KR20220042687A - Method of Determining Whether A Smart Farm Sensor has failed using a Recurrent Neural Network(RNN) - Google Patents

Abstract

The present invention relates to a method for determining whether a smart farm sensor has failed using a recurrent neural network (RNN) to increase reliability of smart farms. More specifically, the method comprises: a sensor information acquisition step of acquiring sensor information by a sensor; a learning checking step of checking, by a management server, whether learning of an RNN is performed; a failure determination step of using, by the management server, the sensor information on the basis of the RNN to determine whether the sensor has failed when the RNN is not being learned; and a determination result transmission step of transmitting, by the management server, a result of determining whether the sensor has failed to an administrator terminal.

Description

Method of Determining Whether A Smart Farm Sensor has failed using a Recurrent Neural Network (RNN)}

The present invention relates to a method for determining whether a smart farm sensor is faulty using a recurrent neural network (RNN). And it relates to a method for determining whether a smart farm sensor malfunctions using a predictive recurrent neural network (RNN).

Recently, in domestic agriculture, the deterioration of farmhouse management is accelerating due to the shortage of manpower, material cost, and cost of distribution channels. To overcome this, the government is developing a smart farm that combines agriculture with information and communication (IT) including artificial intelligence (AI) and big data, biotechnology (BT), and nanotechnology (NT) as a 4th industry and new deal policy. Active efforts are being made to form

In this process, the number of farms where smart farms are installed and operated is increasing. The core of the smart farm is to manage and control the farmhouse based on data collected from various types of sensors attached to the facilities and devices in the farmhouse. There is a problem that can lead to economic loss and personal injury.

In general industrial field failure diagnosis, studies that apply fuzzy theory and statistical reasoning are active, but smart farms are used in general industrial fields due to various data measured from various devices, environmental factors of humid farms, and complexity of requirements. There is a limit to directly applying the failure diagnosis technology of the smart farm, and accordingly, a method for diagnosing sensor failure for smart farms or a decision support method for maintenance is required.

Related Document 1 relates to a smart farm information management system, a smart farm information management method using the same, and a recording medium for performing the same, and generates product sales information including sales amount information that changes according to the growth information of a crop to be managed, or Because it is possible to create packaged product sales information, sellers can earn higher profits with the managed crop, but it is not possible to determine whether the sensor that derives the growth information of the managed crop is actually malfunctioning, thus proving the reliability of the growth information. There are downsides to not being able to.

Related Document 2 relates to an artificial intelligence smart farm management system. After learning using an artificial intelligence-based environment control algorithm, a plurality of control target devices can be controlled or updated periodically, but various sensors in the smart farm The focus is on acquiring data from the data acquisition sensor that can realize the automation of smart farms, and accordingly, it is not possible to diagnose and predict the failure of the sensor.

KR 10-1726257 KR 10-2018-0076766

The present invention is to solve the above problems. Using a plurality of pattern maps of a model learned in a recurrent neural network (RNN) series method through each activation function according to individual characteristics of a plurality of sensors in a smart farm, failure Sensor information is used based on the Jacquard similarity processing technique to diagnose whether there is a problem and improve the reliability of the smart farm. The purpose of this is to obtain a method for determining whether a smart farm sensor is faulty using a recurrent neural network (RNN) that is diagnosed according to the level and transmitted to the manager.

In addition, the present invention predicts the state of a plurality of sensors in advance so that the manager can replace and repair the sensors in advance, a plurality of pattern maps ( The predicted level at which the sensor failure is predicted is derived according to the feature map), and if the predicted level is above the stored reference level, it is classified as a stable level, and if it is less than the reference level, it is classified as a dangerous level and transmitted to the terminal of the manager. The purpose of this study is to obtain a method for determining whether a smart farm sensor is faulty using a recurrent neural network (RNN).

In order to achieve the above object, the method for determining whether a smart farm sensor is faulty using a recurrent neural network (RNN) of the present invention includes, by a sensor, a sensor information acquisition step in which sensor information is acquired; A learning confirmation step in which, by the management server, whether or not learning of the recurrent neural network (RNN) is checked; a failure determination step of determining whether the sensor is malfunctioning by using the sensor information based on the recurrent neural network (RNN) if the management server is not learning the recurrent neural network (RNN); and a determination result transmission step in which, by the management server, the determination result of whether the sensor is faulty is transmitted to the manager terminal.

As described above, in the present invention, sensor information is used based on the Jacquard similarity processing technique, and the sensor level indicating the degree of sensor failure is compared with a pre-stored reference level, and then the sensor level is classified, and the failure is according to the classified sensor level. By providing the diagnosis and transmission to the manager, there is an effect of diagnosing a malfunction in consideration of the individual characteristics of a plurality of sensors in the smart farm and improving the reliability of the smart farm.

The present invention is based on a number of pattern maps (feature maps) of a model learned in a recurrent neural network (RNN) series method through each activation function in the recurrent neural network (RNN), the predicted level of the sensor failure is derived, If the predicted level is higher than the stored reference level, it is classified as a stable level, and if it is less than the reference level, it is classified as a dangerous level and transmitted to the manager's terminal. It has the effect of being able to replace and repair.

1 is a flowchart of a method for determining whether a smart farm sensor is faulty using a recurrent neural network (RNN) of the present invention.
2 is a configuration diagram of a system for determining whether a smart farm sensor is faulty using a recurrent neural network (RNN) according to an embodiment of the present invention.
3 is a view showing a management server and a database according to an embodiment of the present invention.
4 is a detailed flowchart of the failure diagnosis step of the present invention.
5 is a view showing a risk level, a stable level, and a reference level according to an embodiment of the present invention.
6 is a detailed flowchart of the failure prediction step of the present invention.
7 is a case in which the prediction information according to an embodiment of the present invention is outside the error range (a) and within the error range (b) based on the prediction information previously learned from the recurrent neural network (RNN). It is a drawing.
8 is a diagram illustrating a Seq2Seq-based neural network model including a plurality of long-term dependency determining units and a variation auto-encoder according to an embodiment of the present invention.
9 is a diagram showing the structure of a variation auto encoder according to an embodiment of the present invention.

The terms used in this specification have been selected as currently widely used general terms as possible while considering the functions in the present invention, but these may vary depending on the intention or precedent of a person skilled in the art, the emergence of new technology, and the like. In addition, in a specific case, there is a term arbitrarily selected by the applicant, and in this case, the meaning will be described in detail in the description of the corresponding invention. Therefore, the term used in the present invention should be defined based on the meaning of the term and the overall content of the present invention, rather than the name of a simple term.

Unless defined otherwise, all terms used herein, including technical and scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. Terms such as those defined in a commonly used dictionary should be interpreted as having a meaning consistent with the meaning in the context of the related art, and should not be interpreted in an ideal or excessively formal meaning unless explicitly defined in the present application. does not

Hereinafter, a method for determining whether a smart farm sensor malfunctions using a recurrent neural network (RNN) of the present invention will be described in detail with reference to the accompanying drawings in accordance with an embodiment according to the present invention. 1 is a flowchart of a method for determining whether a smart farm sensor is faulty using a recurrent neural network (RNN) of the present invention. 2 is a configuration diagram of a system for determining whether a smart farm sensor is faulty using a recurrent neural network (RNN) according to an embodiment of the present invention. 3 is a view showing the management server 200 and the database 300 according to an embodiment of the present invention.

2 to 3, in order to execute the method for determining whether a smart farm sensor is faulty using a recurrent neural network (RNN) of the present invention, there are sensors 100a, 100b, 100n in a plurality of smart farm complexes, the sensors ( 100a, 100b, 100n) may include an integrated controller and an internal server. Here, a plurality of sensors may be provided within the single smart farm complex. Next, a relay server for receiving sensor information from the plurality of smart farm complexes and transmitting it to the management server 200 may be included. The relay server may be provided as a web or mobile software so as to transmit sensor information in both wired and wireless directions, and the manager terminal 400 is connected to the wired/wireless communication method to access the sensor information. Next, it may include a management server 200 that is connected to communicate with the relay server and processes the sensor information. The management server 200 may include a sensor information collection unit, a preprocessor, a sensor information learning unit, and a failure determination unit. The sensor information collection unit may receive and collect sensor information such as temperature, solar radiation, humidity, CO2, soil humidity, rainfall, wind direction, wind speed, EC, pH, and earth temperature from the sensors 100a, 100b, and 100n. The pre-processing unit may perform pre-processing including sampling, FFT, and filter processing for the sensor information so that the management server 200 can easily use it. The sensor information learning unit may include a recurrent neural network (RNN) to extract the characteristics of each sensor, and may learn by using the sensor information. In addition, the failure determination unit may determine whether the sensors 100a, 100b, and 100n have failures using the previously learned sensor information and the sensor information. In this case, in determining whether the failure occurs, it is possible to diagnose whether the sensors 100a , 100b , 100n are currently malfunctioning or to predict a failure in the future.

The management server 200 may store or call information generated after processing the sensor information in the connected database 300 .

1 , the method for determining whether a smart farm sensor is faulty using a recurrent neural network (RNN) of the present invention includes a sensor information acquisition step (S100), a learning status check step (S200), a failure determination step (S300), and a determination step It includes a transmission step (S400).

More specifically, in the sensor information acquisition step (S100), sensor information is acquired by the sensors 100a, 100b, and 100n.

Most preferably, the sensors 100a, 100b, 100n may be installed adjacent to the equipment in the smart farm, and the temperature, insolation, humidity, CO2, soil humidity, rainfall, wind direction, wind speed of the smart farm from the installed location as a starting point , pH, temperature, etc. are measured and then sensor information is created. In addition, the sensor information may be unstructured data including time series data. In general, unstructured data refers to unstructured data such as pictures, images, and documents, unlike numeric data with a certain standard or shape.

In addition, the sensor information may include a sensor ID of the corresponding sensor. Accordingly, in the sensor information acquisition step (S100), the sensor ID is used to access the table of the sensor 100 pre-stored in the database 300, and the sensors 100a, 100b, 100n) type, manufacturer, installation time, and installation location may be obtained as the sensor information.

Next, in the learning whether or not checking step (S200), by the management server 200, whether the learning of the recurrent neural network (RNN) is checked. The recurrent neural network (RNN) learns input data in real time or for a certain period of time. If new sensor information is received during learning, a collision may occur between the previously inputted information and the sensor information, which results in accurate output Information cannot be derived. Therefore, the learning whether or not checking step (S200) is to check whether the recurrent neural network (RNN) is learning before the failure determining step (S300) to prevent the above problems from occurring.

In other words, in the learning whether or not checking step (S200), if the recurrent neural network (RNN) is learning, it waits without going to the failure determining step (S300). In addition, it is checked whether the recurrent neural network (RNN) is learning at regular time intervals. If the recurrent neural network (RNN) has finished learning or is not learning, it goes to the failure determination step (S300).

Meanwhile, the recurrent neural network (RNN) may generate numerous learning models in the process of repeating the real-time learning process multiple times. Such a learning model judges the bias of the result, and a model with low bias is applied to multiple pattern maps of a later learned model, or a model with high bias is applied to multiple pattern maps of a later learned model. ) can be repeated. Accordingly, there is an effect that can improve the performance of accuracy for the fault diagnosis of the sensor of the present recurrent neural network (RNN).

Next, in the failure determination step (S300), if the recurrent neural network (RNN) is not being learned by the management server 200, the sensor information is used based on the recurrent neural network (RNN) and the sensor It is judged whether (100a, 100b, 100n) is faulty.

In this case, the failure determination step (S300) may include at least one of a failure diagnosis step (S310) and a failure prediction step (S320).

In the failure diagnosis step (S310), the sensor information is used so that the current state of the sensors 100a, 100b, and 100n can be diagnosed, so that the failure of the sensors 100a, 100b, 100n can be diagnosed.

Next, in the failure prediction step (S320), the sensor information is used so that the future state of the sensors 100a, 100b, 100n can be predicted, and the failure of the sensors 100a, 100b, 100n is predicted. can be

Next, in the determination result transmission step ( S400 ), the determination result of the failure of the sensors 100a , 100b , 100n is transmitted to the manager terminal 400 by the management server 200 .

Failure diagnosis

Hereinafter, the failure diagnosis step ( S310 ) will be described in more detail with reference to the accompanying drawings according to an embodiment of the present invention. 4 is a detailed flowchart of the fault diagnosis step 310 of the present invention. 5 is a view showing a risk level, a stable level, and a reference level according to an embodiment of the present invention.

First, referring to FIG. 4, the failure diagnosis step (S310) includes a sensor level deriving step (S311), a sensor level classification step (S312), a diagnosis result deriving step (S313) and a sensor information learning step (S314). can

More specifically, in the sensor level deriving step (S311), the sensor information and the sensor information previously learned in the recurrent neural network (RNN) are processed by the management server 200 by the jacquard similarity processing technique, and the sensor ( A sensor level in which the degree of failure of 100a, 100b, 100n) is indicated may be derived.

Here, the Jaccard similarity processing technique generally indicates the ratio of the intersection to the union of two sets when there are two sets, and has a value between 0 and 1. That is, if the two sets are the same, the maximum value of 1 can be derived because the intersection and the union are the same. If the two sets are completely different, the minimum value of 0 can be derived because the intersection is 0.

In other words, in the sensor level deriving step (S311), the sensor information (A) obtained from the sensor information obtaining step (S100) and the sensor information (B) previously learned in the recurrent neural network (RNN) are each set ( As A and B), the intersection ratio for the union is derived, which is called the sensor level. Most preferably, the sensor level may have a value between 0 and 1.

Next, in the sensor level classification step (S312), by the management server 200, if the sensor level is above the pre-stored reference level, it may be classified as a stable level, and if it is less than the reference level, it may be classified as a dangerous level.

Referring to FIG. 5 , when an arbitrary parameter is input by a user to the management server 200, a range between 0 and 1, which is a jacquard similarity range, may be arbitrarily divided according to the parameter, and accordingly, a plurality of levels may be set. there is. In addition, a reference level may be arbitrarily set by a user in the management server 200 , and a plurality of levels between 0 and 1 may be divided into a dangerous level and a stable level according to the reference level.

That is, between 0 and the reference level may be a risk level, and the closer the risk level is to 0, the less the similarity between the sensor information (A) and the sensor information (B) previously learned in the recurrent neural network (RNN). can be high And between the reference level and 1 may be a stable level, and the closer the stable level is to 1, the higher the similarity between the sensor information (A) and the sensor information (B) previously learned in the recurrent neural network (RNN), so the stability is can be high

Next, in the step of deriving the diagnosis result (S313), if the sensor level is a dangerous level by the management server 200, the sensors 100a, 100b, 100n are diagnosed as malfunctioning, and the severity of the risk level is reflected Diagnostic results can be derived. As an example using FIG. 5 , in the step of deriving the diagnosis result ( S313 ), if the sensor level is classified as risk level 5 from the sensor level classification step ( S312 ), the diagnosis result indicates that the sensor is “failed” and a diagnosis result such as “severity level of 5 out of 6” can be derived.

Next, in the sensor information learning step (S314), after the diagnosis result is derived by the management server 200, the sensor information may be learned in a recurrent neural network (RNN). That is, in the sensor information learning step S314, the sensor information may be learned so that the performance of the recurrent neural network (RNN) can be continuously improved and updated, rather than deriving only a diagnosis result and discarding or simply storing the sensor information. However, the sensor information learning step S314 may be performed after the diagnosis result is derived. This is because, as described above, when the recurrent neural network (RNN) is learned in the process of deriving a diagnosis result, an accurate diagnosis result cannot be derived.

Failure Prediction

Hereinafter, the failure prediction step (S320) will be described in more detail with reference to the accompanying drawings according to an embodiment of the present invention. 6 is a detailed flowchart of the failure prediction step of the present invention. 7 is a case in which the prediction information according to an embodiment of the present invention is outside the error range (a) and within the error range (b) based on the prediction information previously learned from the recurrent neural network (RNN). It is a drawing.

First, referring to FIG. 6, the failure prediction step (S320) includes a prediction information output step (S321), a prediction level deriving step (S322), a prediction level classification step (S323), and a prediction result deriving step (S324). can

More specifically, in the step of outputting the prediction information ( S321 ), when the sensor information is input to the recurrent neural network (RNN) by the management server 200, the prediction information may be output.

In this case, when the sensor information is input, the recurrent neural network (RNN) may output prediction information using one of several types of activation functions. In general, an activation function in a recurrent neural network (RNN) is a function that generates an output signal by processing an input signal, and it is an appropriate activation function because the types are very diverse and affect the output signal depending on which activation function is used. It is necessary to use

Next, in the prediction level deriving step (S322), at least one of a jacquard similarity processing technique or a cosine similarity processing technique is used by the management server 200 according to an activation function in the recurrent neural network (RNN), and the prediction information and the prediction information previously learned in the recurrent neural network (RNN) may be processed, and a prediction level at which the sensor failure is predicted may be derived.

For example, in the step of deriving the prediction level (S322), if the activation function is a Rectified Linear Unit (ReLU), a sigmoid function, or the like, a jacquard similarity processing technique is used to output the prediction information (S321) The prediction information (C) output from the and the prediction information (D) previously learned in the recurrent neural network (RNN) become sets (B, D), respectively, and an intersection ratio for the union is derived, which is called a prediction level. .

)와 상기 순환 신경망(RNN)에 기 학습된 예측 정보(

)의 두 벡터의 방향이 완전히 동일한 경우에는 1의 값, 90도를 이루면 0의 값, 방향이 반대이면 -1의 값이 도출되고 이를 예측레벨이라고 한다. 이때, 1에 가까울수록 유사도가 높다고 할 수 있다.On the other hand, in the step of deriving the prediction level (S322), if the activation function is a tanh function (Hyperbolic Tangent Function), a Leaky ReLU function, or the like, a cosine similarity processing technique is used, and the prediction information obtained from the step of outputting the prediction information (S321) (

) and the prediction information (

), a value of 1 is derived when the directions of the two vectors are exactly the same, a value of 0 when the direction is 90 degrees, and a value of -1 when the directions are opposite, which is called the prediction level. In this case, it can be said that the closer to 1, the higher the similarity.

In addition, in the step of deriving the prediction level ( S322 ), if the activation function is a Max Out function, an identity function, or the like, both the cosine similarity and the jacquard similarity values are derived, and a larger value among them is called a prediction level.

Next, in the prediction level classification step ( S323 ), if the prediction level is higher than or equal to a pre-stored reference level, it may be classified as a stable level, and if it is less than the reference level, it may be classified as a dangerous level.

If the prediction level is derived by the jacquard similarity processing technique, it can have a value between 0 and 1, so that a plurality of levels can be set by arbitrarily dividing between 0 and 1. Meanwhile, when an arbitrary parameter is input by the user, a range between 0 and 1, which is a jacquard similarity range, may be arbitrarily divided according to the parameter. In addition, the reference level may be arbitrarily set by the user, and a plurality of levels between 0 and 1 may be divided into a dangerous level and a stable level according to the reference level.

)와 상기 순환 신경망(RNN)에 기 학습된 예측 정보(

)의 유사도 적다는 것이므로 심각도가 높을 수 있다. 그리고 기준레벨과 1 사이는 안정레벨일 수 있고, 상기 안정레벨이 1에 가까울수록 상기 예측 정보(C)와 상기 순환 신경망(RNN)에 기 학습된 예측 정보(D)의 유사도가 높은 것이므로 안정도가 높을 수 있다.That is, between 0 and the reference level may be a risk level, and as the risk level is closer to 0, the prediction information (

) and the prediction information (

), the degree of similarity is small, so the severity may be high. And between the reference level and 1 may be a stable level, and the closer the stable level is to 1, the higher the degree of similarity between the prediction information (C) and the prediction information (D) previously learned in the recurrent neural network (RNN), so the stability is can be high

In addition, if the prediction level is derived using a cosine processing technique, it can have a value between -1 and 1, so that a plurality of levels can be set by arbitrarily dividing between -1 and 1. Meanwhile, when an arbitrary parameter is input by the user, a range between -1 and 1, which is a cosine similarity range, may be arbitrarily divided according to the parameter. In addition, a reference level may be arbitrarily set by the user, and a plurality of levels between -1 and 1 may be divided into a dangerous level and a stable level according to the reference level.

)와 상기 순환 신경망(RNN)에 기 학습된 예측 정보(

)의 두 벡터의 방향의 유사도 적다는 것이므로 심각도가 높을 수 있다. 그리고 기준레벨과 1 사이는 안정레벨일 수 있고, 상기 안정레벨이 1에 가까울수록 상기 예측 정보(

)와 상기 순환 신경망(RNN)에 기 학습된 예측 정보(

)의 유사도가 높은 것이므로 안정도가 높을 수 있다.That is, between -1 and the reference level may be a risk level, and as the risk level is closer to -1, the prediction information (

) and the prediction information (

), the similarity of the directions of the two vectors is also small, so the severity may be high. And between the reference level and 1 may be a stable level, and as the stable level is closer to 1, the prediction information (

) and the prediction information (

), the similarity is high, so the stability may be high.

Next, in the prediction result deriving step (S324), if the prediction level is a risk level, it is predicted that the sensors 100a, 100b, 100n will fail, and a prediction result reflecting the severity of the risk level may be derived.

As an example in Fig. 5, in the prediction result deriving step (S324), if the prediction level is classified as a risk level 5 from the prediction level classification step (S323), it is predicted that the sensor will fail as a result of the prediction, Predictive results such as “severity of 5 out of 6” can be derived.

Next, in the failure prediction step (S320) of the present invention, when the recurrent neural network (RNN) is learned, the learning determination step (S325) and the prediction information learning step (S326) may be further included.

More specifically, in the learning determination step (S325), after the prediction result is derived by the management server 200, the prediction information is circulated based on the prediction information previously learned from the recurrent neural network (RNN). Whether to learn may be determined in a neural network (RNN).

Next, in the prediction information learning step (S326), if the prediction information is within an error range based on the prediction information previously learned from the recurrent neural network (RNN) by the management server 200, the recurrent neural network (RNN) can be learned

On the other hand, in the prediction information learning step (S326), if the prediction information previously learned from the recurrent neural network (RNN) is out of an error range as a reference, the prediction information is deleted without being learned by the recurrent neural network (RNN) do it with Accordingly, there is an effect that the performance of the recurrent neural network (RNN) can be further improved by allowing only meaningful prediction information to be learned without learning prediction information such as noise with a large error.

7 (a), the prediction information (green graph) indicates a case outside the error range based on the prediction information (blue graph) previously learned from the recurrent neural network (RNN), and the prediction information ( The green graph) is arbitrary as noise, and may be deleted without being trained by the recurrent neural network (RNN).

Referring to FIG. 7B, the prediction information (green graph) indicates a case within an error range based on the prediction information (blue graph) previously learned from the recurrent neural network (RNN), and the prediction information ( The green graph) is meaningful information, so it can be learned by the recurrent neural network (RNN).

That is, in the process of repeating the real-time learning process a number of times in the recurrent neural network (RNN), numerous learning models may be generated. Such a learning model judges the bias of the result, and a model with low bias is applied to multiple pattern maps of a later learned model, or a model with high bias is applied to multiple pattern maps of a later learned model. ) can be repeated. Accordingly, there is an effect that can improve the performance of accuracy for the prediction of the failure of the sensor of the present recurrent neural network (RNN).

Recurrent Neural Network (RNN)

Hereinafter, a recurrent neural network (RNN) of the present invention will be described with reference to the accompanying drawings according to an embodiment of the present invention. 8 is a diagram illustrating a Seq2Seq-based neural network model including a plurality of long-term dependency determining units and a variation auto-encoder according to an embodiment of the present invention. 9 is a diagram showing the structure of a variation auto encoder according to an embodiment of the present invention.

On the other hand, for the recurrent neural network (RNN) of the present invention, if the sensor information is in an unstructured format, a Seq2Seq-based neural network model may be used, and if the sensor information is in a formal format, an LSTM-based neural network model may be used.

Here, the Seq2Seq-based neural network model is generally the most advanced architecture of a recurrent neural network, and is a specialized model for processing complex and vast sequence data by stacking cells in a recurrent neural network deep and wide, and in the case of other models, input and output dimensions In this defined state, learning and prediction are possible, but the Seq2Seq-based neural network model has a feature that is relatively free from the definition of input and output dimensions. In addition, unstructured data generally refers to unstructured data that has a different shape and structure, such as pictures, images, and documents, unlike numeric data having a certain standard or shape.

Therefore, if the sensor information of the present invention is in an unstructured format, a Seq2Seq-based neural network model is used to easily process sensor information made up of complex and vast sequence data in an unstructured format to derive a prediction level.

8 to 9, the Seq2Seq-based neural network model of the present invention has a plurality of long-term dependency determining units and a variation auto encoder ( Variational Auto Encoder (VAE) may be included. Most preferably, it may include one or more long-term dependency determiners and a variation auto-encoder (VAE).

The long-term dependency determining unit is based on a parameter set by a user or a performance indicator result of the recurrent neural network (RNN) to solve the long-term dependency problem of the recurrent neural network (RNN) model. Sensor information previously learned in the recurrent neural network (RNN) Alternatively, the long-term dependence may be determined according to the degree to which the prediction information disappears.

In addition, the variation autoencoder (VAE) reduces the dimension of the encoder part of the recurrent neural network (RNN) through a context switch when long-term dependency is determined from the long-term dependency determining unit, and the probability space Z ) to proceed with decoding, which is a restoration process again. Accordingly, there is an effect that can solve the problem of long-term dependence of the recurrent neural network (RNN), which is a conventional problem that the learning ability is rather deteriorated as the dimension of the recurrent neural network (RNN) increases.

In general, the recurrent neural network (RNN) model has a long-term dependency problem, and as the number of hidden layers of the neural network increases, a phenomenon occurs in which past information cannot be transmitted until the end. That is, when the distance between information and information increases significantly, the amount of change for backpropagation gradually decreases, and the learning ability of the recurrent neural network (RNN) deteriorates. A Long Short-Term Memory (LSTM) model has been proposed to compensate for this phenomenon, but it does not provide a complete solution.

In order to solve this problem, the recurrent neural network (RNN) of the present invention is most preferably a plurality of organs to enable context switch when the vector of the encoder part exceeds a certain dimension in order to process sensor information in an atypical format. It may be a Seq2Seq-based neural network model including a dependency determiner and a variational auto encoder (VAE).

Next, the LSTM-based neural network model is a recurrent neural network architecture used for long-term memory, and unlike a standard neural network, feedback is possible, so it is a model that can process not only single data but also entire data sequences such as voice or video. In addition, structured data is data configured to be stored according to a predetermined format and structure. Spreadsheet data with a well-known format or clear data structure for easy application to a system, CSV data with a comma structure determined etc. may be included. Therefore, if the sensor information is in a formal format, an LSTM-based neural network model is used, so that the sensor information is quickly processed.

As described above, although the embodiments have been described with reference to the limited embodiments and drawings, various modifications and variations are possible from the above description by those skilled in the art. For example, the described techniques are performed in an order different from the described method, and/or the described components of the system, structure, apparatus, circuit, etc. are combined or combined in a different form than the described method, or other components Or substituted or substituted by equivalents may achieve an appropriate result.

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

100a, 100b, 100n.. sensor
200. Management Server
300.. database
400.. Administrator terminal

Claims (5)

A sensor information acquisition step in which the sensor information is acquired by the sensor;
a learning confirmation step of checking whether the recurrent neural network (RNN) is learning by the management server;
a failure determination step of determining whether the sensor is malfunctioning by using the sensor information based on the recurrent neural network (RNN) when, by the management server, the recurrent neural network (RNN) is not being learned; and
A method for determining whether a smart farm sensor is faulty using a recurrent neural network (RNN) comprising a; a determination result transmission step of transmitting, by the management server, a determination result of whether the sensor is faulty to a manager terminal. The method of claim 1,
The failure determination step is,
a failure diagnosis step of diagnosing whether the sensor is defective by using the sensor information to diagnose the current state of the sensor; and
a failure predicting step in which the sensor information is used to predict a future state of the sensor to predict whether the sensor has a failure; A method of determining whether a smart farm sensor is faulty using a recurrent neural network (RNN), characterized in that it includes at least one of. 3. The method of claim 2,
The fault diagnosis step is,
A sensor level deriving step in which the sensor information and the sensor information previously learned in the recurrent neural network (RNN) are processed by the management server by the Jacquard similarity processing technique, and a sensor level indicating the degree of failure of the sensor is derived;
a sensor level classification step of classifying, by the management server, as a stable level if the sensor level is higher than a pre-stored reference level, and classified as a dangerous level if less than the reference level;
a diagnosis result derivation step in which, by the management server, if the sensor level is a dangerous level, the sensor is diagnosed as a failure, and a diagnosis result reflecting the severity of the risk level is derived; and
Smart farm sensor failure using a recurrent neural network (RNN), comprising: by the management server, a sensor information learning step in which the sensor information is learned in a recurrent neural network (RNN) after the diagnosis result is derived How to determine whether or not. 3. The method of claim 2,
The failure prediction step is,
a prediction information output step of outputting prediction information when the sensor information is input to the recurrent neural network (RNN) by the management server;
By the management server, at least one of a jacquard similarity processing technique or a cosine similarity processing technique is used according to an activation function in the recurrent neural network (RNN), and the prediction information and the prediction information previously learned in the recurrent neural network (RNN) are processed and a prediction level deriving step in which a prediction level at which the failure of the sensor is predicted is derived;
a prediction level classification step of classifying, by the management server, as a stable level if the prediction level is higher than or equal to a pre-stored reference level, and classified as a risk level if it is less than the reference level; and
Recurrent neural network (RNN) comprising: by the management server, predicting that the sensor will fail if the prediction level is a risk level, and deriving a prediction result reflecting the severity of the risk level; A method for determining whether a smart farm sensor is faulty using 5. The method of claim 4,
The failure prediction step is,
a learning determination step in which, after the prediction result is derived, by the management server, whether the prediction information is learned in the recurrent neural network (RNN) is determined based on the prediction information previously learned from the recurrent neural network (RNN); and
The method further comprises a; predictive information learning step of being learned by the management server by the recurrent neural network (RNN) if the prediction information is within an error range based on the prediction information previously learned from the recurrent neural network (RNN);
The predictive information learning step is,
Smart farm sensor using a recurrent neural network (RNN), characterized in that if the prediction information is outside the error range based on the prediction information previously learned from the recurrent neural network (RNN), it is deleted without being learned by the recurrent neural network (RNN) How to determine if there is a malfunction.

Images