the clustering results.

Additionally, it would be useful to determine the importance of meteorological parameters, i.e. which parameters may have the greatest impact on residential links.

Additionally, by observing the regional dependence of important characteristics, the control unit 150 can confirm that location-specific insurance policies are required in the insurance product.

Additionally, the control unit 150 uses a method similar to feature ablation.

That is, the control unit 150 sets one meteorological parameter to zero in order to eliminate random information flow with respect to the parameters for the trained neural network and observes the change in accuracy.

Additionally, the control unit 150 additionally trains two additional DSPPs indicated as DSPP1 and DSPP2 in addition to the deep sigma point process (DSPP). Here, DSPP1 is trained with the second most important parameters (e.g. precipitation) removed, and DSPP2 is trained with the second and third most important parameters (e.g. precipitation, temperature) removed.

Additionally, DSPP1 and DSPP2 are used to further analyze the influence of the meteorological parameters.

Additionally, the control unit 150 may update previously selected parameters (including, for example, wind speed, precipitation, temperature, etc.) according to changes in accuracy.

Additionally, the control unit 150 uses two accuracy measures: Mean Absolute Percentage Error (MAPE) and R-squared value.

Additionally, the control unit 150 uses the MAPE to determine the importance of parameters. Here, the MAPE is calculated using the actual value and the predicted value, and the coefficient of determination (R ² ) is a regression line obtained through a linear regression model, indicating the goodness of fit of the regression equation between 0 and 1, and the closer it is to 1, the more suitable it is. indicates that Additionally, the coefficient of determination can be calculated as SSR/SST = 1 - SSE/SST. Here, the SSE (Sum of Squared Error) is the square of the difference between the estimated value (or predicted value) and the actual value, and the SSR (Sum of Squared Regression) is the squared difference between the average value of the estimated value (or predicted value) and the actual value. It is one value, and the SST (Sum of Squared Total) represents the sum of the SSE and SSR.

In this way, the control unit 150 can select (or select) a preset number of key parameters according to the type of parametric insurance.

In addition, in this way, based on the learning results of the deep sigma point process (DSPP) (or artificial neural network), the control unit 150 selects data (or damage data/parameters) that is not correlated with damage probability prediction among the collected data. ) are excluded.

That is, if too many risk factors (or parameters) are used in the DSPP, it becomes difficult to detect damage and it is difficult for the insured to understand the process, so the control unit 150 operates according to the type of parametric insurance. Perform a learning function, select key parameters for each type of parametric insurance based on the results of the learning function, and perform a learning function targeting the selected key parameters to increase overall system efficiency and determine whether a trigger has occurred with higher accuracy. You can judge.

Additionally, the control unit 150 estimates the importance of parameters based on a trained deep sigma point process.

As shown in Figure 4, this confirms that each state has its own unique meteorological parameters that are most effective for dropout rates.

That is, as shown in FIG. 4, it was confirmed that each state differed in terms of whether wind speed, precipitation, or temperature were the strongest predictors of internet connectivity dropout.

Additionally, as shown in Figure 5, in terms of modeling accuracy, the deep sigma point process demonstrates better performance than conventional Bayesian regression.

That is, the control unit 150 calculates the MAPE and the coefficient of determination using 1000 test examples randomly selected for each state.

In addition, it can be seen that the change in accuracy varies depending on the state, as shown in [Table 1] and [Table 2].

State DSPP DSPP1 DSPP2 Bayesian a1 a2 a3 A.K. 0.122 0.177 0.273 0.254 W P T AL 0.04 0.26 0.312 0.132 P T W AR 0.033 0.177 0.208 0.108 P T W A-Z 0.034 0.074 0.202 0.098 T W P CA 0.06 0.092 0.186 0.163 W P T C.O. 0.042 0.106 0.11 0.108 T W P CT 0.062 0.203 0.231 0.173 W T P D.C. 0.043 0.106 0.217 0.09 T W P D.E. 0.06 0.206 0.359 0.193 T W P FL 0.06 0.137 0.164 0.155 P T W GA 0.042 0.089 0.118 0.107 T P W HI 0.073 0.291 0.348 0.208 W T P IA 0.041 0.071 0.174 0.125 T P W ID 0.042 0.12 0.223 0.108 T W P IL 0.047 0.261 0.271 0.134 P T W IN 0.05 0.088 0.294 0.136 T P W K.S. 0.06 0.24 0.24 0.193 T W P KY 0.04 0.182 0.184 0.111 P T W L.A. 0.043 0.329 0.426 0.169 P T W M.A. 0.069 0.12 0.188 0.179 T P W M.D. 0.052 0.127 0.214 0.13 T W P M.E. 0.049 0.131 0.176 0.14 P T W MI 0.052 0.223 0.226 0.171 P T W M.N. 0.039 0.172 0.184 0.107 P T W M.O. 0.045 0.172 0.193 0.125 P T W

where a1 represents the first important parameter, a2 represents the second most important parameter, a3 represents the third most important parameter, W represents the wind speed, P represents the precipitation, and T represents temperature. At this time, the first important parameter, the second important parameter, and the third important parameter can be set differently for each state.

State DSPP DSPP1 DSPP2 Bayesian a1 a2 a3 M.S. 0.032 0.05 0.297 0.084 T P W MT 0.046 0.255 0.255 0.12 T W P NC 0.055 0.101 0.274 0.137 T P W N.D. 0.053 0.143 0.14 0.155 T W P NE 0.038 0.116 0.113 0.113 T W P NH 0.068 0.104 0.222 0.224 T P W NJ 0.061 0.128 0.211 0.157 T W P NM 0.04 0.084 0.179 0.09 T W P NV 0.061 0.205 0.204 0.186 W T P NY 0.057 0.176 0.211 0.127 P T W OH 0.061 0.265 0.265 0.15 P T W OK 0.043 0.174 0.202 0.134 P T W OR 0.063 0.135 0.179 0.126 P T W PA 0.058 0.162 0.173 0.153 P T W R.I. 0.077 0.255 0.36 0.244 T W P S.C. 0.048 0.095 0.139 0.151 T P W SD 0.061 0.19 0.178 0.186 T W P TN 0.033 0.246 0.244 0.111 P T W TX 0.044 0.181 0.192 0.122 P T W UT 0.037 0.161 0.163 0.104 W T P VA 0.047 0.198 0.215 0.131 P T W VT 0.083 0.194 0.267 0.219 T W P WA 0.062 0.103 0.13 0.127 W P T WI 0.055 0.089 0.21 0.133 T P W W.V. 0.046 0.095 0.244 0.112 T P W WY 0.036 0.114 0.175 0.108 T W P

This is because the extra nonlinearity considered in the deep sigma point process (DSPP) compared to the Bayesian regression model has different effects in various states.

Overall, the distribution of dropout probabilities is state dependent, which means regional dependencies must be taken into account.

For a more detailed analysis, Figure 6 shows two-dimensional (temperature and wind speed) views of dropout probabilities estimated from DSPP and Bayesian regression for California and Texas.

In particular, the temperature and wind speed with the highest dropout probability are indicated by circles.

Qualitatively, DSPP's estimates show strong agreement with the ground truth, with a high dropout probability for the weather conditions in the circular region.

However, classic Bayesian regression does not approach high probabilities well in the circular region and shows poor performance in capturing some extreme events.

Overall, DSPP according to embodiments of the present invention has more accurate risk modeling than classical Bayesian regression.

Figure 7 is a diagram showing a three-dimensional (z-axis represents dropout probability) view with uncertainty estimations to show a clear distinction between performance for the DSPP and the Bayesian regression.

Figure 8 is a diagram comparing DSPP, DSPP1, and DSPP2 estimates for California and Texas.

As shown in FIG. 8, regions showing the highest dropout probability are indicated by circles.

Additionally, as shown in FIG. 8, in the California example, it can be seen that removal of the second most important parameter (e.g., precipitation) does not have a significant impact on performance.

This is because information about the second most important parameter is appropriately preserved in the other two parameters, whereas direct relationships between parameters may require further analysis.

On the other hand, as shown in Figure 8 above, in the Texas example, removing the second most important parameter results in a significant change in estimation quality.

This degradation in estimation quality means that the impact of the third most important parameter (e.g. temperature) in Texas is relatively small compared to that in California.

Through this, we can suggest the importance of artificial intelligence-based data analysis that can evaluate various data due to geographical differences.

Figures 9 and 10 show three-dimensional views with uncertainty estimates in California and Texas.

As shown in FIGS. 9 and 10, different behaviors according to omitting parameters can be confirmed for each state.

Here, the color and z-axis represent the dropout probability.

For example, as shown in FIGS. 9 and 10 above, to emphasize the difference between estimates, points with high z values are colored red, and points with low z values are colored blue.

Additionally, from this perspective, black errorbars are drawn in randomly selected regions, which correspond to the uncertainty of the estimation.

In addition, as shown in FIGS. 9 and 10, it can be clearly seen that the deep sigma point process (DSPP) according to an embodiment of the present invention can best approximate the ground truth probability. there is.

Additionally, in some cases, such as classical Bayesian regression and DSPP2, high uncertainty values can be seen almost everywhere, meaning that uncertainty estimation is not successful.

The quantitative results of dropout probability estimation are summarized in [Table 1] and [Table 2] above, and represent MAPE from 1000 randomly selected test data with respect to classic Bayesian regression, DSPP, DSPP1, and DSPP2.

Overall, the estimates of DSPP show a MAPE of about 5%, which is the best result among other models.

Additionally, as shown in FIG. 4, regional dependence on the importance of parameters can be confirmed.

Additionally, ignoring the second and third most important parameters has a noticeable impact on the estimation accuracy.

As shown in Figure 11, for greater visibility, Figure 11 includes estimated quality maps for DSPP1 and DSPP2.

These results indicate that DSPP according to an embodiment of the present invention can predict risk with higher accuracy than existing models, which provides more accurate coverage than existing insurance can achieve. This indicates that it can be applied to the data analysis part of parametric insurance.

In addition to better estimation quality, the deep sigma point process (DSPP) according to embodiments of the present invention can provide uncertainty in the estimation, which is robust against overfitting as well as insurance design. It is related to

For example, very uncertain means that the amount of training data on which the model's estimates are based is very insufficient.

On the other hand, high uncertainty can complicate the relationship between risk and parameters.

Although there is no rule-of-thumb for quantifying the quality of uncertainty estimates, embodiments of the present invention suggest several qualitative ways.

For example, almost everywhere models with high uncertainty do not provide useful information.

Additionally, it can be expected that the model should exhibit high uncertainty in meteorological conditions with a small number of measurements.

From this perspective, as shown in FIGS. 7, 9, and 10, it can be confirmed that DSPP according to an embodiment of the present invention is better for uncertainty estimation compared to other methods.

Additionally, the Bayesian regression model and DSPP without some parameters exhibit high uncertainty at various measurement points.

In contrast, the uncertainty value of the DSPP model tends to be proportional to the density of measurement points, following qualitative expectations.

Additionally, this information about uncertainty can be crucial for the insurance industry.

Higher uncertainty suggests that there is a greater likelihood that actual phenomena will deviate from the model's expectations, and customizing the current model may require a cautious, conservative policy or additional data about those expectations.

The parametric insurance model according to an embodiment of the present invention consists of data analysis and risk estimation procedures as well as trigger and payout considerations.

In other words, the amount of risk expected is highly correlated with the probability of dropout.

In addition, considering the maximum amount of reserves (or reserves) and the number of policyholders, the parametric insurance model (or the deep sigma point process (DSPP)) according to embodiments of the present invention can be used to trigger payments. You can propose (or devise/create/construct/set) a threshold for the dropout probability to be considered.

Additionally, the dropout probability estimated from the DSPP may be clustered in some areas.

Additionally, regions where the minimum dropout probability is greater than some threshold provide important information about triggers.

That is, if all weather parameters are included in the area, releasing payout can be considered.

In this way, the control unit 150 performs machine learning using one or more parameters included in the collected damage data (or raw data) as input, using the deep sigma point process (DSPP) according to an embodiment of the present invention. And, based on the results of machine learning, a preset number of key parameters (or standard damage data/risk parameters) that serve as trigger criteria in relation to a specific parametric insurance are selected (or set/selected/extracted), and selected. Using the key parameters provided, the deep sigma point process can be trained with respect to the parametric insurance.

Parametric insurance described in embodiments of the present invention is attracting attention because it has advantages over existing insurance products in some aspects, such as a quick and easy claim procedure.

However, one of the main problems in parametric insurance is the problem of estimating basis risk.

In an embodiment of the present invention, the potential use of Bayesian neural networks to reduce the underlying risk is explored.

That is, embodiments of the present invention use data on dropouts from residential links in the United States that are affected by weather conditions.

Additionally, in an embodiment of the present invention, by using the deep sigma point process (DSPP), the dropout probability can be estimated with a much smaller amount of error compared to the existing Bayesian regression model.

In addition, it can be seen that there is a clear regional dependence on the importance of weather parameters, which means that geographical differences must be considered when designing parametric insurance.

As such, the deep sigma point process (DSPP) according to embodiments of the present invention not only develops a data-based parametric insurance model for Internet connectivity dropouts in the United States, but also provides various insurance policies that conventional insurance cannot adequately cover. It is expected that it can be applied to develop a parametric insurance model for risk.

Additionally, the control unit 150 performs trigger design for each type of parametric insurance.

That is, the control unit 150 sets (or selects) a threshold value for each parametric insurance type according to the amount of insurance money each insurance provider wishes to provide and the expected scale of damage for key parameters that generate risk for each type of parametric insurance. /configure/create/suggest).

Predicting risk probability in this way must include steps such as reducing the influence of noise in the data, performing mathematical modeling to predict risk probability, and selecting appropriate risk factors (or risk parameters).

In the embodiment of the present invention, the processes necessary for predicting risk probability are presented, and how each process can be successfully performed can be confirmed based on actual data.

In addition, in the embodiment of the present invention, the DSPP is mainly explained as an example of an artificial intelligence neural network (or a pre-trained machine learning model), but it is not limited thereto, and other artificial intelligence neural networks (or other machine learning models) in addition to the DSPP This can be used to measure risk factors related to a specific parametric insurance and select triggers based on the measured risk factors.

Additionally, the control unit 150 controls the collection unit 110 to collect damage data from the information provision server.

Additionally, the control unit 150 performs preprocessing (or preprocessing function) on the collected damage data.

That is, the control unit 150 performs binning included in a preset Monte Carlo sampling method on the collected damage data to remove noise included in the damage data. Here, the binning refers to the process of collecting the average of several cases with similar conditions to the collected damage data and reducing the influence of noise contained in individual cases.

Additionally, the control unit 150 extracts (or classifies) the collected damage data (or damage data from which the noise has been removed) into individual damage data. At this time, if the collected damage data (or damage data from which the noise has been removed) consists of a plurality of damage data, the control unit 150 may combine (or extract/classify) at least two damage data among the plurality of damage data. there is.

In addition, the control unit 150 performs artificial intelligence-based machine learning based on the preprocessed damage data, etc., and determines the network (or Generate (or calculate/calculate) the probability of connectivity dropout for the network. Here, the connectivity drop-out (or connectivity drop-out for the network/network) is a network access server (or Includes information on the presence or absence of connectivity (or presence or absence of connection/connection) between network servers (not shown) (including, for example, Internet connection, Internet dropout/disconnection, etc.).

That is, the control unit 150 performs machine learning (or artificial intelligence/deep learning) by using the preprocessed damage data as an input value of the pre-trained (or set) artificial intelligence neural network (or pre-trained machine learning model). perform, and generate (or calculate/calculate).

At this time, the control unit 150 performs the artificial intelligence-based machine learning for each pre-processed individual damage data (or for each combination of at least two pieces of pre-processed damage data), and determines the individual damage based on the machine learning results. The connectivity dropout probability for the network (or the connectivity dropout probability for the network for each individual damage data) may be generated for each data (or for each combination of at least two pieces of damage data).

Here, the control unit 150 uses the preprocessed damage data as an input value of the deep sigma point process (or the Bayesian neural network) included in the pre-trained (or set) artificial intelligence neural network to perform machine learning (or artificial intelligence). /deep learning) and generate connectivity dropout status information for the network (or networks) related to the damage data (or the preprocessed damage data), which is a machine learning result (or artificial intelligence result/deep learning result). It may be possible. Here, the connectivity dropout status information includes presence or absence of connectivity dropout, connectivity dropout maintenance time, etc.

In addition, the control unit 150 determines the connectivity dropout probability for the network (or network) in relation to the generated (or calculated) corresponding damage data (or the preprocessed damage data) as a main trigger for parametric insurance. (or major trigger related to specific parametric insurance) to determine (or confirm/determine) whether it can be utilized.

That is, the control unit 150 determines the connectivity dropout probability for the network (or network) in relation to the generated (or calculated) damage data (or the preprocessed damage data) in advance in response to a specific parametric insurance. Determine (or confirm) whether the set threshold is exceeded. Here, the threshold is set differently for each parameter (or for each risk factor) depending on the type of parametric insurance, and if a preset trigger (or trigger condition) exists in relation to a specific parametric insurance, the corresponding trigger (or It may be set (or updated) based on the standard damage data included in the trigger condition.

At this time, the control unit 150 determines the damage probability for each damage data generated for each individual damage data (or for each combination of at least two pieces of preprocessed damage data) (or the damage probability for each combination of damage data/network for each individual damage data). It may also be determined whether the connectivity dropout probability for) exceeds the threshold (or the threshold corresponding to the specific parametric insurance).

Here, the control unit 150 detects the connectivity dropout status information for the network (or network) related to the generated damage data (or the preprocessed damage data) at a preset time (e.g., related to the specific parametric insurance). For example, it may be determined (or confirmed) whether or not it is maintained for more than a preset time (for example, 2 hours or more) within 24 hours.

As a result of the determination (or the confirmation result), if the damage data (or the pre-processed damage data) cannot be used as a main trigger for the specific parametric insurance, the control unit 150 previously related to the damage data. Information indicating that it cannot be used as a main trigger for the specific parametric insurance is output through the display unit 130 and/or the audio output unit 140.

That is, if the determination result (or the confirmation result), the connectivity dropout probability for the corresponding network (or network) does not exceed a preset threshold corresponding to the specific parametric insurance, the control unit 150 Regarding the damage data, information indicating that it cannot be used as a main trigger for the specific parametric insurance is output through the display unit 130 and/or the audio output unit 140.

At this time, the threshold is determined among the damage probability for each generated individual damage data (or damage probability for each combination of damage data/connectivity dropout probability for the network for each individual damage data). If there is at least one individual damage data that does not exceed the value, the control unit 150 provides information indicating that the at least one individual damage data cannot be used as a main trigger for the specific parametric insurance. It is output through the display unit 130 and/or the audio output unit 140.

Here, the connectivity dropout status information for the network (or network) related to the generated damage data (or the preprocessed damage data) is provided within a preset time (for example, 24 hours) in relation to the specific parametric insurance. If it is not maintained for more than a preset time (for example, more than 2 hours), the control unit 150 displays information indicating that the damage data cannot be used as a main trigger for the specific parametric insurance. It may also be output through 130 and/or the audio output unit 140.

In addition, the control unit 150 maps (or matches/links) information indicating that the damage data cannot be used as a main trigger for the specific parametric insurance to the storage unit 120. ) to save it.

Additionally, the control unit 150 returns to the previous process of collecting damage data and controls the collection unit 110 to perform a process of collecting new damage data.

In addition, if the judgment result (or the confirmation result) is that the damage data (or the pre-processed damage data) can be used as a main trigger for the specific parametric insurance, the control unit 150 may In relation to , information indicating that it can be used as a main trigger for the specific parametric insurance is output through the display unit 130 and/or the audio output unit 140.

Additionally, the control unit 150 adds conditions related to the corresponding damage data (or risk factor/risk factor) to a preset trigger (or trigger condition) in response to the specific parametric insurance. Here, the trigger (or trigger condition/content/parameter for determining the trigger condition) consists of different parameters (or risk factors) depending on the type of parametric insurance, and each parameter (or risk factor) is different. Other thresholds may be set.

For example, in the case of the cyber insurance, it includes a first threshold according to wind speed, precipitation, temperature, etc., and in the case of the storm and water disaster insurance, it includes a second threshold according to wind speed, precipitation, temperature, etc., and the earthquake insurance includes a second threshold according to the wind speed, precipitation, temperature, etc. Includes a third threshold depending on the case progress, etc.

That is, if the determination result (or the confirmation result), the connectivity dropout probability for the corresponding network (or network) is greater than or equal to a preset threshold corresponding to the specific parametric insurance, the control unit 150 In response to a specific parametric insurance, conditions related to the corresponding damage data are added to the preset trigger.

At this time, the threshold is determined among the damage probability for each generated individual damage data (or damage probability for each combination of damage data/connectivity dropout probability for the network for each individual damage data). If there is at least one other individual damage data exceeding the value, the control unit 150 sets a condition related to at least one other individual damage data (e.g., corresponding to a preset trigger corresponding to the specific parametric insurance). (including connectivity dropout probability for each individual damage data) is added.

Here, the connectivity dropout status information for the network (or network) related to the generated damage data (or the preprocessed damage data) is provided within a preset time (for example, 24 hours) in relation to the specific parametric insurance. If it is maintained for more than a preset time (for example, more than 2 hours), the control unit 150 may add a condition related to the damage data to a preset trigger in response to the specific parametric insurance.

In this way, the control unit 150 measures risk factors through artificial intelligence (or machine learning) in relation to a specific parametric insurance, and selects (or sets) a trigger that can be used for the specific parametric insurance. .

In addition, the control unit 150 calculates an event occurrence probability density function distribution related to the corresponding damage data (or corresponding individual damage data), and outputs (or displays) the event occurrence probability density function distribution related to the calculated corresponding damage data. do.

For example, all events have a probability of occurring under certain conditions, and in the case of an earthquake, when 'seismic intensity' and 'distance to the epicenter' are risk factors, the 'probability of damage due to an earthquake' is the probability of occurrence of the event. If it is a density function, if the seismic intensity is 9 and the distance to the epicenter is 100km, the probability of damage is 95%, if the seismic intensity is 3 and the distance to the epicenter is 100km, the damage probability is 4%, and if the seismic intensity is 3 and the distance to the epicenter is 1km, the damage probability is 4%. If the probability is 80% and the distance to the epicenter is 10,000 km even if the seismic intensity is 9, the probability of damage is almost 0%, so it can have (or be constructed) a distribution by the event occurrence probability density function.

In addition, if the determination result (or the confirmation result), the connectivity dropout probability for the corresponding network (or network) is greater than or equal to a preset threshold corresponding to the specific parametric insurance, the control unit 150 It is determined that the damage data meets the insurance payment conditions, and information related to the damage data can be provided to the insurance company's server (not shown).

In addition, the insurance company server calculates the insurance amount based on information related to the damage data provided from the parametric insurance providing device 100 and information on the terms and conditions of the parametric insurance related to the damage data. Here, the insurance company server may calculate the insurance amount by considering the amount of insurance money and the expected amount of damage that each insurance provider wishes to provide in relation to the type of parametric insurance, together with the information on the terms and conditions of the parametric insurance related to the damage data. there is.

At this time, the connectivity dropout status information for the network (or network) related to the generated corresponding damage data (or the preprocessed damage data) is a preset trigger condition, within a preset time (for example, 24 hours). If it is maintained for more than an hour (for example, more than 2 hours), the insurance company server may calculate the insurance amount based on information on the terms and conditions of the parametric insurance related to the damage data.

In addition, the insurance company server links with a database (not shown) (or the storage unit 120) to extract (or confirm) insurance subscriber information corresponding to at least one user who has subscribed to parametric insurance related to the damage data. )do. Here, the insured subscriber information includes basic information (e.g., contract number, contract status, insurance period, etc.), subscriber information (e.g., policyholder, insured person, mobile phone number, landline phone number, mail reception method, mail receipt address, etc.) , name of bank receiving insurance money, bank account information for receiving insurance money, etc.), beneficiary information (including death beneficiary, other beneficiaries, maturity beneficiary, etc.), contract manager (including business/department name, name, contact information, etc.), etc. Includes. At this time, in order to pay insurance money to users who have subscribed to insurance in the area related to the damage data, the insurance company server is used to determine if the damage data has occurred (or been collected) among a plurality of users who have subscribed to parametric insurance related to the damage data. ) At least one user related to the region may be identified, and insurance subscriber information corresponding to the at least one confirmed user may be extracted.

In addition, the insurance company server is linked with a bank server (not shown), and a bank account included in the insurance subscriber information corresponding to at least one user who has subscribed to parametric insurance related to the extracted (or confirmed) corresponding damage data. The calculated (or confirmed) insurance money is automatically deposited (or provided) to (e.g., insurance money receiving bank name/insurance money receiving bank account information). At this time, the insurance company server may deposit different insurance money depending on the insurance subscription period for each policyholder.

In an embodiment of the present invention, the wind speed, precipitation, temperature, etc. that are optimal for the deep sigma point process included in the artificial neural network for cyber insurance among parametric insurance from the collected weather information (or meteorological information), etc. are calculated from the corresponding reference damage data ( or risk parameter/trigger/trigger condition), but is not limited to this, and the control unit 150 performs the in-depth analysis on damage data (or raw data) collected in relation to a specific parametric insurance. Machine learning is performed using the sigma point process, and damage data with a high correlation with the deep sigma point process are used (or extracted/ settings) can be set.

In this way, the level of risk caused by a specific disaster can be accurately predicted using artificial intelligence based on accumulated data.

Hereinafter, the method for providing data-based parametric insurance according to the present invention will be described in detail with reference to FIGS. 1 to 12.

Figure 12 is a flowchart showing a method of providing data-based parametric insurance according to an embodiment of the present invention.

First, the collection unit 110 collects damage data from an information provision server (not shown). Here, the information provision server may be a server that collects various information and shares (or provides) the collected information for a fee or free of charge. In addition, the damage data is set according to the type of parametric insurance. In the case of cyber insurance, it includes personal information leakage, hacking, weather information, location of occurrence, date of occurrence, etc., and in the case of storm and water disaster insurance, weather information. , location of occurrence, date of occurrence, etc., and in the case of earthquake insurance, it includes seismic intensity, location of occurrence, date of occurrence, etc. Here, the weather information includes wind speed, precipitation, temperature, cloud level, visibility/visibility, altimeter settings, sea level, station pressure, dew point, wind direction, gusts, snowfall, etc.

As an example, the first collection unit 110 receives first weather information (e.g., first wind speed, first precipitation, first temperature, first wind speed, first precipitation, first temperature, Includes first cloud level, first visibility, first altimeter setting, first sea level, first station pressure, first dew point, first wind direction, first wind gust, first snow amount, etc.), first occurrence location (e.g. Florida), first damage data including the first occurrence date (e.g., September 28, 2022), etc. are collected.

As another example, the second collection unit 110 receives the second seismic intensity (including, for example, magnitude 2.1, etc.) and the second occurrence location according to the earthquake damage that occurred near Geomundo Island, Yeosu-si, Jeollanam-do in relation to earthquake insurance from the information provision server ( For example, secondary damage data including the 92km sea area southeast of Geomundo, Yeosu-si, Jeollanam-do) and the second occurrence date (e.g., November 8, 2022) are collected (S1210).

Afterwards, the control unit 150 performs preprocessing (or preprocessing function) on the collected damage data.

That is, the control unit 150 performs binning included in a preset Monte Carlo sampling method on the collected damage data to remove noise included in the damage data. Here, the binning refers to the process of collecting the average of several cases with similar conditions to the collected damage data and reducing the influence of noise contained in individual cases.

Additionally, the control unit 150 extracts (or classifies) the collected damage data (or damage data from which the noise has been removed) into individual damage data. At this time, if the collected damage data (or damage data from which the noise has been removed) consists of a plurality of damage data, the control unit 150 may combine (or extract/classify) at least two damage data among the plurality of damage data. there is.

As an example, the first control unit 150 removes noise from the collected first damage data through the segmentation, and selects the first wind speed included in the first weather information from the first damage data from which the noise has been removed. , extract first precipitation, first temperature, first cloud height, first visibility, first altimeter setting, first sea level, first station pressure, first dew point, first wind direction, first wind gust, first snow amount, etc., respectively. do.

As another example, noise is removed from the collected second damage data through the segmentation, and a second magnitude (including, for example, magnitude 2.1, etc.) is extracted from the second damage data from which the noise has been removed ( S1220).

Thereafter, the control unit 150 performs artificial intelligence-based machine learning based on the preprocessed damage data, etc., and determines the network (or Generate (or calculate/calculate) the probability of connectivity dropout for the network. Here, the connectivity drop-out (or connectivity drop-out for the network/network) is a network access server (or Includes information on the presence or absence of connectivity (or presence or absence of connection/connection) between network servers (not shown) (including, for example, Internet connection, Internet dropout/disconnection, etc.).

That is, the control unit 150 performs machine learning (or artificial intelligence/deep learning) by using the preprocessed damage data as an input value of the pre-trained (or set) artificial intelligence neural network (or pre-trained machine learning model). perform, and generate (or calculate/calculate).

At this time, the control unit 150 performs the artificial intelligence-based machine learning for each pre-processed individual damage data (or for each combination of at least two pieces of pre-processed damage data), and determines the individual damage based on the machine learning results. The connectivity dropout probability for the network (or the connectivity dropout probability for the network for each individual damage data) may be generated for each data (or for each combination of at least two pieces of damage data).

Here, the control unit 150 uses the preprocessed damage data as an input value of the deep sigma point process (or the Bayesian neural network) included in the pre-trained (or set) artificial intelligence neural network to perform machine learning (or artificial intelligence). /deep learning) and generate connectivity dropout status information for the network (or networks) related to the damage data (or the preprocessed damage data), which is a machine learning result (or artificial intelligence result/deep learning result). It may be possible. Here, the connectivity dropout status information includes presence or absence of connectivity dropout, connectivity dropout maintenance time, etc.

As an example, the first control unit may control the preprocessed first wind speed, first precipitation, first temperature, first cloud height, first visibility, first altimeter setting, first sea level, first station pressure, first dew point, first Machine learning is performed using the first wind direction, first gust, first snowfall, etc. as input values of the first deep sigma point process included in the artificial intelligence neural network, and the first machine learning result is each machine learning result. Based on the through twelfth machine learning results, the first through twelfth connectivity dropout probabilities are generated for each input value, respectively.

As another example, the second control unit performs machine learning by using the preprocessed second progress (for example, including magnitude 2.1, etc.) as a second deep sigma point process input value included in the artificial intelligence neural network, and 21 Based on the machine learning results, a 21st connectivity dropout probability (e.g., 8%) related to the corresponding second damage data is generated (S1230).

Thereafter, the control unit 150 determines the connectivity dropout probability for the network (or network) in relation to the generated (or calculated) corresponding damage data (or the preprocessed damage data) as a main trigger for parametric insurance. (or major trigger related to specific parametric insurance) to determine (or confirm/determine) whether it can be utilized.

That is, the control unit 150 determines the connectivity dropout probability for the network (or network) in relation to the generated (or calculated) damage data (or the preprocessed damage data) in advance in response to a specific parametric insurance. Determine (or confirm) whether the set threshold is exceeded. Here, the threshold is set differently for each parameter (or for each risk factor) depending on the type of parametric insurance, and if a preset trigger (or trigger condition) exists in relation to a specific parametric insurance, the corresponding trigger (or It may be set (or updated) based on the standard damage data included in the trigger condition.

At this time, the control unit 150 determines the damage probability for each damage data generated for each individual damage data (or for each combination of at least two pieces of preprocessed damage data) (or the damage probability for each combination of damage data/network for each individual damage data). It may also be determined whether the connectivity dropout probability for) exceeds the threshold (or the threshold corresponding to the specific parametric insurance).

Here, the control unit 150 detects the connectivity dropout status information for the network (or network) related to the generated damage data (or the preprocessed damage data) at a preset time (e.g., related to the specific parametric insurance). For example, it may be determined (or confirmed) whether or not it is maintained for more than a preset time (for example, 2 hours or more) within 24 hours.

As an example, the first control unit determines that the first connectivity dropout probability (e.g., 95%) related to the first wind speed, which is the first to twelfth machine learning result, is preset in relation to the corresponding cyber insurance. 1 determine whether a threshold (e.g. 70%) is exceeded, and a second connectivity dropout probability (e.g. 92%) associated with the first precipitation is greater than or equal to the first threshold (e.g. 70%). and determine whether a third connectivity dropout probability (e.g., 93%) related to the first temperature exceeds the first threshold (e.g., 70%), Determine whether a fourth connectivity dropout probability (e.g., 15%) associated with the first visibility exceeds the first threshold (e.g., 70%), and determine whether a fifth connectivity dropout probability associated with the first visibility Determine whether a connectivity dropout probability (e.g., 45%) exceeds the first threshold (e.g., 70%), and determine a sixth connectivity dropout probability (e.g., 45%) relative to the first altimeter setting. determine whether the probability (e.g., 23%) exceeds the first threshold (e.g., 70%), and determine whether a seventh connectivity dropout probability (e.g., 35%) associated with the first sea level is Determine whether a first threshold (e.g. 70%) is exceeded, and determine whether an eighth connectivity dropout probability (e.g. 43%) associated with the first station pressure exceeds the first threshold (e.g. For example, 70%), and whether the ninth connectivity dropout probability (for example, 49%) associated with the first dew point exceeds the first threshold (for example, 70%). Determine whether a 10th connectivity dropout probability (e.g., 49%) associated with the first wind direction exceeds the first threshold (e.g., 70%), and determine whether the first gust Determine whether an 11th connectivity dropout probability (e.g., 49%) associated with exceeds the first threshold (e.g., 70%), and determine whether a 12th connectivity drop associated with the first snow amount It is respectively determined whether the out probability (eg, 25%) exceeds the first threshold (eg, 70%).

As another example, the second control unit determines that the 21st connectivity dropout probability (e.g., 8%) related to the second damage data, which is the second machine learning result, is set to a preset third threshold related to the corresponding earthquake insurance ( For example, determine whether it exceeds 75% (S1240).

As a result of the determination (or the confirmation result), if the damage data (or the pre-processed damage data) cannot be used as a main trigger for the specific parametric insurance, the control unit 150 previously related to the damage data. Information indicating that it cannot be used as a main trigger for the specific parametric insurance is output through the display unit 130 and/or the audio output unit 140.

That is, if the determination result (or the confirmation result), the connectivity dropout probability for the corresponding network (or network) does not exceed a preset threshold corresponding to the specific parametric insurance, the control unit 150 Regarding the damage data, information indicating that it cannot be used as a main trigger for the specific parametric insurance is output through the display unit 130 and/or the audio output unit 140.

At this time, the threshold is determined among the damage probability for each generated individual damage data (or damage probability for each combination of damage data/connectivity dropout probability for the network for each individual damage data). If there is at least one individual damage data that does not exceed the value, the control unit 150 provides information indicating that the at least one individual damage data cannot be used as a main trigger for the specific parametric insurance. It is output through the display unit 130 and/or the audio output unit 140.

Here, the connectivity dropout status information for the network (or network) related to the generated damage data (or the preprocessed damage data) is provided within a preset time (for example, 24 hours) in relation to the specific parametric insurance. If it is not maintained for more than a preset time (for example, more than 2 hours), the control unit 150 displays information indicating that the damage data cannot be used as a main trigger for the specific parametric insurance. It may also be output through 130 and/or the audio output unit 140.

In addition, the control unit 150 maps (or matches/links) information indicating that the damage data cannot be used as a main trigger for the specific parametric insurance to the storage unit 120. ) to save it.

Additionally, the control unit 150 returns to the previous process of collecting damage data and controls the collection unit 110 to perform a process of collecting new damage data.

As an example, the first control unit may configure a first cloud level, a first visibility, a first altimeter setting, a first sea level, and a first station pressure related to a dropout probability of not exceeding the first threshold (e.g., 70%). , the first dew point, the first wind direction, the first gust, the first snow amount, etc., display information indicating that they cannot be used as the main trigger of the cyber insurance through the first display unit 130 and the first audio output unit 140. output through

In addition, the first control unit controls the cyber system in relation to the first cloud height, first visibility, first altimeter setting, first sea level, first station pressure, first dew point, first wind direction, first gust, first snow amount, etc. Information indicating that it cannot be used as a major trigger for insurance is mapped with the corresponding data and stored in the first storage unit 120.

As another example, the second connectivity dropout probability (e.g., 8%) related to the second damage data, which is the second machine learning result, is set to a preset third threshold (e.g., 75%) related to the corresponding earthquake insurance. ), the second control unit outputs information indicating that a trigger has not occurred in relation to the second damage data through the second display unit 130 and the second audio output unit 140.

Additionally, the second control unit maps information indicating that a trigger has not occurred in relation to the second damage data with the second damage data and stores it in the second storage unit 120 (S1250).

In addition, if the judgment result (or the confirmation result) is that the damage data (or the pre-processed damage data) can be used as a main trigger for the specific parametric insurance, the control unit 150 may In relation to , information indicating that it can be used as a main trigger for the specific parametric insurance is output through the display unit 130 and/or the audio output unit 140.

Additionally, the control unit 150 adds conditions related to the damage data to a preset trigger (or trigger condition) in response to the specific parametric insurance. Here, the trigger (or trigger condition/content/parameter for determining the trigger condition) consists of different parameters (or risk factors) depending on the type of parametric insurance, and each parameter (or risk factor) is different. Other thresholds may be set.

That is, if the determination result (or the confirmation result), the connectivity dropout probability for the corresponding network (or network) is greater than or equal to a preset threshold corresponding to the specific parametric insurance, the control unit 150 In response to a specific parametric insurance, conditions related to the corresponding damage data are added to the preset trigger.

At this time, the threshold is determined among the damage probability for each generated individual damage data (or damage probability for each combination of damage data/connectivity dropout probability for the network for each individual damage data). If there is at least one other individual damage data exceeding the value, the control unit 150 sets a condition related to at least one other individual damage data (e.g., corresponding to a preset trigger corresponding to the specific parametric insurance). (including connectivity dropout probability for each individual damage data) is added.

Here, the connectivity dropout status information for the network (or network) related to the generated damage data (or the preprocessed damage data) is provided within a preset time (for example, 24 hours) in relation to the specific parametric insurance. If it is maintained for more than a preset time (for example, more than 2 hours), the control unit 150 may add a condition related to the damage data to a preset trigger in response to the specific parametric insurance.

In this way, the control unit 150 measures risk factors through artificial intelligence (or machine learning) in relation to a specific parametric insurance, and selects (or sets) a trigger that can be used for the specific parametric insurance. .

In addition, the control unit 150 calculates an event occurrence probability density function distribution related to the corresponding damage data (or corresponding individual damage data), and outputs (or displays) the event occurrence probability density function distribution related to the calculated corresponding damage data. do.

As an example, the first control unit may be used as a main trigger in the case of the first wind speed, first precipitation amount, first temperature, etc. related to the dropout probability exceeding the first threshold (e.g. 70%). The indicated information is output through the first display unit and the first audio output unit.

Additionally, the first control unit includes the first wind speed, the first precipitation amount, and the first temperature in a preset trigger of cyber insurance.

In addition, the first control unit calculates a first event occurrence probability density function distribution related to the first wind speed, calculates a second event occurrence probability density function distribution related to the first precipitation, and calculates a first event occurrence probability density function distribution related to the first temperature. 3 An event occurrence probability density function distribution is calculated, and the calculated first event occurrence probability density function distribution to a third event occurrence probability density function distribution are output through the first display unit and the first audio output unit (S1260).

As described above, the embodiment of the present invention provides a framework for insurance providers who wish to introduce index-type insurance by accurately predicting the degree of risk due to a specific disaster using artificial intelligence based on accumulated data. can be provided, and the trigger selection process according to the degree of risk due to the specific disaster can be efficiently performed.

Anyone skilled in the art to which the present invention pertains can make modifications and changes to the above-described content without departing from the essential characteristics of the present invention. Accordingly, the embodiments disclosed in the present invention are not intended to limit the technical idea of the present invention, but are for illustrative purposes, and the scope of the technical idea of the present invention is not limited by these embodiments. The scope of protection of the present invention should be interpreted in accordance with the claims below, and all technical ideas within the equivalent scope should be construed as being included in the scope of rights of the present invention.

100: Data-based parametric insurance provision device
110: collection unit 120: storage unit
130: display unit 140: audio output unit
150: control unit

Claims (10)

A collection unit that collects damage data from the information provision server; and
Preprocessing is performed on the collected damage data, artificial intelligence-based machine learning is performed based on the preprocessed damage data, and connectivity dropout probability for the network is determined based on the machine learning results in relation to the damage data. Generate, determine whether the generated connectivity dropout probability can be used as a main trigger related to a specific parametric insurance, and as a result of the determination, determine whether the generated connectivity dropout probability can be used as a main trigger related to a specific parametric insurance. A data-based parametric insurance providing device including a control unit that adds conditions related to the damage data to a preset trigger in response to the specific parametric insurance, when possible. According to claim 1,
The damage data is,
It is set according to the type of parametric insurance. In the case of cyber insurance, it includes personal information leakage, hacking, weather information, location of occurrence, and date of occurrence, and in the case of storm and water disaster insurance, it includes weather information, location of occurrence, and date of occurrence. Includes, and in the case of earthquake insurance, includes seismic intensity, location of occurrence, and date of occurrence,
The weather information above is,
A data-driven parametric insurance providing device comprising at least one of wind direction, precipitation, temperature, cloud level, visibility/visibility, altimeter settings, sea level, station pressure, dew point, wind speed, wind gust, and snow depth. According to claim 1,
The control unit,
A data-based parametric insurance providing device, characterized in that for calculating an event occurrence probability density function distribution related to the damage data and controlling the calculated event occurrence probability density function distribution related to the damage data to be displayed on a display unit. Collecting damage data from an information provision server by a collection unit;
Preprocessing the collected damage data by a control unit;
performing artificial intelligence-based machine learning based on the preprocessed damage data, by the control unit, and generating a connectivity dropout probability for a network with respect to the damage data based on the machine learning results;
determining, by the control unit, whether the generated connectivity dropout probability can be utilized as a main trigger related to a specific parametric insurance; and
As a result of the determination, when the generated connectivity dropout probability can be used as a main trigger related to a specific parametric insurance, the control unit sets conditions related to the damage data to a preset trigger corresponding to the specific parametric insurance. A method of providing data-driven parametric insurance that includes additional steps. According to claim 4,
The step of performing preprocessing on the collected damage data is,
A process of removing noise included in the damage data by performing binning included in a preset Monte Carlo sampling method on the collected damage data; and
A method for providing data-based parametric insurance, comprising extracting preset standard damage data from the damage data from which the noise has been removed. According to claim 4,
The step of generating a connectivity dropout probability for the network in relation to the damage data is,
A process of extracting the preprocessed damage data into individual damage data; and
Data comprising the process of performing the artificial intelligence-based machine learning for each extracted individual damage data and generating a connectivity dropout probability for the network for each individual damage data based on the machine learning results. How to provide based parametric insurance. According to claim 4,
The step of determining whether the generated connectivity dropout probability can be used as a main trigger related to a specific parametric insurance is,
A method for providing data-based parametric insurance, characterized in that it determines whether the connectivity dropout probability for the network generated in relation to the damage data exceeds a preset threshold in response to a specific parametric insurance. According to claim 4,
As a result of the determination, when the generated connectivity dropout probability cannot be used as a main trigger related to a specific parametric insurance, it cannot be used by the control unit as a main trigger related to the specific parametric insurance with respect to the damage data. A method of providing data-based parametric insurance, further comprising displaying information representing through a display unit. According to claim 4,
The trigger is,
A method of providing data-based parametric insurance, characterized in that it consists of different parameters depending on the type of parametric insurance, and different threshold values are set for each parameter. According to claim 4,
calculating, by the control unit, an event occurrence probability density function distribution related to the damage data; and
A method for providing data-based parametric insurance, further comprising displaying, by the control unit, an event occurrence probability density function distribution related to the calculated damage data on a display unit.

Images