KR102556173B1 - Method AND DEVICE FOR extracting abnormal data from machine learning model - Google Patents

Abstract

A method for extracting anomaly data of a machine learning model is provided. A method for extracting ideal data from a machine learning model according to some embodiments of the present invention includes obtaining sensing data, outputting first result data from a first machine learning model driving the sensing data as an input, and using the sensing data as an input. Outputting second result data from a second machine learning model driven by an input but different from the first machine learning model, comparing the sensing data and the second result data, whether or not the sensing data is ideal data based on the comparison result and when the sensing data is determined to be abnormal data, storing log information of target data that is the sensing data determined to be abnormal data.

Description

Method and device for extracting abnormal data from machine learning model {Method AND DEVICE FOR extracting abnormal data from machine learning model}

The present invention relates to a method for extracting abnormal data from a machine learning model, and more particularly, extracts abnormal data based on the recall of an autoencoder and uses log information of the extracted abnormal data to improve transfer learning performance of the machine learning model. It's about how to improve.

Fires occurring in human residences, offices, factories, etc. cause enormous property damage and human casualties, and as a result, the importance of fire detection algorithms that enable quick response by determining the occurrence of a fire at an early stage has been highlighted.

Existing fire detectors judged fire by detecting a specific temperature value using a single temperature sensor or detecting a specific amount (concentration) of smoke using a single smoke sensor based on a simple rule base. These fire detectors have the advantage of being inexpensive in terms of cost, but have the disadvantage of being weak against unwanted alarms. In order to reduce the rate of non-fire alarms, that is, false alarms caused by factors other than fire, a fire detection technique using various complex sensors can be used. In particular, recently, a fire detection method using machine learning techniques is required for complex judgment .

Machine learning is used for various purposes such as image processing, voice recognition, natural language processing, and fire detection. The advantage of a supervised learning-based machine learning model is that it can perform classification with high accuracy through learning from a large number of data. The supervised learning-based machine learning model collects data on the situation to be classified and assigns an appropriate label to train it. Then, the trained machine learning model can determine which situation the received data belongs to. there is. For example, a machine learning model for fire detection can classify non-fire and fire situations by collecting sensor data for non-fire and fire situations and assigning respective labels to the machine learning model. However, an error in judgment may occur in an abnormal situation (eg, sensor error, interference, malfunction, environmental noise, etc.) that was not considered in the learning stage. In particular, in the case of fire detection, such an error in judgment may cause damage due to non-fire alarm. For this reason, a technique for reducing judgment errors is required, and one of the techniques is a transfer learning technique that performs re-learning from continuously collected data.

On the other hand, the labeling task for transfer learning consumes a lot of time and cost due to the huge amount of data collected. In addition, data similar to training data included in the vast amount of continuously collected data does not significantly affect the performance improvement of the machine learning model, resulting in waste in the labeling process. On the other hand, if transfer learning of a machine learning model is performed using anomaly data that occurs when faced with a situation that was not considered when learning a machine learning model, a relatively high performance improvement can be expected.

A technical problem to be solved by the present invention is to provide a method and apparatus for increasing transfer learning performance of a machine learning model by determining whether a sensing value is abnormal and storing a log thereof.

Another object of the present invention is to provide a method and apparatus for improving the performance of a machine learning model while reducing the learning time by reducing the amount of training data used for transfer learning.

The technical problems of the present invention are not limited to the technical problems mentioned above, and other technical problems not mentioned will be clearly understood by those skilled in the art from the following description.

An ideal data extraction method of a machine learning model according to some embodiments of the present invention for achieving the above technical problem includes obtaining sensing data, first result data in a first machine learning model driving the sensing data as an input. Outputting, driving the sensing data as an input, but outputting second result data from a second machine learning model different from the first machine learning model, comparing the sensing data and the second result data, based on the comparison result The method may include determining whether or not the sensing data is abnormal data, and storing log information of target data that is the sensing data determined to be abnormal data when the sensing data is determined to be abnormal data.

According to some embodiments, the second machine learning model is an undercomplete stacked autoencoder, and the step of comparing the sensing data and the second resultant data includes determining a recall of the undercomplete stacked autoencoder. can do.

According to some embodiments, the step of determining whether the sensed data is abnormal data may include determining whether the data is abnormal based on whether or not the recall is greater than or equal to a predefined threshold value.

In this case, the step of determining whether the sensing data is abnormal data may include determining the sensing data as abnormal data when the recall is less than a threshold value.

According to some embodiments, the log information includes target data, first target result data output by a first machine learning model using the target data as an input, and second target result output by a second machine learning model using the target data as an input. It may include information on at least one of data, an input time of target data, and a sensing time of target data.

According to some embodiments, the method may further include performing transfer learning of the first machine learning model by using log information as training data.

According to some embodiments, the sensing data may include information on at least one of temperature, humidity, smoke amount, and gas amount, and outputting the first result data includes a result of whether or not a fire occurs in the target space. It may include outputting the first result data.

According to some embodiments, the step of storing the log information of the target data may include determining that the sensing data is abnormal data by a second machine learning model, and first result data generated by the first machine learning model in the target space. A step of storing log information of the target data may be included when information indicating that a fire has occurred is included.

According to some embodiments, the step of acquiring state information of at least one of a heater operating state, a window opening state, and weather information at the time of sensing the target data, and a first machine learning model by using the log information and the state information as learning data. It may further include the step of performing transfer learning of.

According to some embodiments, acquiring the sensing data may include acquiring a plurality of sensing data and generating preprocessed data obtained by scaling the plurality of sensing data to a predefined interval.

In this case, the first and second machine learning models may output first and second result data, respectively, with preprocessing data as an input.

According to some embodiments of the present invention, it may be implemented as a computer program stored in a computer readable recording medium in order to execute the method in combination with hardware.

Details of other embodiments are included in the detailed description and drawings.

1 is a diagram for explaining an apparatus to which a method for extracting abnormal data according to some embodiments of the present invention is applied.
2 is a diagram for explaining a determination unit according to some embodiments of the present invention.
3 is a diagram for explaining a second machine learning model according to some embodiments of the present invention.
4 is a diagram for explaining an autoencoder according to some embodiments of the present invention.
5 is a diagram for explaining an input unit according to some embodiments of the present invention.
6 is a flowchart illustrating a method of extracting abnormal data from a machine learning model according to some embodiments of the present invention.
7 is a flowchart illustrating an embodiment of determining abnormal data based on sensing data and a recall of second result data.
8 is a flowchart for explaining an embodiment of performing transfer learning of a first machine learning model using log information of target data.
FIG. 9 is a block diagram for explaining the embodiment of FIG. 8 .
10 is a flowchart illustrating an embodiment of determining whether to store log information based on sensing data and first result data.

Hereinafter, various embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that this is not intended to limit the present invention to the specific embodiments, but to cover various modifications, equivalents, and/or alternatives of the embodiments of the present invention.

Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with the meaning in the context of the related art, and unless explicitly defined in this specification, it should not be interpreted in an ideal or excessively formal meaning. don't

Various embodiments of this document and terms used therein are not intended to limit the technical features described in this document to specific embodiments, and should be understood to include various modifications, equivalents, or substitutes of the embodiments. In connection with the description of the drawings, like reference numbers may be used for like or related elements. The singular form of a noun corresponding to an item may include one item or a plurality of items, unless the relevant context clearly dictates otherwise. In this document, "A or B", "at least one of A and B", "at least one of A or B", "A, B or C", "at least one of A, B and C" and "A, Each of the phrases such as "at least one of B or C" may include any one of the items listed together in the corresponding one of the phrases, or all possible combinations thereof. Terms such as "first" and "second" may simply be used to distinguish a corresponding component from other corresponding components, and do not limit the corresponding components in other respects (eg, importance or order).

The embodiments may be implemented in various types of products such as personal computers, laptop computers, tablet computers, smart phones, televisions, smart home appliances, intelligent vehicles, kiosks, and wearable devices. Hereinafter, embodiments will be described in detail with reference to the accompanying drawings. Like reference numerals in each figure indicate like elements.

Hereinafter, a method for extracting abnormal data from a machine learning model according to some embodiments of the present invention will be described with reference to FIGS. 1 to 10 .

1 is a diagram for explaining an apparatus to which a method for extracting abnormal data according to some embodiments of the present invention is applied.

Machine learning is a learning method that implements human learning capabilities in machines such as computers. The methods for implementing machine learning include neural networks, data mining, decision trees, Genetic Algorithm, Case Based Reasoning, Pattern Recognition, Reinforcement Learning, Deep Learning, etc. may be included. Learning expressed as machine learning in this specification may be implemented in any one of the above methods rather than a specific implementation method.

An artificial neural network is a computing system that mimics the way the human brain processes information.

A deep neural network is one method of implementing an artificial neural network, and may include a plurality of layers. For example, a deep neural network has an input layer to which input data is applied, an output layer that outputs result values derived through prediction based on input data based on learning, and a gap between the input layer and the output layer. contains multiple hidden layers.

A deep neural network is classified into a convolutional neural network, a recurrent neural network, and the like according to an algorithm used to process information.

A method of learning artificial neural networks is called deep learning, and as described above, various algorithms such as convolutional neural networks and recurrent neural networks may be used in deep learning.

In this case, learning the artificial neural network may refer to determining and updating weights and biases between layers or weights and biases between a plurality of neurons belonging to different layers among adjacent layers.

For example, a plurality of hierarchical structures, a plurality of layers, or weights and biases between neurons may be collectively referred to as connectivity of an artificial neural network. Thus, learning an artificial neural network can refer to building and learning connectivity.

The apparatus 10 for extracting ideal data of a machine learning model according to some embodiments of the present invention generates a neural network, trains (or learns) the neural network, or retrains the neural network. It may correspond to a computing device having various processing functions such as For example, the apparatus 10 for extracting abnormal data of a machine learning model may be implemented in various types of devices such as a personal computer (PC), a server device, and a mobile device.

Referring to FIG. 1 , an apparatus 10 for extracting abnormal data of a machine learning model according to some embodiments of the present invention may include an input unit 100, a determination unit 200, and a storage unit 300, and an input unit ( The sensing data DT_SS input through the 100) is determined (or learned) by the determination unit 200, and the log information of the target data DT_TG, which is the sensing data determined to be abnormal data DT_ABN based on the determination result (LOG_TG) is stored in the storage unit 300, and the performance of the machine learning model can be improved by using it.

Although not shown, the storage unit 300 stores the acquisition time of the target data DT_TG, the first result data DT1 for the target data DT_TG, and the recall (or comparison data COMP) of the target data DT_TG. It may include a logging module generated with log information (LOG_TG) and a data storage in which the generated log information (LOG_TG) is stored.

Each of the components (input unit 100, determination unit 200 and storage unit 300) of the ideal data extraction device 10 of the machine learning model are separately shown in the drawing to show that they can be functionally and logically separated. It is indicated, and does not mean that it is physically a separate component or implemented as a separate code.

2 is a diagram for explaining a determination unit according to some embodiments of the present invention, FIG. 3 is a diagram for explaining a second machine learning model according to some embodiments of the present invention, and FIG. 4 is a diagram for some embodiments of the present invention. It is a diagram for explaining an autoencoder according to an example.

Referring to FIG. 2 , the determination unit 200 may include a first machine learning model 210 and a second machine learning model 220, and the first machine learning model 210 and the second machine learning model ( 220) may receive the sensing data DT_SS and output first result data DT1 and comparison data COMP based on the sensing data DT_SS. That is, the first and second machine learning models 210 and 220 may be learned using the same sensing data DT_SS as learning data.

The first and second machine learning models 210 and 220 may generate a trained neural network by repeatedly training (learning) a given initial neural network. Creating a trained neural network can mean determining neural network parameters. Here, the parameters may include, for example, various types of data input/output to the neural network, such as input/output activations, weights, and biases of the neural network. As the iterative training of the neural network progresses, the parameters of the neural network can be tuned to compute a more accurate output for a given input.

According to some embodiments, the determination unit 200 may be included in a mobile device, an embedded device, or the like. According to some embodiments, the determination unit 200 may be dedicated hardware for driving a neural network.

According to some embodiments, the first machine learning model 210 may determine whether a fire has occurred based on the sensing data DT_SS. As the first machine learning model 210, various machine learning models such as DNN, RNN, and CNN may be used, and an embodiment to which the present invention is applied is not limited to a specific machine learning model. The number of neurons in the input layer of the first machine learning model 210 is the number of sensor data such as temperature, humidity, smoke, and gas received from the fire detection sensor, and the number of neurons in the output layer indicates fire or non-fire. It can consist of two neurons representing

Each component of the determination unit 200 (the first machine learning model 210 and the second machine learning model 220) is shown separately in the drawing to show that it can be functionally and logically separated, and the physical This does not necessarily mean that it is a separate component or implemented as separate code.

Referring to FIGS. 3 and 4 , a second machine learning model 220 according to some embodiments of the present invention may include an autoencoder 221 and a comparison module 222 . According to some embodiments, the autoencoder 221 outputs second result data DT2 based on the sensing data DT_SS received by the input unit 100, and the comparison module 222 outputs the second result data ( Comparison data COMP may be output based on DT2) and sensing data DT_SS.

Each of the components (autoencoder 221 and comparison module 222) of the second machine learning model 220 are shown separately in the drawing to show that they can be functionally and logically separated, and must be physically separate. It does not mean that it is a component of or implemented as a separate code.

The autoencoder 221 is a neural network that is a type of unsupervised learning that copies input as output as it is, and efficiently converts data through a hidden layer between the input layer and the output layer. It is a neural network that learns how to represent That is, the autoencoder 221 generates second result data DT2 obtained by reproducing the sensing data DT_SS by taking the sensing data DT_SS as an input, and the sensing data DT_SS and the second result data DT2 are It may be similar data with a high recall rate or dissimilar data with a low recall rate.

According to some embodiments, the autoencoder 221 may be implemented as an undercomplete stacked autoencoder. Like the first machine learning model 210, the input layer receives sensing data (eg, temperature, humidity, smoke, gas, etc.) sensed by the fire detection sensor or preprocessing data of the sensing data as input. Provided, the output layer may have the same structure as the input layer (the number of neurons or the type of output data).

According to some embodiments, first machine learning model 210 and autoencoder 221 are trained using the same training data. If sufficiently trained, the first machine learning model 210 will be able to make a correct judgment in a situation similar to the learning data, and the autoencoder 221 will produce a second result from the output layer when data in a situation similar to the learning data is input. The reproducibility of the data DT2 will be high. In the present invention, it is assumed that learning of the first machine learning model 210 and the autoencoder 221 is completed, and the learning method of the first machine learning model 210 and the autoencoder 221 is not limited.

The comparison module 222 may output comparison data COMP based on the sensing data DT_SS and the second result data DT2. The comparison data COMP includes information on a comparison result of input data (sensing data DT_SS and second result data DT2) of the comparison module 222 . That is, the similarity between the sensing data DT_SS and the second result data DT2 may be determined, and information on the recall determined based on the similarity may be included. The reproducibility may be a numerical value representing the degree to which the second result data DT2 reproduces the sensing data DT_SS, and a determination condition for this may be set in various ways. That is, if the recall is high, the sensing data DT_SS is highly likely to be data similar to the pre-learned learning data, and accordingly, the sensing data DT_SS is highly likely to be normal data with low outliers. Therefore, it means that the reliability of the first result data DT1, which is the judgment result of the first machine learning model 210, is high. Conversely, when the recall is low, the sensed data DT_SS is highly likely to be abnormal data with a high outlier value due to external factors, and thus the reliability of the first result data DT1 is low. In this specification, "abnormal data DT_ABN" means data whose recall ratio is lower than a predefined threshold, and "normal data DT_NM" means data other than the abnormal data DT_ABN.

For example, information on “fire occurrence” is included in the first result data DT1 output by the first machine learning model 210, and comparison data output by the second machine learning model 220 ( COMP) contains information that the recall is high, the result of "fire outbreak" is likely not a non-fire report. In contrast, the first result data DT1 output by the first machine learning model 210 includes information on "fire occurrence", and the comparison data COMP output by the second machine learning model 220 ) includes information that recall is low, the result of "fire occurrence" is highly likely to show no fire.

In addition, since sensing data with a high recall rate, that is, normal data (DT_NM), has a small difference from the learning data of the first machine learning model 210, even if it is used as learning data for re-learning (or transfer learning) later, the first A significant effect on performance improvement of the machine learning model 210 cannot be expected. On the other hand, in the case of anomalous data (DT_ABN) having a low recall rate, when used as training data of the first machine learning model 210, effective learning is possible, and thus the performance of the first machine learning model can be improved.

5 is a diagram for explaining an input unit according to some embodiments of the present invention.

Referring to FIG. 5 , the input unit 100 may include a data collection module 110 receiving sensing data DT_SS and a data preprocessing module 120 performing a preprocessing operation on the sensing data DT_SS. In this specification, "sensing data DT_SS" may mean preprocessed data DT_PREP that has undergone a preprocessing process. That is, data received by the determination unit 200 may be preprocessed data DT_PREP.

Each of the components (data collection module 110 and data preprocessing module 120) of the input unit 100 are separately shown in the drawing to show that they can be functionally and logically separated, and are physically separate components It is not meant to be an element or implemented as separate code.

6 is a flowchart illustrating a method of extracting abnormal data from a machine learning model according to some embodiments of the present invention.

Referring to FIG. 6 , in a method for extracting abnormal data according to some embodiments of the present invention, first and second result data DT1 and DT2 are generated based on sensing data DT_SS input from the outside, and the sensing data It is possible to determine whether (DT_SS) is the abnormal data (DT_ABN) and store the log information (LOG_TG) of the target data (DT_TG) that is the abnormal data (DT_ABN).

In step S1000, the input unit 100 may obtain sensing data DT_SS from the outside. As described above, the sensing data DT_SS may be data sensed by a fire detection sensor. For example, the sensing data DT_SS may include at least one of temperature, humidity, smoke amount, and gas amount.

In step S2000, the first result data DT1 may be output from the first machine learning model 210 driven by the sensing data DT_SS as an input. The sensing data DT_SS may be preprocessing data DT_PREP generated by the input unit 100 . For example, the preprocessing data DT_PREP may be data obtained by scaling each sensing data DT_SS having different units to the same range. The first result data DT1 may be data determined by determining whether the sensing data DT_SS indicates fire or non-fire.

In step S3000, the second result data DT2 may be output from the second machine learning model 220 driven by the sensing data DT_SS as an input. At this time, the second result data DT2 may be data output by the autoencoder 221 included in the second machine learning model 220 .

In step S4000, the sensed data DT_SS and the second result data DT2 may be compared. Specifically, the comparison module 222 included in the second machine learning model 220 may compare the sensed data DT_SS and the second result data DT2.

In step S5000, it may be determined whether the sensing data DT_SS is abnormal data DT_ABN. Specifically, the recall of the sensing data DT_SS and the second result data DT2 is determined based on the comparison data COMP determined by the comparison module 222, and the sensing data DT_SS is abnormal based on the recall. It can determine whether it is data (DT_ABN). That is, the comparison data COMP output from the comparison module 222 may include information on whether the sensing data DT_SS is the abnormal data DT_ABN.

In step S6000, when the sensing data DT_SS is determined to be abnormal data DT_ABN, log information LOG_TG of the target data DT_TG may be stored. Specifically, the sensing data DT_SS determined as the abnormal data DT_ABN may be defined as target data DT_TG, and log information LOG_TG of the target data DT_TG may be stored in the storage unit 300 .

According to some embodiments, the log information (LOG_TG) is i) target data (DT_TG), ii) first target result data output by the first machine learning model 210 with the target data (DT_TG) as an input, iii) At least one of second target result data output by the second machine learning model 220 with the target data DT_TG as an input, iv) an input time of the target data DT_TG, and v) a sensing time of the target data DT_TG of information may be included.

7 is a flowchart illustrating an embodiment of determining abnormal data based on sensing data and a recall of second result data.

Referring to FIGS. 6 and 7 , target data DT_TG may be determined based on the recall of the sensing data DT_SS and the second result data DT2.

Step S4000 of FIG. 6 may include comparing the sensing data DT_SS and the second result data DT2 to determine a recall (S4100).

Subsequently, in step S5100, it is determined whether the recall rate is less than a predefined threshold value, and in step S5200, when the recall rate is less than the threshold value, the sensing data DT_SS may be determined as abnormal data DT_ABN. In other words, the sensing data DT_SS may be determined as the target data DT_TG. The threshold value may be predefined as a value between 0% and 100% of the recall rate of the autoencoder 221 .

When the recall is greater than or equal to the critical value, the sensing data DT_SS may be determined as normal data DT_NM in step S5300. That is, high recall means that there is a high probability that noise is not generated by external factors, and since the result of the first machine learning model 210 is reliable, the sensing data DT_SS is determined as normal data DT_NM. .

Thereafter, when the sensing data DT_SS is the target data DT_TG, log information LOG_TG of the target data DT_TG is stored in the storage unit (S6000), and when the sensing data DT_SS is not the target data DT_TG process can be terminated.

8 is a flowchart for explaining an embodiment of performing transfer learning of a first machine learning model using log information of target data, and FIG. 9 is a block diagram for explaining the embodiment of FIG. 8 .

Referring to FIGS. 8 and 9 , a step of performing transfer learning of the first machine learning model 210 using log information (LOG_TG) (S7000) may be further included. That is, the log information (LOG_TG) may be used as training data when re-learning of the first machine learning model 210 is performed. For example, it may be used as learning data of the first machine learning model 210 by examining whether a heater is used, whether a window is opened or not, whether a gas stove is used, etc. at the time when the target data DT_TG is sensed.

Transfer learning is a learning technique that reuses a model trained in a data-rich field to build a model in an area lacking learning data. In the present invention, log information (LOG_TG) of target data (DT_TG) determined as abnormal data (DT_ABN) can be collected, and the collected information can be used as learning data for transfer learning. According to the embodiment of the present invention, as the log information (LOG_TG) of the target data (DT_TG) is used as learning data for transfer learning, the first machine learning model 210 is used as learning data that is not similar to the previously learned learning data. This is learned, and thus more efficient learning is possible. That is, data that has little effect on improving the performance of the machine learning model (eg, normal data (DT_NM)) is excluded, and data that helps improve performance during retraining (eg, abnormal data (DT_ABN) or Re-learning efficiency can be increased by using the target data DT_TG as learning data.

10 is a flowchart illustrating an embodiment of determining whether to store log information based on sensing data and first result data.

Referring to FIG. 10, among the sensing data DT_SS determined to be abnormal data DT_ABN, only the sensing data DT_SS in which the first result data DT1 includes “fire” information is determined as the target data DT_TG , log information (LOG_TG) of the target data (DT_TG) may be stored.

More specifically, the comparison data COMP generated by the second machine learning model 220 includes information indicating that the recall of the second result data DT2 is less than the threshold value, and the first result data DT1 is " The corresponding sensing data DT_SS may be determined as the target data DT_TG only when the information of "fire" is included. By storing only the log information (LOG_TG) of the sensing data (DT_SS) when it is determined to be a fire, it is possible to reduce the possibility of non-fire alarm and increase the relearning efficiency of the first machine learning model 210.

The embodiments described above may be implemented as hardware components, software components, and/or a combination of hardware components and software components. For example, the devices, methods, and components described in the embodiments may include, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate (FPGA). array), programmable logic units (PLUs), microprocessors, or any other device capable of executing and responding to instructions. A processing device may run an operating system (OS) and one or more software applications running on the operating system. A processing device may also access, store, manipulate, process, and generate data in response to execution of software. For convenience of understanding, there are cases in which one processing device is used, but those skilled in the art will understand that the processing device includes a plurality of processing elements and/or a plurality of types of processing elements. It can be seen that it can include. For example, a processing device may include a plurality of processors or a processor and a controller. Other processing configurations are also possible, such as parallel processors.

Software may include a computer program, code, instructions, or a combination of one or more of the foregoing, which configures a processing device to operate as desired or processes independently or collectively. The device can be commanded. Software and/or data may be any tangible machine, component, physical device, virtual equipment, computer storage medium or device, intended to be interpreted by or provide instructions or data to a processing device. , or may be permanently or temporarily embodied in a transmitted signal wave. Software may be distributed on networked computer systems and stored or executed in a distributed manner. Software and data may be stored on one or more computer readable media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer readable medium. The computer readable medium may include program instructions, data files, data structures, etc. alone or in combination. Program commands recorded on the medium may be specially designed and configured for the embodiment or may be known and usable to those skilled in computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floptical disks. - includes hardware devices specially configured to store and execute program instructions, such as magneto-optical media, and ROM, RAM, flash memory, and the like. Examples of program instructions include high-level language codes that can be executed by a computer using an interpreter, as well as machine language codes such as those produced by a compiler.

The term "module" used in this document may include a unit implemented in hardware, software, or firmware, and may be used interchangeably with terms such as logic, logic block, component, or circuit, for example. A module may be an integrally constructed component or a minimal unit of components or a portion thereof that performs one or more functions. For example, according to one embodiment, the module may be implemented in the form of an application-specific integrated circuit (ASIC).

Various embodiments of this document may be implemented as software including one or more instructions stored in a machine. For example, the device may call at least one command among one or more commands stored from a storage medium and execute it. This enables the device to be operated to perform at least one function according to the at least one command invoked. The one or more instructions may include code generated by a compiler or code executable by an interpreter. The device-readable storage medium may be provided in the form of a non-transitory storage medium. Here, 'non-temporary' only means that the storage medium is a tangible device and does not contain a signal (e.g. electromagnetic wave), and this term refers to the case where data is stored semi-permanently in the storage medium. It does not discriminate when it is temporarily stored.

According to one embodiment, the method according to various embodiments disclosed in this document may be included and provided in a computer program product. Computer program products may be traded between sellers and buyers as commodities. A computer program product is distributed in the form of a device-readable storage medium (eg compact disc read only memory (CD-ROM)), or through an application store or between two user devices (eg smartphones). It can be distributed (e.g., downloaded or uploaded) directly or online. In the case of online distribution, at least part of the computer program product may be temporarily stored or temporarily created in a device-readable storage medium such as a manufacturer's server, an application store server, or a relay server's memory.

According to various embodiments, each component (eg, module or program) of the components described above may include a singular object or a plurality of entities. According to various embodiments, one or more components or operations among the aforementioned components may be omitted, or one or more other components or operations may be added. Alternatively or additionally, a plurality of components (eg modules or programs) may be integrated into a single component. In this case, the integrated component may perform one or more functions of each of the plurality of components identically or similarly to those performed by a corresponding component of the plurality of components prior to the integration. . According to various embodiments, operations performed by modules, programs, or other components are executed sequentially, in parallel, iteratively, or heuristically, or one or more of the operations are executed in a different order, omitted, or , or one or more other operations may be added.

100: input unit 110: data collection module
120: data pre-processing module 200: determination unit
210: first machine learning model 220: second machine learning model
221: autoencoder 222: comparison module
300: storage unit

Claims (10)

A method for extracting abnormal data of a machine learning model by a judgment unit of an abnormal data extraction device of a machine learning model,
obtaining sensing data through an input unit;
outputting first result data from a first machine learning model that drives the sensing data as an input;
driving the sensing data as an input and outputting second result data from a second machine learning model different from the first machine learning model;
comparing the sensing data and the second result data;
determining whether the sensing data is abnormal data based on the comparison result; and
When the sensing data is determined to be abnormal data, storing log information of target data that is the sensing data determined to be abnormal data;
performing learning of a first machine learning model by using the sensing data as learning data;
obtaining state information including an air heater operating state, a window opening/closing state, and a gas stove operating state at the time of sensing the target data; and
Further comprising performing re-learning of the first machine learning model by using the log information and the status information as learning data,
The sensing data includes at least one information of temperature, humidity, smoke amount, and gas amount,
The step of outputting the first result data includes outputting the first result data including a result of whether a fire occurs in a target space,
The step of storing the log information of the target data,
When the sensing data is determined as abnormal data by the second machine learning model and the first result data generated by the first machine learning model includes information that a fire has occurred in the target space, the target data Including the step of storing the log information of,
The log information includes the target data, first target result data output by the first machine learning model using the target data as an input, and second target output by the second machine learning model using the target data as an input. A method for extracting abnormal data of a machine learning model, including result data, an input time of the target data, a sensing time of the target data, and a recall of the target data.

The method of claim 1,
The second machine learning model is an undercomplete stacked autoencoder,
Comparing the sensing data and the second result data includes determining a recall of the under-perfect stacked autoencoder,
The method of extracting abnormal data from a machine learning model, wherein the step of determining whether the sensing data is abnormal data includes determining whether or not the abnormal data is based on whether the recall is equal to or greater than a predefined threshold value.
The method of claim 2,
The step of determining whether the sensing data is abnormal data,
And determining the sensing data as the abnormal data when the recall is less than the threshold value.
delete delete delete delete delete The method of claim 1,
The obtaining of the sensing data may include obtaining a plurality of sensing data; And generating preprocessed data by scaling the plurality of sensing data to a predefined interval,
Wherein the first and second machine learning models output the first and second result data, respectively, with the preprocessing data as an input.
A computer program stored in a computer readable recording medium in order to execute the method of any one of claims 1 to 3 and claim 9 in combination with hardware.

Images