KR20230165995A - Method for Providing Artificial Intelligence Module Optimized Analysis/Prediction by Using Time Series Data - Google Patents

Abstract

The present invention relates to a method of providing an artificial intelligence module optimized for analysis/prediction using time series data. The method of providing an artificial intelligence module optimized for analysis/prediction using time series data executed through an operation server according to the present invention is an industrial S sensing data sensed in time series through S (S ≥ 1) sensors installed in fields or agricultural fields, and C captured in time series through C (C ≥ 1) cameras installed in the field. Among the image data, D(1≤D≤(S+C)) designated data are collected and preprocessed to create T(T≥1) time series models related to the time series of the process or growth performed in the field and within the allowable range. A first step of configuring N (1 ≤ N ≤ D) time series data modeled to enable matching, and matching with a specified periodicity related to at least one process or growth performed in the field based on the N time series data. A second step of confirming t (1≤t≤T) time series models with periodicity and n (1≤n≤N) time series data with periodicity that can be matched within a specified tolerance range, and the confirmed n time series data A third step of processing into frequency domain-based time series data corresponding to specified frequency characteristics, and n' (1≤n'≤ Construct a plurality of learning data containing N'(1≤N'≤N, n'⊂N') time series data containing n) pieces of time series data in a specified structure, and use the plurality of configured learning data. A fourth step of learning M (M≥2) artificial intelligence modules to which different artificial intelligence algorithms are applied, and after learning of the M artificial intelligence modules, are collected through the processes corresponding to the first to third steps, and Among n time series data preprocessed and then processed into the designated frequency domain, N' time series data, including n' time series data for designated analysis or prediction, are substituted into the learned M artificial intelligence modules for designated analysis or prediction. A fifth step of checking the result information for each of the M artificial intelligence modules, a sixth step of checking the actual measurement result information measured for the M result information of the artificial intelligence modules, and the confirmed actual measurement result information and the M number of artificial intelligence modules. A seventh step of matching and analyzing the result information for each artificial intelligence module to calculate M evaluation information in which at least one of the accuracy, precision, and recall rate for each M artificial intelligence module for the specified analysis or prediction is evaluated according to the specified evaluation rule; , and an eighth step of determining at least one m (1≤m≤M) artificial intelligence module optimized for a specified analysis or prediction based on the calculated M evaluation information.

Description

Method for Providing Artificial Intelligence Module Optimized Analysis/Prediction by Using Time Series Data}

The present invention relates to a method of providing an artificial intelligence module optimized for analysis/prediction using time series data. The present invention relates to a method of providing an artificial intelligence module optimized for analysis/prediction using time series data, by preprocessing data collected in time series through various sensors or cameras provided in a smart farm or smart factory to N (N ≥ 1) While modeling time series data, n(1≤n≤N) time series data with a series of periodic characteristics are processed into the frequency domain, and r(r≥1) known artificial intelligence algorithms (e.g. LSTM, CNN) are used. etc.), prepare M (M≥1) artificial intelligence modules applied to specific purposes/uses by combining one or more of the following, and n' (1≤n'≤n) time series data processed into the frequency domain. After learning N'(1≤N'≤N, n'⊂N') time series data by applying it to M artificial intelligence modules, separate N' for which the result information for specific purpose/use is already known. Compare the result information generated by substituting the time series data to the M artificial intelligence modules with the already known result information to check at least one of the accuracy, precision, and recall rate, and determine the accuracy and precision of the result information among the M artificial intelligence modules. , relates to a method of determining m (1≤m≤M) artificial intelligence modules that have the highest recall rate or a combination of two or more or are included in a preset ranking.

In general, 1st and 2nd generation smart farms minimize the impact of various environmental changes outside the greenhouse and set user control settings for the greenhouse's facilities and equipment to provide an optimal cultivation and watering environment for crops inside the greenhouse. It is operated in this way.

However, the control of greenhouse facilities and equipment to provide an optimal cultivation and positive water environment for crops in the greenhouse is realistically difficult to systematize standardized cultivation knowledge due to the variety of greenhouse structures and equipment operation methods, and it is difficult to systematize standardized cultivation knowledge, and it is difficult to systematize the user's knowledge and the target cultivation environment. Because it largely relies on experience and ability, there may be significant differences between farms. In addition, in order to accurately respond to situations where crop damage is expected, such as pests and diseases, physiological disorders, and preparation for abnormal weather, it is necessary to rely on the judgment and decision-making of experts (consultants). You can't help but rely heavily on it.

Meanwhile, as prior art, Republic of Korea Patent Publication No. 10-2021-015853 (published on December 31, 2021) relates to a system and method for generating farmland cultivation maps for artificial intelligence-based agricultural robots, using aerial images. After performing polygon rendering, the boundaries between cultivated fields are identified, the actual cultivable space according to the type of crop is identified, a cultivation map is created, and then provided to the robot agricultural machinery, so that cultivation work is accurately performed even in the cultivable area near the border between cultivated fields. The goal is to make it happen.

However, the above prior art performs polygon rendering using aerial images, generates a cultivation map that accurately identifies the boundary between farmlands through the density difference between polygons, and then transmits it to a robotic agricultural machine to place crops at the exact location through the robotic agricultural machine. It is just cultivating, and various environments for actual cultivation management in smart farms (cultivation environment management, water management, crop growth and physiological management, pest and physiological disorder management, agricultural work management, production and quality management, energy management, etc. ) could not be processed.

Therefore, while overcoming the situation where existing farms strongly prefer face-to-face consulting methods and have relatively low participation or satisfaction with non-face-to-face and online consulting projects, non-face-to-face online consulting for domestic facility farmers is possible under the global pandemic situation. There is a need to find new ways to meet needs and demands.

The purpose of the present invention is to model N (N ≥ 1) time series data by pre-processing the collected data collected in time series through various sensors or cameras installed in smart farms or smart factories, and to model N (N ≥ 1) time series data with a series of n (1) with periodic characteristics. ≤n≤N) time series data are processed into the frequency domain, and any one or two or more of the r (r≥1) known artificial intelligence algorithms (e.g. LSTM, CNN, etc.) are applied to a specific purpose/use. Prepare M(M≥1) artificial intelligence modules and N'(1≤N'≤N, n'⊂N' containing n'(1≤n'≤n) time series data processed into the frequency domain. ) time series data are individually applied to M artificial intelligence modules to learn them, and then N' separate time series data for which the result information for a specific purpose/use is already known are applied to each of M artificial intelligence modules to generate result information. and already known result information to check at least one of accuracy, precision, and recall, and among the M artificial intelligence modules, at least one or a combination of two or more of the accuracy, precision, and recall of the result information is the highest or in a preset ranking. The aim is to provide a method of providing artificial intelligence modules optimized for analysis/prediction using time series data to determine the m (1≤m≤M) number of artificial intelligence modules included.

The method of providing an artificial intelligence module optimized for analysis/prediction using time series data executed through an operation server according to the present invention is a method of providing an artificial intelligence module optimized for analysis/prediction using time series data through S (S ≥ 1) sensors installed in industrial or agricultural fields. Among the S sensed data and C image data captured in time series through C (C≥1) cameras installed at the site, designated D (1≤D≤(S+C)) collected data are collected. A first step of preprocessing and constructing T (T≥1) time series models related to the time series of processes or growth performed in the field and N (1≤N≤D) time series data modeled to enable matching within the allowable range. and t (1≤t≤T) time series models with a periodicity that can be matched with a specified periodicity related to at least one process or growth performed in the field based on the N time series data and a periodicity that can be matched within a specified tolerance range. A second step of confirming n (1≤n≤N) time series data with, and a third step of processing the confirmed n time series data into frequency domain-based time series data corresponding to specified frequency characteristics. step, and N'(1≤N'≤N, n'⊂N) containing n'(1≤n'≤n) time series data required for specified analysis or prediction among the n time series data processed into the frequency domain. A fourth method that configures a plurality of learning data containing ') time series data in a designated structure, and uses the plurality of learning data to learn M (M ≥ 2) artificial intelligence modules to which different artificial intelligence algorithms are applied. step, and after learning of the M artificial intelligence modules, n' for the specified analysis or prediction among the n time series data collected and preprocessed through the process corresponding to the first to third steps and then processed into the specified frequency domain. A fifth step of substituting N' time series data containing time series data into the learned M artificial intelligence modules to check result information for each M artificial intelligence module for a specified analysis or prediction, and the M artificial intelligence A sixth step of checking the actual measurement result information for each module, and matching and analyzing the confirmed actual measurement result information and the M result information for each artificial intelligence module to determine the M number of artificial intelligence modules for the specified analysis or prediction. A seventh step of calculating M evaluation information in which at least one of accuracy, precision, and recall is evaluated according to specified evaluation rules, and at least one product optimized for analysis or prediction specified based on the calculated M evaluation information. It may include an eighth step of determining m(1≤m≤M) artificial intelligence module.

According to the present invention, the preprocessing may include an averaging procedure to reveal the representativeness of a periodic pattern or a certain time interval of the D collected data with a specified periodicity.

According to the present invention, the preprocessing may include a procedure for correcting temporal influence corresponding to bias weighted by data from a previous time on data corresponding to a designated independent variable among the D collected data.

According to the present invention, the T time series models include a time series model that can match the time series characteristics of processes performed through devices installed in industrial or agricultural fields, time series characteristics related to the growth of crops grown in agricultural fields, and It may include at least one time series model among the time series models that can be matched.

According to the present invention, the N pieces of time-series data may include time-domain-based time-series data based on time-series characteristics of collected data generated by time-series sensing or time-series photography.

According to the present invention, the periodicity is a periodicity corresponding to at least one process repeatedly performed at regular intervals through a device provided at an industrial or agricultural field, and at least one process performed through a device provided at an industrial or agricultural field. Periodicity corresponds to a schedule associated with a process, periodicity corresponds to a change in the daily interval (or the Earth's rotation interval), periodicity to a change in a trend over a specified period, periodicity to correspond to a seasonal change, and a yearly interval (or the Earth's revolution). It may include at least one periodicity among the periodicities corresponding to changes in the interval.

According to the present invention, the t time series models include the time series characteristics matched with the time series characteristics of the process or growth performed in the field, and the periodic characteristics or specified changes (or trends) of the process performed including a designated period. A periodic time series model containing multiple periodicities matching the periodic characteristics of growth affected by can be included.

According to the present invention, the second step analyzes the N time series data based on at least one periodicity time series model to generate n time series data with a periodicity that can match at least one specified periodicity time series model and a specified tolerance range. It may include a confirmation step.

According to the present invention, the second step is to apply a specified exponential smoothing method to the time series data in the case of time series data in which the data change according to time is slower than a specified reference value, and then at least one periodic time series model and within a specified tolerance range are performed. It may include the step of identifying n time series data with matchable periodicity.

According to the present invention, the second step analyzes the N time series data based on at least one periodic time series model, and corresponds to a random element or residual/remainer element included in the time series data. It may include the step of separately analyzing the data to identify at least one periodicity time series model and n time series data with a periodicity that can be matched within a specified allowable range.

According to the present invention, the frequency feature may include a model-based period (or frequency) feature corresponding to a time series model with periodicity.

According to the present invention, the frequency characteristic may include a data-based period (or frequency) confirmed by reading time series data with periodicity.

According to the present invention, the frequency feature is statistical processing of the feature of the model-based cycle (or frequency) corresponding to a time series model with periodicity and the feature of the data-based cycle (or frequency) confirmed by reading time series data with periodicity. Alternatively, it may include characteristics of the period (or frequency) confirmed through numerical analysis processing.

According to the present invention, the n time series data processed into the frequency domain are a cycle that matches the periodic characteristics of the process performed, including the cycle designated in the field, or the periodic characteristics of growth affected by a designated change (or trend). It may include time series data processed into the frequency domain corresponding to frequency characteristics that can be matched within the performance tolerance range.

According to the present invention, the n' time series data are within an acceptable range and periodicity matching the periodic characteristics of the process performed including the designated cycle in the field or the periodic characteristics of growth affected by a designated change (or trend). Among n time series data processed into the frequency domain corresponding to matchable frequency characteristics, at least one time series data required for specified analysis or prediction may be included.

According to the present invention, the N' time series data includes n' time series data required for specified analysis or prediction among the n time series data processed into the frequency domain, and the specified analysis or prediction among the (N-n) time series data. It may include i(1≤i≤(N-n)) time series data required for .

According to the present invention, the fourth step may further include preprocessing the N' time series data so that they can be used as learning data for a designated artificial intelligence module.

According to the present invention, the preprocessing includes a data change procedure for changing the character or category attribute value of at least one time series data included in the N' time series data into numeric data, and the N' time series data included in the N' time series data. It may include at least one data scaling procedure that adjusts the range difference between each time series data to within a specified range.

According to the present invention, the preprocessing equalizes the periodicity of at least one time series data included in the N' time series data to a certain time period in a specified time range, thereby producing the result of the next data through k (k≥1) previous data. It may include procedures for processing values so that they can be derived.

According to the present invention, the preprocessing may include common preprocessing that is commonly applicable to time series data included in the learning data of the M artificial intelligence modules.

According to the present invention, the preprocessing may further include at least one individual preprocessing that can be individually applied to time series data included in the learning data of the M artificial intelligence modules.

According to the present invention, the learning data includes the n' time series data as data corresponding to observation features or feature vectors for artificial intelligence learning, and observation results related to the n' time series data or It may contain j(i≥1) pieces of data corresponding to the label.

According to the present invention, the learning data includes the n' time series data and includes (1 ≤ i ≤ (N-n)) time series data required for specified analysis or prediction among the (N-n) time series data. Includes '+i) time series data as data corresponding to observation features or feature vectors for artificial intelligence learning, and an observation result or label related to the specified time series data among the (n'+i) time series data. It may contain j(i≥1) pieces of data corresponding to (Label).

According to the present invention, the learning data may include M learning data composed of an optimized structure for each of the M artificial intelligence modules designated for the learning data including the n' time series data.

According to the present invention, the learning data may include integrated learning data composed of a structure that can be integrated and applied to M artificial intelligence modules designated for the learning data including the n' time series data.

According to the present invention, the fifth step may further include preprocessing the N' time series data so that they can be substituted into the learned M artificial intelligence modules.

According to the present invention, the result information for each of the M artificial intelligence modules is result information corresponding to artificial intelligence-based analysis or prediction using frequency domain-based time series data, and artificial intelligence using multiple frequency domain-based time series data and time domain-based time series data. It may include result information corresponding to intelligence-based analysis or prediction.

According to the present invention, after determining the mth artificial intelligence module, the n time series data collected and preprocessed through the process corresponding to the first to third steps are processed into a specified frequency domain for a specified analysis or prediction. It may further include a ninth step of verifying result information of the mth artificial intelligence module for a specified analysis or prediction by substituting N' time series data including n' time series data into the mth artificial intelligence module.

According to the present invention, through a method of providing an artificial intelligence module optimized for analysis/prediction using time series data, future data such as facility preventive maintenance, marketing evidence, and environmental change prediction are collected by using previous values in the form of past, present, and future. There is an advantage in being able to predict.

In addition, according to the present invention, there is an advantage of being able to predict abnormal cause values, such as predictive maintenance of equipment, prediction of crop growth/physiology, prediction of causes of pests and diseases, and maintenance of IoT environment by detecting patterns with repeated periodic characteristics.

Figure 1 is a diagram showing the configuration of an artificial intelligence module provision system optimized for analysis/prediction using time series data according to the implementation method of the present invention.
Figure 2 is a flowchart showing a process of processing time series data confirmed to have a periodicity that can match a specified periodicity related to process or growth into frequency domain-based time series data according to the implementation method of the present invention.
Figure 3 shows the process of substituting time series data for designated analysis or prediction into M learned artificial intelligence modules according to the implementation method of the present invention and confirming result information for each M artificial intelligence module for designated analysis or prediction. This is a flow chart.
Figure 4 shows the process of determining an artificial intelligence module optimized for a specified analysis or prediction based on the M evaluation information calculated by matching and analyzing the actual measurement result information and the result information for each M artificial intelligence module according to the implementation method of the present invention. This is a flow chart showing.

Hereinafter, the operating principle of a preferred embodiment of the present invention will be described in detail with reference to the accompanying drawings and description. However, the drawings shown below and the description below are for preferred implementation methods among various methods for effectively explaining the characteristics of the present invention, and the present invention is not limited to the drawings and description below.

That is, the following examples correspond to preferred union examples among the numerous examples of the present invention, and in the following examples, specific configurations (or steps) are omitted, or specific configurations (or steps) are omitted. An embodiment of dividing an implemented function into a specific configuration (or step), or an embodiment of integrating a function implemented in two or more configurations (or steps) into one configuration (or step), of a specific configuration (or step) It is clearly stated that all embodiments that replace the operation sequence fall within the scope of the present invention, even if not specifically mentioned in the following embodiments. Therefore, based on the examples below, it is clearly stated that various embodiments corresponding to subsets or complements can be divided retroactively to the filing date of the present invention.

Additionally, in the following description of the present invention, if it is determined that a detailed description of a related known function or configuration may unnecessarily obscure the gist of the present invention, the detailed description will be omitted. The terms described below are terms defined in consideration of functions in the present invention, and may vary depending on the intention or custom of the user or operator. Therefore, the definition should be made based on the overall content of the present invention.

As a result, the technical idea of the present invention is determined by the claims, and the following examples are a means to efficiently explain the advanced technical idea of the present invention to those skilled in the art to which the present invention pertains. It's just that.

Figure 1 is a diagram showing the configuration of an artificial intelligence module provision system optimized for analysis/prediction using time series data according to the implementation method of the present invention.

In more detail, Figure 1 models N (N ≥ 1) time series data by pre-processing the collected data collected in time series through various sensors or cameras installed in smart farms or smart factories, and models n with a series of periodic characteristics. Process (1≤n≤N) time series data into the frequency domain and use any one or two or more of the known r(r≥1) artificial intelligence algorithms (e.g. LSTM, CNN, etc.) for a specific purpose/use. Prepare M(M≥1) artificial intelligence modules applied to N'(1≤N'≤N, n'⊂) containing n'(1≤n'≤n) time series data processed into the frequency domain. After learning N') time series data by applying them to M artificial intelligence modules, N' separate time series data for which the result information for specific purposes are already known are applied to M artificial intelligence modules to generate Compare the result information with the already known result information to check at least one of accuracy, precision, and recall rate, and among the M artificial intelligence modules, at least one or a combination of two or more of the accuracy, precision, and recall rate of the result information is the highest or preset. It shows the system configuration for determining m (1≤m≤M) artificial intelligence modules included in the ranking. Those skilled in the art should refer to and/or modify Figure 1. Thus, it may be possible to infer various implementation methods for the system (e.g., implementation methods in which some components are omitted, subdivided, or combined), but the present invention includes all of the above inferred implementation methods, and this drawing The technical features are not limited only to the implementation method shown in 1.

The system of the present invention substitutes time series data collected from one or more sensors or cameras installed in industrial or agricultural fields into a plurality of learned artificial intelligence modules to provide result information for each artificial intelligence module for specified analysis or prediction. After checking, an operation server 100 that calculates evaluation information on the result information for each artificial intelligence module and determines one or more optimized artificial intelligence modules, and a user who receives the result information from the operation server 100 Includes a terminal 140. Meanwhile, the operating server 100 may be implemented in the form of at least one or a combination of two or more of the following: an independent server, a combination of two or more servers, and a server program installed and run on an equipped server. The present invention is not limited by the method of implementation.

The user terminal 140 is a general term for computer devices used by users that are related to the site and check the result information provided through the operation server 100, and include wired terminals such as computers and laptops connected to a wired network, and wireless terminals such as computers and laptops connected to a wired network. It may include at least one of wireless terminals such as mobile phones, smartphones, tablet PCs, and laptops connected to the network.

Preferably, the user terminal 140 has an application or program installed and running that performs at least one procedure specified for the function of receiving result information for a specified analysis or prediction from the operation server 100. desirable.

Meanwhile, referring to Figure 1, the operation server 100 includes a data configuration unit 105 that configures time series data collected and modeled through sensors or cameras provided in industrial or agricultural fields, and the field a data confirmation unit 110 that verifies time series data with a periodicity that can match a specified periodicity related to at least one process or growth performed in the process, and a data processing unit 115 that processes the time series data into frequency domain-based time series data. and an artificial intelligence learning unit 120 that configures a plurality of learning data containing time series data in a specified structure and trains a plurality of artificial intelligence modules using the configured plurality of learning data, and N' time series data. a result information confirmation unit 125 that substitutes the learned plurality of artificial intelligence modules to check result information for a specified analysis or prediction, and checks actual measurement result information for each artificial intelligence module; and at least one It will be comprised of an evaluation information calculation unit 130 that calculates M evaluation information by evaluating the artificial intelligence modules according to designated evaluation rules, and a module determination unit 135 that determines at least one optimized artificial intelligence module. You can.

According to Figure 1, the data configuration unit 105 includes S sensing data sensed in time series through S (S ≥ 1) sensors provided in industrial or agricultural fields and C provided in the field. Among C image data captured in time series through (C≥1) cameras, designated D (1≤D≤(S+C)) collected data are collected and preprocessed to create a time series of processes or growth performed at the site. It is possible to construct T(T≥1) time series models related to performance and N(1≤N≤D) time series data modeled to enable matching within the allowable range.

Typically, the preprocessing of the collected data involves any or a combination of two or more of data cleaning, data transformation, data filtering, data integration, and data reduction at least once. It may include performing

Here, the data purification includes a procedure for filling missing data with default values or replacing them with average/median values, a procedure for identifying and correcting or removing abnormal data corresponding to an abnormal state, and a procedure for smoothing noisy data. It may include any one or a combination of two or more.

In addition, the data conversion may include a procedure for converting data into a form that is easy to analyze through any one or a combination of two or more of normalization, aggregation, summarization, and hierarchy creation. You can.

Additionally, the data filtering may include any one or a combination of two or more of an error data discovery procedure, an error data correction procedure, an error data deletion procedure, a duplicate data confirmation procedure, a duplicate data correction procedure, and a duplicate data deletion procedure.

Additionally, the data integration may include a procedure for integrating similar data or linked data to facilitate data analysis.

Additionally, the data reduction may include a procedure for removing unused data from collected data.

Meanwhile, according to the implementation method of the present invention, the preprocessing of the collected data may include an averaging procedure to reveal the representativeness of a periodic pattern or a certain time interval of the collected data with a specified periodicity among the D collected data.

In addition, according to the implementation method of the present invention, the preprocessing of the collected data is a procedure of correcting the temporal influence corresponding to the bias weighted by data of the previous time with respect to the data corresponding to the designated independent variable among the D collected data. may include.

Meanwhile, T (T ≥ 1) time series models related to the time series characteristics of processes or growth performed in the field are time series models that can match the time series characteristics of processes performed through devices installed in industrial or agricultural fields. It may include at least one time series model among time series models that can match time series characteristics related to the growth of crops grown in the field.

In addition, the N (1≤N≤D) time series data modeled to enable matching within the above allowable range is a time domain based on the time series characteristics of the collected data generated by time series sensing or time series shooting. Can include time series data based.

According to Figure 1, when N time series data is configured through the data configuration unit 105, the data confirmation unit 110 determines at least one process or growth performed in the field based on the N time series data. You can check t(1≤t≤T) time series models with periodicity that can be matched with the specified periodicity related to and n(1≤n≤N) time series data with periodicity that can be matched within the specified allowable range.

According to the method of carrying out the present invention, the periodicity is a periodicity corresponding to at least one process repeatedly performed at regular intervals through a device provided at an industrial or agricultural field, through a device provided at an industrial or agricultural field. A periodicity corresponding to a schedule associated with at least one process to be performed, a periodicity corresponding to a change in the daily interval (or the Earth's rotation interval), a periodicity corresponding to a change in a trend over a specified period, a periodicity corresponding to a seasonal change, a yearly interval ( or at least one periodicity corresponding to a change in the Earth's orbital interval).

Here, it is clear that the periodicity does not include periodicity related to simple vibration of the device sensed through a conventional vibration sensor and periodicity related to the frequency of electromagnetic waves sensed through an electromagnetic wave sensor.

According to the implementation method of the present invention, t (1≤t≤T) time series models with periodicities that can match the specified periodicity related to the at least one process or growth are the time series characteristics of the process or growth performed in the field. It can include a periodic time series model that includes multiple periodicities that match the periodic characteristics of the process performed, including a specified period, or the periodic characteristics of growth affected by a specified change (or trend), while including time series matched to the specified period. .

In addition, the n (1≤n≤N) time series data with the periodicity may include time series data with a specified periodicity related to at least one process or growth performed in the field and a periodicity matched within a specified tolerance range. there is.

In addition, the n time series data may include a periodicity time series model that can match a specified periodicity related to at least one process or growth performed in the field and time series data with a periodicity that matches within a specified allowable range.

Meanwhile, according to the implementation method of the present invention, the data confirmation unit 110 analyzes the N time series data based on at least one periodic time series model and can match at least one specified periodic time series model within a specified allowable range. You can check n time series data with periodicity.

In addition, according to the implementation method of the present invention, in the case of time series data where the data change due to time change is slower than a specified reference value, the data confirmation unit 110 applies a specified exponential smoothing method to the time series data and then performs at least one You can check n time series data with periodicity that can match the periodic time series model and within the specified tolerance range.

In addition, according to the implementation method of the present invention, the data confirmation unit 110 analyzes the N time series data based on at least one periodic time series model, and random elements or residuals included in the time series data (Residual/Remainder) By separating and analyzing the data corresponding to the element, it is possible to identify at least one periodicity time series model and n time series data with a periodicity that can be matched within the specified allowable range.

According to Figure 1, when n (1≤n≤N) time series data with the periodicity are confirmed through the data confirmation unit 110, the data processing unit 115 selects the confirmed n time series data. It can be processed into frequency domain-based time series data corresponding to specified frequency characteristics.

According to the implementation method of the present invention, the frequency feature may include a model-based period (or frequency) feature corresponding to a time series model with periodicity.

Additionally, according to the implementation method of the present invention, the frequency characteristic may include a data-based period (or frequency) confirmed by reading time series data with periodicity.

In addition, according to the implementation method of the present invention, the frequency feature is the characteristic of a model-based period (or frequency) corresponding to a time series model with periodicity and a data-based period (or frequency) confirmed by reading time series data with periodicity. The characteristics of may be statistically processed (e.g., average, correction, etc.) or numerical analysis process (e.g., average, correction, interpolation, etc.) may be performed to include the identified period (or frequency) characteristics.

Meanwhile, according to the implementation method of the present invention, the n time series data processed into the frequency domain are the periodic characteristics of the process performed including the specified cycle in the field or the periodic characteristics of growth affected by the specified change (or trend). It may include time series data processed into a frequency domain corresponding to periodicity matching the characteristics and frequency characteristics that can be matched within an allowable range.

Additionally, the n pieces of time series data processed into the frequency domain do not include time series data related to simple vibration of the device sensed through a vibration sensor or time series data related to the frequency of electromagnetic waves sensed through an electromagnetic wave sensor.

Here, time series data related to the simple vibration of the device sensed through the vibration sensor or time series data related to the frequency of electromagnetic waves sensed through the electromagnetic wave sensor are the periodic characteristics or specified changes of the process performed including the specified period at the site. It is not included in the n time series data processed into the frequency domain corresponding to the periodicity that matches the periodic characteristics of growth affected by (or trend) and the frequency characteristic that matches within the allowable range.

In addition, time series data related to the simple vibration of the device sensed through the vibration sensor or time series data related to the frequency of electromagnetic waves sensed through the electromagnetic wave sensor are the periodic characteristics or specified changes of the process performed including the specified cycle at the site. Identified as time series data corresponding to a separate frequency domain that is distinguished from n time series data processed into a frequency domain corresponding to a periodicity that matches the periodic characteristics of growth affected by (or trend) and a frequency characteristic that matches within an acceptable range. Or it can be managed.

According to Figure 1, when the time series data is processed into frequency domain-based time series data through the data processing unit 115, the artificial intelligence learning unit 120 performs a specified analysis among the n time series data processed into the frequency domain. Or, a plurality of learning data containing N'(1≤N'≤N, n'⊂N') time series data including n'(1≤n'≤n) time series data required for prediction in a specified structure. It is possible to configure and learn M (M ≥ 2) artificial intelligence modules to which different artificial intelligence algorithms are applied using the plurality of configured learning data.

According to the implementation method of the present invention, the n' time series data is a periodicity that matches the periodic characteristics of the process performed including the designated cycle in the field or the periodic characteristic of growth affected by the designated change (or trend). It may include at least one time series data required for specified analysis or prediction among n time series data processed into the frequency domain corresponding to frequency characteristics that can be matched within the allowable range.

Meanwhile, the N' time series data may include n' time series data required for specified analysis or prediction among the n time series data processed into the frequency domain.

In addition, according to the implementation method of the present invention, the N' time series data includes n' time series data required for specified analysis or prediction among the n time series data processed into the frequency domain, and (N-n) time series data. It may include i(1≤i≤(N-n)) time series data required for specified analysis or prediction.

Here, the (N-n) time series data may include time domain-based time series data.

Meanwhile, according to the implementation method of the present invention, the artificial intelligence learning unit 120 can preprocess the N' time series data so that it can be used as learning data for a designated artificial intelligence module.

The preprocessing may perform any one or a combination of two or more of data purification, data conversion, data filtering, data integration, and data reduction on at least one time series data included in the N' time series data at least once.

According to the implementation method of the present invention, the preprocessing includes a data change procedure for changing the character or category attribute value of at least one time series data included in the N' time series data into numeric data, the N' time series data It may include at least one data scaling adjustment procedure for adjusting the range difference between each time series data included in to within a specified range.

Here, the preprocessing equalizes the periodicity of at least one time series data included in the N' time series data to a certain time period in a specified time range to derive the result value of the next data through k (k≥1) previous data. Possible processing procedures may be included.

Additionally, the preprocessing may include common preprocessing that is commonly applicable to time series data included in the learning data of the M artificial intelligence modules.

In addition, the preprocessing may further include at least one individual preprocessing that can be individually applied to time series data included in the learning data of the M artificial intelligence modules.

Meanwhile, according to the implementation method of the present invention, the learning data includes the n' time series data as data corresponding to observation features or feature vectors for artificial intelligence learning, and the n' time series data It may include j(i≥1) pieces of data corresponding to observation results or labels related to .

In addition, according to the implementation method of the present invention, the learning data includes the n' time series data, and among the (N-n) time series data, i (1 ≤ i ≤ (N-n)) time series data required for specified analysis or prediction. Includes (n'+i) time series data as data corresponding to observation features or feature vectors for artificial intelligence learning, and includes the designated time series data among the (n'+i) time series data. It may contain j(i≥1) data corresponding to related observation results or labels. Here, the i time series data may include time domain-based time series data.

Here, the learning data includes M learning data composed of optimized structures for each of the M artificial intelligence modules designated for the learning data including the n' time series data, or the n' time series data. It may include integrated learning data composed of a structure that can be integrated and applied to the designated M artificial intelligence modules for learning data containing.

Additionally, the i time series data may include time domain-based time series data.

Meanwhile, according to the implementation method of the present invention, the artificial intelligence module may include an LSTM (Long Short-Term Memory) series algorithm optimized for time series analysis.

According to Figure 1, the result information confirmation unit 125 trains M artificial intelligence modules using a plurality of learning data including time series data through the artificial intelligence learning unit 120, and then trains the M artificial intelligence modules. After learning of the artificial intelligence modules, N' time series data including n' time series data for specified analysis or prediction among the n time series data collected, preprocessed, and processed into the specified frequency domain are divided into the learned M pieces of time series data. By substituting into the artificial intelligence module, you can check the result information for each M artificial intelligence module for the designated analysis or prediction, and check the actual measurement result information for the result information for each of the M artificial intelligence modules.

Here, the n' time series data for the specified analysis or prediction are periodicity characteristics that match the periodic characteristics of the process performed in the field, including the designated cycle, or the periodic characteristics of growth affected by the designated change (or trend). It may include at least one time series data required for specified analysis or prediction among n time series data processed into the frequency domain corresponding to frequency characteristics that can be matched within the allowable range.

Additionally, the N' time series data including the n' time series data may include n' time series data required for specified analysis or prediction among the n time series data processed into the frequency domain.

In addition, the N' time series data including the n' time series data includes n' time series data required for specified analysis or prediction among the n time series data processed into the frequency domain, and (N-n) time series data. It may include i(1≤i≤(N-n)) time series data required for specified analysis or prediction. Here, the (N-n) time series data may include time domain-based time series data.

Meanwhile, according to the implementation method of the present invention, the result information confirmation unit 125 can preprocess the N' time series data so that they can be substituted into the learned M artificial intelligence modules.

Here, the preprocessing may perform any one or a combination of two or more of data purification, data conversion, data filtering, data integration, and data reduction on at least one time series data included in the N' time series data at least once. there is.

In addition, the preprocessing includes a data change procedure for changing character or category attribute values of at least one time series data included in the N' time series data into numeric data, and each time series data included in the N' time series data. It may include at least one of the data scaling procedures for adjusting the range difference between the data within a specified range.

In addition, the preprocessing equalizes the periodicity of at least one time series data included in the N' time series data to a certain time period in a specified time range to obtain the result value of the next data through k (k≥1) previous data. It may include procedures for processing so that it can be derived.

In addition, the preprocessing may include common preprocessing that is commonly applicable to time series data to be applied to the designated M artificial intelligence modules.

In addition, the preprocessing may further include at least one individual preprocessing that can be individually applied to time series data to be applied to the M artificial intelligence modules.

In addition, the result information confirmation unit 125 can confirm the result information for each M artificial intelligence module for the specified analysis or prediction by substituting the N' preprocessed time series data into the learned M artificial intelligence modules. .

Meanwhile, the result information for each of the M artificial intelligence modules is result information corresponding to artificial intelligence-based analysis or prediction using frequency domain-based time series data, and artificial intelligence-based analysis using multiple frequency domain-based time series data and time domain-based time series data. It may also include result information corresponding to the prediction.

Here, the result information for each of the M artificial intelligence modules may not include result information corresponding to artificial intelligence-based analysis or prediction using only time domain-based time series data.

Meanwhile, according to the implementation method of the present invention, the result information confirmation unit 125 stores the result information for each of the M artificial intelligence modules in a designated storage medium, and sends the M to a designated user terminal 140 related to the site. Result information for each artificial intelligence module can be provided.

Here, the result information for each of the M artificial intelligence modules includes result information of analyzing the occurrence of an abnormality, result information of predicting the occurrence of an abnormality, result information of predicting the future state, and control or adjustment of at least some processes performed at the site. At least one result information may be included among the result information being fed back.

According to the implementation method of the present invention, the result information confirmation unit 125 can receive actual measurement result information corresponding to the result information for each of the M artificial intelligence modules from the user terminal 140 corresponding to the field.

In addition, according to the implementation method of the present invention, the result information confirmation unit 125 generates/manages data corresponding to actual measurement result information corresponding to the result information for each of the M artificial intelligence modules or derives the actual measurement result information. The data can be received from a server (not separately shown) that generates/manages data to be used, and the actual measurement result information can be confirmed based on the received data.

In addition, according to the implementation method of the present invention, the result information confirmation unit 125 senses data corresponding to the actual measurement result information among sensors provided in the field or selects data to be used to derive the actual measurement result information. Sensing data sensed through a sensing sensor can be received, and the actual measurement result information can be confirmed based on the received sensing data.

According to Figure 1, when the actual measurement result information measured for the result information for each of the M artificial intelligence modules is confirmed through the result information confirmation unit 125, the evaluation information calculation unit 130 determines the confirmed actual measurement result. By matching and analyzing the information and the result information for each M artificial intelligence module, M evaluation information can be calculated by evaluating at least one of the accuracy, precision, and recall rate for each M artificial intelligence module for the specified analysis or prediction according to the specified evaluation rule. You can.

Here, the accuracy may include the ratio of accurately predicted actual 'true' and actual 'false', and the precision may include the ratio of actual 'true' among those predicted as 'true', , the recall rate may include the ratio of what is predicted to be ‘true’ among what is actually ‘true’.

In addition, the evaluation rule is a rule for scoring at least one of accuracy, precision, and recall for each of the M artificial intelligence modules, and a rule for quantifying at least one of accuracy, precision, and recall for each of the M artificial intelligence modules into a value within a specified range. It may include at least one rule, a rule for grading at least one of accuracy, precision, and recall rate for each of the M artificial intelligence modules.

In addition, the evaluation rule may further include a rule for assigning a predetermined weight related to the object to be analyzed or predicted through the artificial intelligence module to a designated item among accuracy, precision, and recall.

According to Figure 1, when the M evaluation information evaluated according to the evaluation rule specified through the evaluation information calculation unit 130 is calculated, the module determination unit 135 determines the M evaluation information based on the calculated M evaluation information. At least one m(1≤m≤M) artificial intelligence module optimized for analysis or prediction can be determined.

According to the implementation method of the present invention, the module determination unit 135 compares the M pieces of evaluation information and optimizes the artificial intelligence module corresponding to the highest evaluated (or highest ranked) evaluation information for the specified analysis or prediction. It can be decided by any one of the mth artificial intelligence modules.

In addition, according to the implementation method of the present invention, the module determination unit 135 compares the M pieces of evaluation information and selects at least one artificial intelligence module corresponding to the evaluation information within the specified ranking optimized for the specified analysis or prediction. The decision can be made with at least one mth artificial intelligence module.

Meanwhile, when the mth artificial intelligence module is determined, the result information confirmation unit 125 selects n' time series data for specified analysis or prediction among the n time series data collected, preprocessed, and then processed into the designated frequency domain. By substituting N' time series data including into the mth artificial intelligence module, the result information of the mth artificial intelligence module for the specified analysis or prediction can be confirmed.

Here, the n' time series data is a periodicity that matches the periodic characteristics of the process performed including the designated cycle in the field or the periodic characteristics of growth affected by a designated change (or trend), and a frequency that can be matched within an acceptable range. Among n time series data processed into the frequency domain corresponding to the feature, at least one time series data required for specified analysis or prediction may be included.

Additionally, the N' time series data may include n' time series data required for specified analysis or prediction among the n time series data processed into the frequency domain.

In addition, the N' time series data includes n' time series data required for a specified analysis or prediction among the n time series data processed into the frequency domain, and i required for a specified analysis or prediction among the (N-n) time series data. It may contain (1≤i≤(N-n)) time series data.

Additionally, the (N-n) pieces of time series data may include time domain-based time series data.

In addition, the result information confirmation unit 125 substitutes the N' time series data into the learned mth artificial intelligence module in the process of checking the result information of the mth artificial intelligence module for the specified analysis or prediction. Possibly preprocessed.

Here, the preprocessing involves performing any one or a combination of two or more of data purification, data conversion, data filtering, data integration, and data reduction at least once on at least one time series data included in the N' time series data. It can be included.

In addition, the preprocessing includes a data change procedure for changing character or category attribute values of at least one time series data included in the N' time series data into numeric data, and each time series data included in the N' time series data. It may include at least one data scaling adjustment procedure for adjusting the range difference between the data within a specified range.

In addition, the preprocessing equalizes the periodicity of at least one time series data included in the N' time series data to a certain time period in a specified time range to obtain the result value of the next data through k (k≥1) previous data. It may include procedures for processing so that it can be derived.

Meanwhile, the result information confirmation unit 125 substitutes the preprocessed N' time series data into the mth artificial intelligence module in the process of checking the result information of the mth artificial intelligence module for the specified analysis or prediction. You can check the result information of the m artificial intelligence module for the specified analysis or prediction.

Here, the result information of the mth artificial intelligence module is result information corresponding to artificial intelligence-based analysis or prediction using frequency domain-based time series data, and artificial intelligence-based analysis using multiple frequency domain-based time series data and time domain-based time series data. It may also include result information corresponding to the prediction.

Additionally, the result information of the mth artificial intelligence module may not include result information corresponding to artificial intelligence-based analysis or prediction using only time domain-based time series data.

In the process of checking the result information of the mth artificial intelligence module for the specified analysis or prediction, the result information confirmation unit 125 stores the result information of the mth artificial intelligence module in a designated storage medium, and stores the result information of the mth artificial intelligence module in the designated storage medium. The result information of the mth artificial intelligence module can be provided to a designated user terminal related to.

In addition, the result information of the mth artificial intelligence module includes result information of analyzing the occurrence of an abnormality, result information of predicting the occurrence of an abnormality, result information of predicting the future state, and control or adjustment of at least some processes performed at the site. At least one result information may be included among the result information being fed back.

Figure 2 is a flowchart showing the process of processing time series data confirmed to have a periodicity that can match the specified periodicity related to process or growth into frequency domain-based time series data according to the implementation method of the present invention.

In more detail, Figure 2 constitutes time series data collected and modeled by the operation server 100 through sensors or cameras provided in industrial or agricultural fields, and includes at least one process or growth and This diagram shows the process of confirming time series data with a periodicity that can match the relevant specified periodicity and processing the time series data into frequency domain-based time series data. Those skilled in the art will understand this drawing. By referring to and/or modifying 2, various implementation methods for the above process (e.g., implementation methods in which some steps are omitted or the order is changed) may be inferred, but the present invention includes all implementation methods inferred above. The technical features are not limited to the implementation method shown in Figure 2.

Referring to Figure 2, the operation server 100 stores S sensing data sensed in time series through S (S ≥ 1) sensors provided in industrial or agricultural fields and C (C) provided in the field. Among C pieces of image data captured in time series through ≥1) cameras, designated D (1≤D≤(S+C)) pieces of collected data are collected (200).

Here, the preprocessing may include an averaging procedure to reveal periodic patterns or representativeness of a certain time interval of the D collected data with a specified periodicity.

In addition, the preprocessing may include a procedure for correcting the temporal influence corresponding to the bias weighted by data of the previous time with respect to the data corresponding to the designated independent variable among the D collected data.

When the D pieces of collected data are collected, the operation server 100 preprocesses the collected D pieces of data to create T (T≥1) time series models related to the time series of the process or growth performed in the field and within the allowable range. Construct N (1≤N≤D) time series data modeled to enable matching (205).

Here, the T time series models are a time series model that can match the time series characteristics of processes performed through devices installed in industrial or agricultural fields, and a time series that can match the time series characteristics related to the growth of crops grown in agricultural fields. At least one time series model may be included among the models.

Additionally, the N pieces of time series data may include time domain-based time series data based on the time series characteristics of collected data generated by time series sensing or time series photography.

When the N time series data are configured, the operation server 100 sets t (1 ≤ t Confirm ≤T) time series models and n (1≤n≤N) time series data with periodicity that can be matched within the specified allowable range (210).

Here, the periodicity corresponds to at least one process repeatedly performed at regular intervals through a device installed in an industrial or agricultural field, and at least one process to be performed through a device provided in an industrial or agricultural field. periodicity corresponding to a schedule related to a periodicity, a periodicity corresponding to a change in a daily interval (or the Earth's rotation interval), a periodicity corresponding to a change in a trend in a specified period, a periodicity corresponding to a seasonal change, and a periodicity corresponding to a yearly interval (or the Earth's rotation interval). It may include at least one periodicity among the periodicities corresponding to change.

In addition, the t time series models include time series characteristics matched with the time series characteristics of the process or growth performed in the field and do not affect the periodic characteristics or specified changes (or trends) of the process performed including the specified period. It is possible to include a periodic time series model that includes multiple periodicities that match the periodic characteristics of the receiving growth.

According to the implementation method of the present invention, the operation server 100 analyzes the N time series data based on at least one periodic time series model and n with a periodicity that can match at least one specified periodic time series model within a specified allowable range. You can check time series data.

In addition, according to the implementation method of the present invention, in the case of time series data in which the data change according to time is slower than a specified reference value, the operation server 100 applies a specified exponential smoothing method to the time series data and then generates at least one periodic time series. You can check n time series data with periodicity that can match the model and within the specified tolerance range.

When the n time series data are confirmed, the operation server 100 processes the confirmed n time series data into frequency domain-based time series data corresponding to the specified frequency characteristic (215).

Here, the frequency feature includes a feature of a model-based period (or frequency) corresponding to a time series model with periodicity, or a data-based period (or frequency) confirmed by reading time series data with periodicity, Or, the characteristics of the model-based cycle (or frequency) corresponding to a time series model with periodicity and the characteristics of the data-based cycle (or frequency) confirmed by reading the time series data with periodicity ( or frequency) characteristics.

In addition, the n time series data processed into the frequency domain have a periodicity and allowable range that match the periodic characteristics of the process performed in the field, including the designated cycle, or the periodic characteristics of growth affected by the designated change (or trend). It can include time series data processed into the frequency domain corresponding to the frequency features that can be matched.

Figure 3 shows the process of substituting time series data for specified analysis or prediction into M learned artificial intelligence modules according to the implementation method of the present invention and confirming result information for each M artificial intelligence module for specified analysis or prediction. This is a flow chart.

In more detail, Figure 3 shows that after the process of Figure 2, the operation server 100 configures a plurality of learning data containing time series data in a designated structure, and uses the configured plurality of learning data to implement slightly different artificial intelligence algorithms. This shows the process of learning the applied M designated artificial intelligence modules, substituting the time series data into the learned M artificial intelligence modules, and confirming the result information for each M artificial intelligence module for the designated analysis or prediction. Those of ordinary skill in the art to which the invention pertains can infer various implementation methods for the above process (for example, implementation methods in which some steps are omitted or the order is changed) by referring to and/or modifying Figure 3. However, the present invention includes all the implementation methods inferred above, and its technical features are not limited to only the implementation method shown in Figure 3.

Referring to Figure 3, the operation server 100 processes the time series data into frequency domain-based time series data corresponding to the specified frequency characteristic through the process of Figure 2, and then the operation server 100 Among n time series data processed into the frequency domain, N'(1≤N'≤N, n'⊂N') time series containing n'(1≤n'≤n) time series data required for specified analysis or prediction. The data is preprocessed so that it can be used as learning data for M artificial intelligence modules to which different artificial intelligence algorithms are applied (300).

Here, the N' time series data includes n' time series data required for a specified analysis or prediction among the n time series data processed into the frequency domain, and i required for a specified analysis or prediction among the (N-n) time series data. It may contain (1≤i≤(N-n)) time series data.

In addition, the preprocessing includes a data change procedure for changing character or category attribute values of at least one time series data included in the N' time series data into numeric data, and each time series data included in the N' time series data. It may include at least one data scaling adjustment procedure for adjusting the range difference between the data within a specified range.

In addition, the preprocessing equalizes the periodicity of at least one time series data included in the N' time series data to a certain time period in a specified time range to derive the result value of the next data through k (k≥1) previous data. Possible processing procedures may be included.

Additionally, the preprocessing may include common preprocessing that is commonly applicable to time series data included in the learning data of the M artificial intelligence modules.

In addition, the preprocessing may further include at least one individual preprocessing that can be individually applied to time series data included in the learning data of the M artificial intelligence modules.

When the N' time series data are preprocessed to be usable as learning data for M artificial intelligence modules, the operation server 100 configures a plurality of learning data including the preprocessed N' time series data in a designated structure. (305).

Here, the learning data includes the n' time series data as data corresponding to observation features or feature vectors for artificial intelligence learning, and the observation results or labels related to the n' time series data. ) may include j(i≥1) pieces of data corresponding to ).

In addition, the learning data includes the n' time series data, and (n'+i) including i (1≤i≤(N-n)) time series data required for specified analysis or prediction among the (N-n) time series data. ) time series data as data corresponding to observation features or feature vectors for artificial intelligence learning, and observation results or labels related to the specified time series data among the (n'+i) time series data. It may include j(i≥1) pieces of data corresponding to .

In addition, the learning data includes M learning data composed of optimized structures for each of the M artificial intelligence modules designated for the learning data including the n' time series data, or the n' time series data. It may include integrated learning data composed of a structure that can be integrated and applied to the designated M artificial intelligence modules for learning data containing.

When the plurality of learning data is configured, the operation server 100 uses the configured plurality of learning data to learn M artificial intelligence modules to which different algorithms are applied (310).

Thereafter, the operation server 100 selects N' time series data including n' time series data for specified analysis or prediction among the n time series data collected, preprocessed, and processed into the designated frequency domain using the learned M. , Substitute the AI modules and check the result information for each M artificial intelligence module for the specified analysis or prediction (315).

Here, the result information for each of the M artificial intelligence modules may not include result information corresponding to artificial intelligence-based analysis or prediction using only time domain-based time series data.

In addition, the result information for each of the M artificial intelligence modules includes result information of analyzing the occurrence of an abnormality, result information of predicting the occurrence of an abnormality, result information of predicting the future state, and control or adjustment of at least some processes performed at the site. At least one result information may be included among the result information being fed back.

Thereafter, the operation server 100 stores the result information for each of the M artificial intelligence modules in a designated storage medium (320) and sends the result information for each of the M artificial intelligence modules to a designated user terminal 140 related to the site. Provided (325).

In addition, the result information for each of the M artificial intelligence modules is result information corresponding to artificial intelligence-based analysis or prediction using frequency domain-based time series data, and artificial intelligence-based analysis using multiple frequency domain-based time series data and time domain-based time series data. It may also include result information corresponding to the prediction.

Figure 4 shows the process of determining an artificial intelligence module optimized for a specified analysis or prediction based on the M evaluation information calculated by matching and analyzing the actual measurement result information and the result information for each M artificial intelligence module according to the implementation method of the present invention. This is a flow chart showing.

In more detail, Figure 4 shows that after the process of Figure 3, the operation server 100 checks the actual measurement result information for each M artificial intelligence module, and combines the confirmed actual measurement result information with the M artificial intelligence information. By matching and analyzing the result information for each module, M evaluation information is calculated by evaluating at least one of the accuracy, precision, and recall rate of each M artificial intelligence module for the specified analysis or prediction according to the specified evaluation rule, and the calculated M pieces of evaluation information are calculated. It illustrates the process of determining at least one m (1≤m≤M) artificial intelligence module optimized for specified analysis or prediction based on evaluation information, and those skilled in the art will understand , various implementation methods for the above process (e.g., implementation methods in which some steps are omitted or the order is changed) can be inferred by referring to and/or modifying Figure 4, but the present invention does not cover all the implementation methods inferred above. It includes, and its technical features are not limited to only the implementation method shown in Figure 4.

Referring to Figure 4, when the operation server 100 confirms the result information for each M artificial intelligence module for the analysis or prediction specified through the process of Figure 3, the M is sent from the user terminal 140 corresponding to the site. Actual measurement result information corresponding to the result information for each artificial intelligence module is received (400).

According to the implementation method of the present invention, the operation server 100 generates/manages data corresponding to actual measurement result information corresponding to the result information for each of the M artificial intelligence modules, or data to be used to derive the actual measurement result information. The data can be received from the server that creates/manages.

In addition, according to the implementation method of the present invention, the operation server 100 is a sensor that senses data corresponding to the actual measurement result information among sensors provided in the field or senses data to be used to derive the actual measurement result information. You can receive sensing data sensed through .

When actual measurement result information corresponding to the result information for each of the M artificial intelligence modules is received, the operation server 100 confirms the actual measurement result information based on the received data (405).

When the actual measurement result information is confirmed, the operation server 100 matches, compares and analyzes the confirmed actual measurement result information and the result information for each M artificial intelligence module (410), and determines the M number of artificial intelligence modules for the specified analysis or prediction. M evaluation information in which at least one of accuracy, precision, and recall is evaluated according to specified evaluation rules is calculated (415).

Thereafter, the operation server 100 determines at least one m (1 ≤ m ≤ M) artificial intelligence module optimized for the specified analysis or prediction based on the calculated M evaluation information (420).

According to the implementation method of the present invention, the operation server 100 compares the M pieces of evaluation information and selects an artificial intelligence module corresponding to the highest evaluated (or highest ranked) evaluation information to any optimized for specified analysis or prediction. It can be decided with a single artificial intelligence module.

In addition, according to the implementation method of the present invention, the operation server 100 compares the M pieces of evaluation information and selects at least one artificial intelligence module corresponding to the evaluation information within the specified ranking to at least one optimized for the specified analysis or prediction. It can be decided with the m artificial intelligence module.

When the mth artificial intelligence module is determined, the operation server 100 selects n' time series for specified analysis or prediction among the n time series data collected and preprocessed through the process of Figure 2 and then processed into a designated frequency domain. N' time series data containing data are substituted into the mth artificial intelligence module to check the result information of the mth artificial intelligence module for the specified analysis or prediction (425).

And, the operation server 100 stores the result information of the mth artificial intelligence module in a designated storage medium and provides the result information of the mth artificial intelligence module to the designated user terminal 140 related to the site ( 430).

Here, the result information of the mth artificial intelligence module includes result information of analyzing the occurrence of an abnormality, result information of predicting the occurrence of an abnormality, result information of predicting the future state, and control or adjustment of at least some processes performed at the site. At least one result information may be included among the result information being fed back.

100: Operation server 105: Data configuration unit
110: data confirmation unit 115: data processing unit
120: Artificial intelligence learning unit 125: Result information confirmation unit
130: Evaluation information calculation unit 135: Module decision unit
140: user terminal

Claims (28)

In the method executed through the operating server,
S sensing data sensed in time series through S (S ≥ 1) sensors installed in industrial or agricultural fields and captured in time series through C (C ≥ 1) cameras installed in the field. Among C image data, designated D (1≤D≤(S+C)) collected data are collected and preprocessed to create T(T≥1) time series models and allowable ranges related to the time series of the process or growth performed in the field. A first step of configuring N (1≤N≤D) time series data modeled to allow for my matching;
Based on the N time series data, t (1≤t≤T) time series models with a periodicity that can be matched with a specified periodicity related to at least one process or growth performed in the field and a periodicity that can be matched within a specified allowable range. A second step of checking n (1≤n≤N) time series data;
A third step of processing the confirmed n time series data into frequency domain-based time series data corresponding to specified frequency characteristics;
Among the n time series data processed into the frequency domain, N'(1≤N'≤N, n'⊂N') pieces containing n'(1≤n'≤n) time series data required for specified analysis or prediction. A fourth step of configuring a plurality of learning data including time series data in a designated structure, and learning M (M≥2) artificial intelligence modules to which different artificial intelligence algorithms are applied using the plurality of configured learning data;
After learning of the M artificial intelligence modules, n' time series data for specified analysis or prediction among n time series data collected and preprocessed through the process corresponding to the first to third steps and processed into a designated frequency domain. A fifth step of substituting N' time series data including the learned M artificial intelligence modules to confirm result information for each M artificial intelligence module for the specified analysis or prediction;
A sixth step of checking actual measurement result information for each of the M artificial intelligence modules;
The confirmed actual measurement result information and the result information for each M artificial intelligence module are matched, compared and analyzed, and at least one of the accuracy, precision and recall rate for each M artificial intelligence module for the specified analysis or prediction is evaluated according to the specified evaluation rule. Seventh step of calculating evaluation information; and
An eighth step of determining at least one m (1≤m≤M) artificial intelligence module optimized for the specified analysis or prediction based on the calculated M evaluation information; analysis/prediction using time series data including How to provide an optimized artificial intelligence module.
The method of claim 1, wherein the pretreatment is,
A method of providing an artificial intelligence module optimized for analysis/prediction using time series data, comprising an averaging procedure to reveal the representativeness of a periodic pattern or a certain time interval of the collected data with a specified periodicity among the D collected data.
The method of claim 1, wherein the pretreatment is,
Optimized for analysis/prediction using time series data, comprising a procedure for correcting temporal effects corresponding to biases weighted by data from previous times with respect to data corresponding to designated independent variables among the D collected data. How to provide artificial intelligence modules.
The method of claim 1, wherein the T time series models are:
A time series model that can match the time series characteristics of processes performed through devices installed in industrial or agricultural fields,
A method of providing an artificial intelligence module optimized for analysis/prediction using time series data, comprising at least one time series model among time series models that can be matched with time series characteristics related to the growth of crops grown in agricultural fields.
The method of claim 1, wherein the N time series data are:
An artificial intelligence module optimized for analysis/prediction using time series data, characterized by including time series data based on the time domain based on the time series characteristics of collected data generated by time series sensing or time series shooting. How to provide.
The method of claim 1, wherein the periodicity is:
Periodicity corresponding to at least one process that is repeatedly performed at regular intervals through a device installed in an industrial or agricultural field,
Periodicity corresponding to a schedule related to at least one process to be performed through a device provided in an industrial or agricultural field,
Periodicity, corresponding to changes in the daily interval (or Earth's rotation interval);
Periodicity, corresponding to changes in trends over a specified period of time;
periodicity in response to seasonal changes;
A method of providing an artificial intelligence module optimized for analysis/prediction using time series data, characterized in that it includes at least one periodicity among the periodicities corresponding to changes in one-year intervals (or Earth orbital intervals).
The method of claim 1, wherein the t time series models are:
Periodicity that matches the periodic characteristics of the process performed at the site or the time-series characteristics of growth, including a specified period, or growth that is affected by a specified change (or trend). A method of providing an artificial intelligence module optimized for analysis/prediction using time series data, characterized by including a periodic time series model containing multiple.
The method of claim 1, wherein the second step is:
A time series comprising the step of analyzing the N time series data based on at least one periodicity time series model and identifying n time series data with a periodicity that can be matched with at least one specified periodicity time series model and within a specified allowable range. Method of providing artificial intelligence modules optimized for analysis/prediction using data.
The method of claim 1 or 8, wherein the second step is:
In the case of time series data where the data change over time is slower than a specified reference value,
Analysis/prediction using time series data, comprising the step of applying a specified exponential smoothing method to the time series data and then identifying n time series data with a periodicity that can be matched with at least one periodicity time series model within a specified allowable range. How to provide an optimized artificial intelligence module.
The method of claim 1 or 8, wherein the second step is:
Analyze the N time series data based on at least one periodic time series model, and separate and analyze data corresponding to random elements or residual/remainder elements included in the time series data to determine at least one periodic time series. A method of providing an artificial intelligence module optimized for analysis/prediction using time series data, comprising the step of checking n time series data with a periodicity that can be matched with the model within a specified allowable range.
The method of claim 1, wherein the frequency characteristic is:
A method of providing an artificial intelligence module optimized for analysis/prediction using time series data, characterized by including features of model-based period (or frequency) corresponding to a time series model with periodicity.
The method of claim 1, wherein the frequency characteristic is:
A method of providing an artificial intelligence module optimized for analysis/prediction using time series data, characterized by including a data-based period (or frequency) confirmed by reading time series data with periodicity.
The method of claim 1, wherein the frequency characteristic is:
The characteristics of the model-based cycle (or frequency) corresponding to the time series model with periodicity and the characteristics of the data-based cycle (or frequency) identified by reading the time series data with periodicity are statistically processed or numerical analysis is performed to determine the cycle (or A method of providing an artificial intelligence module optimized for analysis/prediction using time series data, characterized by including the characteristics of frequency.
The method of claim 1, wherein the n time series data processed into the frequency domain are:
A time series processed into the frequency domain corresponding to a frequency characteristic that can be matched within an acceptable range and a periodicity that matches the periodic characteristics of the process performed at the site, including the specified cycle, or the periodic characteristics of growth affected by the specified change (or trend). A method of providing an artificial intelligence module optimized for analysis/prediction using time series data, characterized by including data.
The method of claim 1, wherein the n' time series data are:
n processed into a frequency domain corresponding to a frequency characteristic that can be matched within an acceptable range and a periodicity that matches the periodic characteristics of the process performed including the specified cycle in the field or the periodic characteristic of growth affected by a specified change (or trend) A method of providing an artificial intelligence module optimized for analysis/prediction using time series data, comprising at least one time series data required for specified analysis or prediction among time series data.
The method of claim 1, wherein the N' time series data are:
Among the n time series data processed into the frequency domain, it includes n' time series data required for specified analysis or prediction,
A method of providing an artificial intelligence module optimized for analysis/prediction using time series data, characterized in that it includes i (1≤i≤(Nn)) time series data required for specified analysis or prediction among (Nn) time series data.
The method of claim 1, wherein the fourth step is:
A method of providing an artificial intelligence module optimized for analysis/prediction using time series data, further comprising the step of preprocessing the N' time series data so that they can be used as learning data for a designated artificial intelligence module.
The method of claim 17, wherein the pretreatment is,
A data change procedure for changing a character or category attribute value for at least one time series data included in the N' time series data into numeric data,
Optimized for analysis/prediction using time series data, comprising at least one of the data scaling procedures for adjusting the range difference between each time series data included in the N' time series data to within a specified range. How to provide an artificial intelligence module.
The method of claim 17, wherein the pretreatment is,
A procedure for equalizing the periodicity of at least one time series data included in the N' time series data to a certain time period in a specified time range so that the result value of the next data can be derived through k (k≥1) previous data. A method of providing an artificial intelligence module optimized for analysis/prediction using time series data, comprising:
The method of claim 17, wherein the pretreatment is,
A method of providing an artificial intelligence module optimized for analysis/prediction using time series data, comprising common preprocessing commonly applicable to time series data included in the learning data of the M artificial intelligence modules.
The method of claim 20, wherein the pretreatment is,
A method of providing an artificial intelligence module optimized for analysis/prediction using time series data, further comprising at least one individual preprocessing that can be individually applied to time series data included in the learning data of the M artificial intelligence modules.
The method of claim 1, wherein the learning data is:
Containing the n' time series data as data corresponding to observation features or feature vectors for artificial intelligence learning,
A method of providing an artificial intelligence module optimized for analysis/prediction using time series data, comprising j (i≥1) data corresponding to observation results or labels related to the n' time series data.
The method of claim 1, wherein the learning data is:
Artificial intelligence analyzes (n'+i) time series data, including the n' time series data, and i (1≤i≤(Nn)) time series data required for specified analysis or prediction among the (Nn) time series data. Included as data corresponding to observed features or feature vectors for learning,
Optimized for analysis/prediction using time series data, characterized in that it includes j (i≥1) data corresponding to observation results or labels related to the specified time series data among the (n'+i) time series data. How to provide an artificial intelligence module.
The method of claim 22 or 23, wherein the learning data is,
An artificial intelligence optimized for analysis/prediction using time series data, characterized in that it includes M learning data composed of an optimized structure for each M artificial intelligence module designated with respect to the learning data including the n' number of time series data. How to provide an intelligence module.
The method of claim 22 or 23, wherein the learning data is,
Artificial intelligence optimized for analysis/prediction using time series data, characterized in that it includes integrated learning data composed of a structure that can be integrated and applied to M artificial intelligence modules designated for the learning data including the n' number of time series data. How to provide modules.
The method of claim 1, wherein the fifth step is:
A method of providing an artificial intelligence module optimized for analysis/prediction using time series data, further comprising preprocessing the N' time series data so that they can be substituted into the learned M artificial intelligence modules.
The method of claim 1, wherein the result information for each of the M artificial intelligence modules is:
Result information corresponding to artificial intelligence-based analysis or prediction using frequency domain-based time series data,
A method of providing an artificial intelligence module optimized for analysis/prediction using time series data, which includes result information corresponding to artificial intelligence-based analysis or prediction using multiple frequency domain-based time series data and time domain-based time series data.
According to clause 1,
After determining the mth artificial intelligence module,
N' time series data including n' time series data for specified analysis or prediction among the n time series data collected and preprocessed through the process corresponding to the first to third steps and then processed into a designated frequency domain. Provides an artificial intelligence module optimized for analysis/prediction using time series data, further comprising a ninth step of checking the result information of the mth artificial intelligence module for the specified analysis or prediction by substituting it into the mth artificial intelligence module. method.