Method and apparatus for monitoring performance of artificial intelligence model

Abstract

The present disclosure relates to an artificial intelligence performance monitoring technology, and provides an apparatus and method for monitoring the performance of an artificial intelligence model, the apparatus comprising: a first abnormality detection unit that stores input data received from a database for a preset period to generate statistical information and compares the statistical information with preset reference statistical information to detect whether an abnormality has occurred in the input data; a second abnormality detection unit that inputs the input data to a preset abnormality detection model and detects whether an abnormality has occurred in the input data on the basis of an output value of the abnormality detection model; and a third abnormality detection unit that inputs the input data to a preset artificial intelligence model and detects whether an abnormality has occurred in the artificial intelligence model by using output data of the artificial intelligence model.

Description

Method and device for monitoring the performance of an artificial intelligence model

[0001] The present disclosure relates to a technique for monitoring the performance of an artificial intelligence model.

[0002] With the advancement of artificial intelligence and deep learning technologies, attempts are being made to integrate them into various industrial settings. For example, in manufacturing settings, input data from various sensors can be used to perform predictive actions. This predicted information can then be used to preemptively identify abnormal situations or be applied to device control.

[0003] In this way, with the advancement of artificial intelligence technology, various artificial intelligence models suitable for each industrial field are being developed and applied in actual fields.

[0004] However, when implementing an AI model in an actual industrial setting, there are limitations in monitoring the AI ​​model.

[0005] For example, in industrial settings, if the characteristics of input data fluctuate or if input data becomes contaminated due to sensor issues or other problems, accurate predictions using AI models can become problematic. Furthermore, if a learning model is left unused for extended periods, its performance deteriorates, directly impacting product quality.

[0006] Therefore, AI models require maintenance after being deployed in industrial sites, but smooth communication between the development department and industrial sites may be limited, and continuous management is difficult.

[0007] Therefore, unified standards are limited, and technology is required to monitor artificial intelligence performance models by monitoring abnormalities in interconnected devices that may occur in industrial settings.

[0008] The present disclosure seeks to provide a technique for monitoring artificial intelligence model performance.

[0009] In one aspect, the present embodiments provide an artificial intelligence model performance monitoring device, comprising: a first anomaly detection unit that stores input data received from a database for a preset period of time to generate statistical information, compares the statistical information with preset reference statistical information, and detects whether an abnormality occurs in the input data; a second anomaly detection unit that inputs the input data to a preset anomaly detection model and detects whether an abnormality occurs in the input data based on an output value of the anomaly detection model; and a third anomaly detection unit that inputs the input data to a preset artificial intelligence model and detects whether an abnormality occurs in the artificial intelligence model using output data of the artificial intelligence model.

[0010] In another aspect, the present embodiments provide an artificial intelligence model performance monitoring method, comprising: a first anomaly detection step of storing input data received from a database for a preset period of time to generate statistical information, comparing the statistical information with preset reference statistical information to detect whether an abnormality has occurred in the input data; a second anomaly detection step of inputting the input data into a preset anomaly detection model and detecting whether an abnormality has occurred in the input data based on an output value of the anomaly detection model; and a third anomaly detection step of inputting the input data into a preset artificial intelligence model and detecting whether an abnormality has occurred in the artificial intelligence model using output data of the artificial intelligence model.

[0011] According to the present disclosure, a technology for monitoring artificial intelligence model performance can be provided.

[0012] FIG. 1 is a diagram illustrating the configuration of an artificial intelligence model performance monitoring device according to one embodiment.

[0013] Figure 2 is a diagram for explaining the flow of input data according to one embodiment.

[0014] FIG. 3 is a drawing for explaining the operation of a first abnormality detection unit according to one embodiment.

[0015] FIG. 4 is a diagram for explaining an operation for detecting anomalies using a difference in probability distribution according to one embodiment.

[0016] FIG. 5 is a drawing for explaining the operation of a second abnormality detection unit according to one embodiment.

[0017] FIG. 6 is a diagram for explaining the operation of an anomaly detection model according to one embodiment.

[0018] Fig. 7 is a diagram for explaining an operation for detecting whether an abnormality has occurred according to one embodiment.

[0019] FIG. 8 is a drawing for explaining the operation of a third abnormality detection unit according to one embodiment.

[0020] FIG. 9 is a drawing for explaining an operation of detecting whether an abnormality has occurred according to the operation of a third abnormality detection unit according to one embodiment.

[0021] Fig. 10 is a drawing for explaining an operation of detecting whether an abnormality has occurred according to the operation of a third abnormality detection unit according to another embodiment.

[0022] FIG. 11 is a diagram illustrating the overall flow of an artificial intelligence model performance monitoring operation according to one embodiment.

[0023] FIG. 12 is a diagram for explaining a method for monitoring artificial intelligence model performance according to one embodiment.

[0024] Hereinafter, some embodiments of the present disclosure will be described in detail with reference to exemplary drawings. When adding reference numerals to components in each drawing, identical components may have the same numerals as much as possible even if they are shown in different drawings. In addition, when describing the present embodiments, if it is determined that a detailed description of a related known configuration or function may obscure the gist of the technical idea of ​​the present invention, the detailed description may be omitted. When "includes," "has," "consists of," etc. are used in this specification, other parts may be added unless "only" is used. When a component is expressed in the singular, it may include a case in which the plural is included unless specifically stated otherwise.

[0025] Additionally, terms such as first, second, A, B, (a), (b), etc. may be used to describe components of the present disclosure. These terms are only intended to distinguish the components from other components, and the nature, order, sequence, or number of the components are not limited by the terms.

[0026] In a description of the positional relationship of components, when it is described that two or more components are "connected," "combined," or "connected," it should be understood that the two or more components may be directly "connected," "combined," or "connected," but that the two or more components may also be further "interposed" with another component to be "connected," "combined," or "connected." Here, the other component may be included in one or more of the two or more components that are "connected," "combined," or "connected" to each other.

[0027] In the description of the temporal flow relationship related to components, operation methods, or manufacturing methods, for example, when the temporal or flow relationship is described as “after”, “following”, “next to”, “before”, etc., it may also include cases where it is not continuous, unless “immediately” or “directly” is used.

[0028] Meanwhile, when numerical values ​​or corresponding information (e.g., levels, etc.) for components are mentioned, even without separate explicit description, the numerical values ​​or corresponding information may be interpreted as including an error range that may occur due to various factors (e.g., process factors, internal or external impact, noise, etc.).

[0029] The embodiments are described in detail with reference to the drawings below.

[0030] Below, the devices and methods according to the present disclosure are described in more detail with reference to the drawings. Each algorithm described below is provided as an example, and various algorithms that perform the same purpose and function can be applied to the present disclosure. Furthermore, programs capable of performing the functions of the present embodiments are also included in the present disclosure, and recording media containing the programs are also construed as being included in the present disclosure.

[0031] The present disclosure described below can be installed and operated in conjunction with an AI model. For example, it can be deployed on a cloud service to monitor AI model performance. Alternatively, it can be deployed on a specific server in an on-premise manner to monitor AI model performance. There are no restrictions on the AI ​​models that can be monitored, and any AI model can be monitored by the system according to the present disclosure. Therefore, the present disclosure has no restrictions on the AI ​​models.

[0032] Additionally, each component of the AI ​​model performance monitoring device according to the present disclosure may be constructed as multiple separate devices or integrated into a single device. For example, when deployed in a cloud service, the AI ​​model performance monitoring device may be constructed as multiple separate functions or devices. In this case, input and output data may be transmitted through a data pipeline between each device.

[0033] If data and learning models with changed characteristics are left unattended in the manufacturing field, model performance deteriorates and product quality can be seriously impacted. AI models require maintenance after deployment, but consistent management is challenging when development departments and application sites differ. Furthermore, even if maintenance is achieved, monitoring is difficult due to individual efforts without a unified standardized system.

[0034] To address these issues, this embodiment aims to provide automated AI model performance monitoring technology. For example, a cloud-based pipeline can be used to collect stream data from a database within a general-purpose cloud environment, store statistics, and detect anomalies in the input data. To achieve this, a model for anomaly detection can be trained, and then newly incoming production data from a new environment can be compared with existing statistics to determine whether there are any anomalies. This allows for automatic monitoring of input data anomalies.

[0035] These technologies can operate on large data sets and multidimensional variables, making them suitable for automated, general-purpose use in industrial environments. Furthermore, they calculate anomaly detection scores at each point in time to detect changes, enabling detection of unexpected spikes and periodic interruptions. Furthermore, they can additionally detect anomalies in AI models and distinguish whether these anomalies are due to input data anomalies or to degradation of the AI ​​model itself, facilitating follow-up measures.

[0036] FIG. 1 is a diagram illustrating the configuration of an artificial intelligence model performance monitoring device according to one embodiment.

[0037] Referring to FIG. 1, the artificial intelligence model performance monitoring device (100) may include a first anomaly detection unit (110) that stores input data received from a database for a preset period of time to generate statistical information, and compares the statistical information with preset reference statistical information to detect whether an anomaly occurs in the input data.

[0038] For example, when input data is input to the artificial intelligence model, the first abnormality detection unit (110) can store it for a preset period of time. That is, the first abnormality detection unit (110) stores the input data for a certain period of time using a buffer to detect whether an abnormality has occurred. In this case, the real-time input data is transmitted to the second abnormality detection unit (120) and the third abnormality detection unit (130). That is, the input data can be copied and stored in the first abnormality detection unit (110) and transmitted to the second abnormality detection unit (120) and the third abnormality detection unit (130).

[0039] The first anomaly detection unit (110) can detect whether an anomaly occurs in the input data using statistical information of the input data. For example, the statistical information may include at least one of the distribution, sum, standard deviation, minimum value, and maximum value of the input data.

[0040] The first anomaly detection unit (110) can detect whether an anomaly has occurred in the input data by comparing the statistical information of the input data with the reference statistical information. To this end, the reference statistical information needs to be set in advance.

[0041] For example, the preset baseline statistical information may include at least one statistical indicator among a baseline distribution, a baseline mean, a baseline sum, a baseline standard deviation, a baseline minimum value, and a baseline maximum value, which are generated using baseline data received from a database before input data is received. Baseline data refers to data generated identically to the input data, and refers to input data that is stored in advance in the same AI model construction situation before the input data is input. In other words, when an AI model is constructed in a specific situation, a test run can be performed using the baseline data for a preset period. The input data processed during the test run period may refer to the baseline data. Therefore, the statistical indicators of the input data and the baseline data must have similarity.

[0042] The first abnormality detection unit (110) can be set to detect whether an abnormality occurs in input data using various statistical indicators.

[0043] For example, the first anomaly detection unit (110) can detect that an abnormality has occurred in the input data if the statistical information and the same statistical indicator of the preset reference statistical information are out of the preset threshold range. For example, the first anomaly detection unit (110) can generate statistical information using input data stored for a certain period of time. The first anomaly detection unit (110) compares the same statistical indicator of the statistical information on the preset reference statistical information and the input data to determine whether the difference value is out of the preset threshold range, thereby determining whether an abnormality has occurred in the input data buffered for a certain period of time. The storage period of the input data can be set to be the same as the storage period for generating statistical information on the baseline data. Alternatively, it can be set differently. This information can be applied to statistical indicators that appear as a single value. For example, if statistical information appears as a single value such as the sum, standard deviation, maximum value, and minimum value, the first anomaly detection unit (110) can compare the difference between the statistical indicators to determine whether the difference is out of the threshold range.

[0044] As another example, the first anomaly detection unit (110) can detect whether an anomaly occurs in the input data by using a function that calculates the difference in probability distribution between the reference distribution and the distribution included in the statistical information. For example, if the statistical indicator is not a single value, the first anomaly detection unit (110) can detect whether an anomaly occurs in the input data by using the distribution among the statistical indicators. For example, the occurrence of an anomaly can be detected by comparing the reference distribution for a certain period of baseline data generated using baseline data with the distribution for a certain period of input data generated using input data.

[0045] To this end, the first anomaly detection unit (110) may use a function to calculate the probability distribution difference between each distribution. For example, the first anomaly detection unit (110) may use a function such as the Kullback-Leibler divergence to determine whether an anomaly has occurred based on whether the difference between probability distributions is above a certain level. In addition, various functions for calculating probability distribution differences may be applied.

[0046] Meanwhile, the artificial intelligence model performance monitoring device (100) may include a second anomaly detection unit (120) that inputs input data into a preset anomaly detection model and detects whether an anomaly occurs in the input data based on an output value of the anomaly detection model.

[0047] For example, the second anomaly detection unit (120) can input input data into a pre-learned and set anomaly detection model. The second anomaly detection unit (120) can detect whether an anomaly has occurred in the input data using the output value output by the anomaly detection model.

[0048] For example, the second anomaly detection unit (120) can detect that an anomaly has occurred in the input data when the output value of the preset anomaly detection model that calculates an anomaly score based on the number of divisions for isolating the input data from the baseline data exceeds a preset threshold value.

[0049] For example, a tree-structured model can be used as an anomaly detection model. The anomaly detection model can be trained using baseline data. Alternatively, the anomaly detection model can calculate the number of divisions for dividing a two-dimensional plane to separate newly input input data from the baseline data based on the distribution of the baseline data. The anomaly detection model can calculate the inverse value of the number of divisions as an outlier score. Accordingly, a higher outlier score can be determined as an abnormality in the input data. To this end, the second anomaly detection unit (120) can compare a threshold value with the outlier score, and if the outlier score is equal to or exceeds the threshold value, it can be determined that an abnormality has occurred in the corresponding input data.

[0050] As another example, the second anomaly detection unit (120) may use the number of segmentations for isolating input data as is, and a lower number of segmentations may be used to determine that an anomaly has occurred in the input data. In this case, the second anomaly detection unit (120) may determine that an anomaly has occurred in the input data if the number of segmentations is lower than a threshold value.

[0051] For example, anomaly detection models can use the Random Forest algorithm. Another example is the Random Cut Forest. In addition, various models can be applied as anomaly detection models, such as those that can detect when specific data is out of sync or deviates from the baseline data by converting this into numerical values.

[0052] Meanwhile, the artificial intelligence model performance monitoring device (100) may include a third abnormality detection unit (130) that inputs input data into a preset artificial intelligence model and detects whether an abnormality occurs in the artificial intelligence model using the output data of the artificial intelligence model.

[0053] For example, the third abnormality detection unit (130) can detect whether an abnormality has occurred by verifying the output data of the artificial intelligence model itself.

[0054] For example, the third anomaly detection unit (130) can input baseline data received from a database into an artificial intelligence model before receiving input data and use the accuracy information generated to set a reference value. In other words, a reference value for detecting the occurrence of an anomaly in the artificial intelligence model can be set using the baseline data.

[0055] The third anomaly detection unit (130) can detect an abnormality in the AI ​​model when the accuracy information of output data for input data is evaluated to be below a reference value. For example, the third anomaly detection unit (130) can input a certain level of baseline data into the AI ​​model and measure accuracy using output data for the baseline data. The reference value can be set in advance by applying a certain level of offset to the measured accuracy information.

[0056] The third abnormality detection unit (130) measures the accuracy of output data for input data, and can determine that an abnormality has occurred if the accuracy falls below a preset reference value.

[0057] Accuracy can be measured by comparing the predicted output data in an industrial setting to the actual values ​​generated after a certain period of time. For example, assuming an AI model that receives temperature data as input data and predicts the rotation speed of a fan heater, the accuracy of the actual AI model can be measured using rotation speed data for the fan heater after a certain period of time. Accordingly, the third abnormality detection unit (130) can use this to detect an abnormality in the AI ​​model if the accuracy falls below a certain level.

[0058] However, there may be cases where the accuracy of output data cannot be measured for various reasons. This may include cases where comparing predicted output data with actual data is difficult, or cases where accuracy is difficult to measure depending on the AI ​​model, such as models that generate actual drive control signals rather than predicted actions. Therefore, considering these cases, the third anomaly detection unit (130) can utilize the deviation between output data.

[0059] For example, the third anomaly detection unit (130) may input baseline data received from a database into an artificial intelligence model before receiving input data, and determine whether the deviation between the output data and the reference output data exceeds a preset threshold value. If the deviation is determined to be equal to or greater than the threshold value, the third anomaly detection unit (130) may detect that an anomaly has occurred in the artificial intelligence model.

[0060] To this end, the third abnormality detection unit (130) can check the output data of the artificial intelligence model in advance for a certain period of time using baseline data, and store the change in the output data over time as reference output data.

[0061] For example, the third anomaly detection unit (130) can detect whether an abnormality has occurred in the artificial intelligence model by storing reference output data for a certain period of time in the time domain and comparing the output data with the reference output data stored in the time domain at regular intervals. That is, the third anomaly detection unit (130) can compare the output data measured in real time by the input data based on the time-dependent flow of the stored reference output data, and detect whether an abnormality has occurred if the deviation between the two output data exceeds a certain level.

[0062] If necessary, the artificial intelligence model performance monitoring device (100) may further include a monitoring unit (140) that monitors at least one of an abnormality occurrence in input data and an abnormality occurrence in the artificial intelligence model, and provides a notification by distinguishing the type of abnormality occurrence.

[0063] For example, the monitoring unit (140) can receive information on whether an abnormality has been detected from the first abnormality detection unit (110), the second abnormality detection unit (120), and the third abnormality detection unit (130). The monitoring unit (140) can generate a control signal so that each piece of received information is output through an output device. In addition, the monitoring unit (140) can process the information on whether an abnormality has occurred so that whether an abnormality has occurred for the same input data is displayed by considering the correspondence between input data for mapping whether each abnormality has been detected.

[0064] For example, the first abnormality detection unit (110) stores input data for a certain period of time, and thus there may be a time difference with the second abnormality detection unit (120) that detects whether an abnormality has occurred for each input data. Taking this into account, the monitoring unit (140) can process and display information on whether an abnormality has occurred so that whether an abnormality has occurred for each input data is displayed correspondingly through an output device. Information on whether an abnormality has occurred for an artificial intelligence model can be displayed correspondingly to the input of input data.

[0065] Through this, when an abnormality is detected in the output data of an artificial intelligence model, the monitoring unit (140) can classify whether the abnormality is due to an abnormality in the input data or a deterioration of the artificial intelligence model. In addition, the monitoring unit (140) can provide the user with the cause and time of the abnormality through the classified results.

[0066] Through the aforementioned operation, the AI ​​model performance monitoring device (100) can monitor for and provide notifications of any abnormalities in the AI ​​model and input data when the AI ​​model is running. Furthermore, if an abnormality occurs in the AI ​​model, the AI ​​model performance monitoring device (100) can provide notifications by distinguishing whether the cause of the abnormality is due to deterioration of the AI ​​model or an abnormality in the input data.

[0067] In addition, even when the artificial intelligence model performance monitoring device (100) detects an abnormality in input data, it can also distinguish and determine whether an abnormality has occurred for each input data and whether an abnormality has occurred gradually in input data over a certain period of time. That is, when a specific input data appears as a jumpy value, the artificial intelligence model performance monitoring device (100) can immediately confirm this through the second abnormality detection unit (120) and determine whether it is due to temporary noise. However, in the case of the second abnormality detection unit (120), continuous input of input data below a reference value cannot be clearly detected. In this case, it is possible to confirm whether an abnormality has occurred in the input data as a whole by checking the statistical bias of the input data for a certain period of time of the first abnormality detection unit (110).

[0068] In this way, the present disclosure can efficiently detect various abnormal situations and monitor the performance of AI models already installed and operating in industrial settings. Furthermore, it can obtain the cause of anomalies, enabling accurate measures to be taken in response to the anomalies.

[0069] Below, various embodiments of the artificial intelligence model performance monitoring operation according to the present disclosure are described in detail with reference to the drawings. The description below is provided using examples for ease of understanding and is not limited to the examples.

[0070] Figure 2 is a diagram for explaining the flow of input data according to one embodiment.

[0071] Referring to FIG. 2, input data of the database (210) can be input equally to the first abnormality detection unit (110), the second abnormality detection unit (120), and the third abnormality detection unit (130). That is, the input data output from the database is input to each abnormality detection unit (110, 120, 130) through three pipelines.

[0072] In addition, the first abnormality detection unit (110), the second abnormality detection unit (120), and the third abnormality detection unit (130) each provide the results of detecting whether an abnormality has occurred to the monitoring unit (140). The results of detecting whether an abnormality has occurred may be provided to the monitoring unit (140) in real time or at regular intervals.

[0073] For example, the first abnormality detection unit (110) stores input data for a certain period of time and uses statistical information to detect whether an abnormality has occurred. The abnormality detection result generated by the first abnormality detection unit (110) can be periodically provided to the monitoring unit (140).

[0074] The second abnormality detection unit (120) and the third abnormality detection unit (130) detect whether an abnormality has occurred for each input data and output data, and the result of detecting whether an abnormality has occurred can be provided to the monitoring unit (140) according to the input of the input data.

[0075] Meanwhile, the monitoring unit (140) can manage the received abnormality detection results based on input data. That is, the reception cycle of the abnormality detection results may vary, and the target of the received abnormality detection results may be different input data and output data, and these can be managed to correspond to each other based on the input data. Through this, the cause of the abnormality can be quickly identified in a specific abnormality situation.

[0076] FIG. 3 is a drawing for explaining the operation of a first abnormality detection unit according to one embodiment.

[0077] Referring to FIG. 3, the first abnormality detection unit stores input data received from a database (210) for a preset period of time to generate statistical information (320), and compares the statistical information with preset reference statistical information (330) to detect whether an abnormality has occurred in the input data.

[0078] For example, the first anomaly detection unit may generate (310) baseline statistical information using baseline data received from the database (210). Baseline data may refer to data for a certain period of time in an environment where an artificial intelligence model is built. For example, baseline data refers to input data input to an artificial intelligence model during a period, such as a test run period, during which a user monitors the operation of the artificial intelligence model after the model is built.

[0079] Therefore, there is a difference in the time domain between operations 310 and 320 and 330. First, baseline statistical information can be generated (310) using baseline data. The baseline statistical information can include at least one of a baseline distribution, a baseline mean, a baseline sum, a baseline standard deviation, a baseline minimum value, and a baseline maximum value.

[0080] When input data is input, the first anomaly detection unit can generate statistical information identical to the reference statistical information (320) using the input data. To this end, the first anomaly detection unit can store input data received over a certain period of time and generate statistical information (320) according to a preset cycle.

[0081] The first abnormality detection unit compares (330) the generated statistical information with the reference statistical information. When comparing, the comparison must be made between identical statistical indicators.

[0082] For example, the first anomaly detection unit can detect an anomaly in the input data by comparing the same statistical indicators of the statistical information and the preset reference statistical information and if the statistical indicators fall outside the preset threshold range. For example, if the statistical indicators are set to a sum, the sum of the input data input over a certain period of time appears as a single value. Similarly, the reference sum also appears as a single value. Therefore, the first anomaly detection unit can detect an anomaly based on whether the difference between the sum of the input data and the reference sum falls outside the preset threshold range.

[0083] Similarly, even when the aggregate information is generated as a single value, such as in the case of the mean, standard deviation, minimum value, and maximum value, the first anomaly detection unit can compare the statistical indicators of the reference statistical information with the statistical indicators of the statistical information for the input data and determine whether an anomaly has occurred based on whether the difference is outside the critical range.

[0084] To achieve this, the baseline data collection period for generating standard statistics and the input data collection period can be set to be the same. Alternatively, they can be set to differ slightly depending on the type of statistical information. For the average, standard deviation, maximum, and minimum values, anomalies can be detected even if the periods are not identical. However, for the sum, the periods must be identical to detect anomalies.

[0085] As another example, the first anomaly detection unit can detect the occurrence of an anomaly in input data using a function that calculates the difference in probability distribution between a reference distribution and a distribution included in the statistical information. This is explained with reference to FIG. 4.

[0086] FIG. 4 is a diagram for explaining an operation for detecting anomalies using a difference in probability distribution according to one embodiment.

[0087] Referring to FIG. 4, the first anomaly detection unit can store input data for a certain period of time and calculate a distribution for the input data. In the case of a reference distribution, distribution information for baseline data is also included.

[0088] The first anomaly detection unit can detect the presence of anomalies in input data by comparing distribution information for input data and baseline data. Figures 410 and 420 each show graphs comparing distributions.

[0089] The X-axis of 410 and 420 represents the values ​​of the input data and baseline data, and the Y-axis represents the frequency of data for each value as a probability density percentage.

[0090] Baseline data, such as 410, may be normally distributed around the value 400. Alternatively, baseline data, such as 420, may be distributed around the values ​​40 and 100, with data occurrences between 70 and 80 occurring less frequently.

[0091] The first anomaly detection unit calculates distributions for the baseline data and input data, and compares these distributions to detect anomalies. Both 410 and 420 indicate normal conditions, as shown by the input data distribution being included in the baseline data.

[0092] Meanwhile, the first anomaly detection unit can use a function that calculates the difference between preset probability distributions to determine whether a difference occurs between distributions. For example, an indicator that measures the similarity between probability distributions, such as the Kullback-Leibler divergence, can be used to detect anomalies.

[0093] Kullback-Leibler divergence (KLD) is a function used to calculate the difference between two probability distributions. It calculates the difference in information entropy that would occur if sampling from an ideal distribution using another distribution that approximates the original distribution. Therefore, Kullback-Leibler divergence uses the entropy difference to determine the similarity between probability distributions, resulting in a value greater than or equal to 0. In addition, various functions can be applied to determine the similarity between probability distributions. For example, functions such as Jensen-Shannon divergence and log likelihood can be applied.

[0094] FIG. 5 is a drawing for explaining the operation of a second abnormality detection unit according to one embodiment.

[0095] Referring to FIG. 5, the second abnormality detection unit inputs input data into a preset abnormality detection model (510) and can detect whether an abnormality has occurred in the input data based on the output value of the abnormality detection model.

[0096] For example, the anomaly detection model (510) may be an unsupervised algorithm model that is learned using baseline data received from a database before input data is received.

[0097] The anomaly detection model (510) can be trained using baseline data. Thereafter, the second anomaly detection unit inputs input data into the trained anomaly detection model (510) to obtain an anomaly score and compares the anomaly score with a threshold value (520) to detect whether an anomaly has occurred.

[0098] For example, the anomaly detection model (510) represents a model that distinguishes between normal samples and abnormal samples. There is no limitation on the anomaly detection model (510) as long as it can distinguish between normal and abnormal. For example, the anomaly detection model (510) can use various categories of anomaly detection algorithms such as Supervised Anomaly Detection, Semi-supervised (One-Class) Anomaly Detection, and Unsupervised Anomaly Detection.

[0099] FIG. 6 is a diagram for explaining the operation of an anomaly detection model according to one embodiment.

[0100] Referring to Fig. 6, an anomaly detection model may be used that obtains an outlier score based on the number of splits for separating input data.

[0101] For example, the Random Cut Forest (RCF) model can be used. The RCF model is an unsupervised algorithm for detecting anomalous data points within a dataset. The RCF model is used to identify observations that deviate from structured or patterned data. For example, outliers can appear as unexpected spikes in time-series data, breaks in periodicity, or classifiable data points. Therefore, it can be useful for detecting anomalies in multidimensional streaming datasets.

[0102] The anomaly detection model can calculate an outlier score based on the number of segmentations required to separate input data points (615, 625). For example, a total of six spatial segmentations are required to separate input data points (615) from existing data, as in 610. In another example, if input data points (625) are separated from other data and pointed to in space, as in 620, the separation can be performed through two spatial segmentations.

[0103] Through this, the anomaly detection model can determine that case 610 is normal and case 620 is abnormal. The outlier score can be calculated by taking the reciprocal of the number of divisions. For example, case 610 can be calculated as 1 / 6, and case 620 can be calculated as 1 / 2.

[0104] The second anomaly detection unit can detect whether an anomaly occurs in the input data by determining whether the anomaly score, which is the output value of the anomaly detection model, exceeds the threshold value. That is, if 610 is 1 / 6 and the threshold value is 1 / 3, the input data of 610 can be determined as normal. Since 620 is 1 / 2, which is a value greater than the threshold value of 1 / 3, it can be determined as an anomaly. Alternatively, the anomaly score can be scored as the number of divisions without taking the reciprocal. Alternatively, the anomaly score can be set to a value changed through various functions based on the number of divisions. There is no limitation on the calculation of the anomaly score.

[0105] Fig. 7 is a diagram for explaining an operation for detecting whether an abnormality has occurred according to one embodiment.

[0106] Referring to Fig. 7, the second anomaly detection unit can determine whether an anomaly has occurred based on an anomaly score for input data input in a time-series manner. The X-axis represents time, and the Y-axis represents an anomaly score. When input data in the form of stream data is continuously input, the anomaly detection model calculates an anomaly score. Once the anomaly score is calculated, it is compared with a threshold value (730), and an anomaly score that exceeds the threshold value can be obtained.

[0107] For example, the second anomaly detection unit may determine an anomaly occurrence section (710, 720) by considering the difference between the anomaly score exceeding the threshold value (730) and the threshold value (730) and / or the frequency of input data exceeding the anomaly score occurring within a certain period of time.

[0108] When an abnormal occurrence section (710, 720) is detected, the second abnormal occurrence detection unit can change the scale to additionally check the abnormal occurrence input data. For example, in order to quickly detect an abnormal occurrence, the second abnormal occurrence detection unit may first determine whether an abnormality has occurred based on the bundled input data, and when an abnormal occurrence section (710, 720) is detected, it may also determine whether an abnormality has occurred for each individual input data.

[0109] The above describes the behavior of anomalies in input data. Below, we describe the monitoring of AI models for anomalies or performance degradation.

[0110] FIG. 8 is a drawing for explaining the operation of a third abnormality detection unit according to one embodiment.

[0111] Referring to FIG. 8, the third abnormality detection unit can input baseline data received from the database (210) into the artificial intelligence model (810) before receiving input data, and use the accuracy information generated to set a reference value (820). The third abnormality detection unit can detect that an abnormality has occurred in the artificial intelligence model (810) when the accuracy information of the output data is evaluated to be lower than the reference value.

[0112] The third anomaly detection unit detects whether an anomaly has occurred using the same artificial intelligence model (810). For example, the output value of the artificial intelligence model (810) can be obtained using baseline data received from the database (210). The output value for the baseline data can include accuracy information or information about the temporal flow of the output value. For example, when the output value for the baseline data is received, accuracy information for the output value can be calculated using each output value and the actual result value. Once the accuracy information is calculated, a reference value can be determined (820) by applying a certain offset to the accuracy information. The determined reference value (820) is used as a comparison value (840) to determine whether an anomaly has occurred in the artificial intelligence model in the future.

[0113] That is, the generation of reference values ​​or reference output data (820) using baseline data is performed and stored before the artificial intelligence model (810) abnormality detection operation using input data.

[0114] When input data is received, the third abnormality detection unit can input it into an artificial intelligence model (810) to obtain output data (830). The third abnormality detection unit can compare (840) the output data (830) with a reference value or reference output data (820) to determine whether an abnormality has occurred in the artificial intelligence model (810).

[0115] For example, if the reference value (820) is generated and stored as 85%, the third abnormality detection unit calculates the accuracy of the output data (830), compares the calculated accuracy with the reference value (820) (840), and if it is evaluated as being lower than the reference value (820), it can detect that an abnormality has occurred in the corresponding artificial intelligence model.

[0116] This operation can be used when accuracy can be determined from the output values ​​of the AI ​​model (810). For example, if the AI ​​model (810) is set to predict data after a certain period of time, and actual data generated after the certain period of time can be obtained, the two can be compared to determine accuracy. However, if accuracy cannot be determined, reference output data can be used. This is exemplarily described with reference to FIG. 9.

[0117] FIG. 9 is a drawing for explaining an operation of detecting whether an abnormality has occurred according to the operation of a third abnormality detection unit according to one embodiment.

[0118] Referring to Figure 9, baseline data can be input into an artificial intelligence model to generate output data as reference output data. For example, the reference output data can be expressed as a flow of output data output in a time-series fashion over a certain period of time.

[0119] The third abnormality detection unit can detect whether an abnormality has occurred by comparing the flow of reference output data with the flow of output data. The X-axis represents time, and the Y-axis represents the value of the output data.

[0120] The third abnormality detection unit can determine that an abnormality has occurred in the artificial intelligence model by comparing the difference between the time series data values ​​of the reference output data and the time series data values ​​of the output data and if the deviation (900) exceeds the threshold value.

[0121] To achieve this, the time-series flow of reference output data relative to baseline data over a certain period is stored. Furthermore, the reference output data and the flow of output data are compared periodically. In other words, the reference output data can be repeated periodically and compared with the flow of output data.

[0122] Fig. 10 is a drawing for explaining an operation of detecting whether an abnormality has occurred according to the operation of a third abnormality detection unit according to another embodiment.

[0123] Referring to Fig. 10, the third abnormality detection unit can detect whether an abnormality has occurred by calculating the difference between the reference output data and the output data and determining whether the difference exceeds a threshold value. The X-axis represents time, and the Y-axis represents the root mean square error (RMSE).

[0124] For example, the third anomaly detection unit can calculate the root mean square error (RMSE) for the reference output data and the output data over time. As shown in Fig. 9, since RMSE is used as one of the methods for calculating the difference, if an RMSE value (1000) exceeds the threshold value, it can be detected that there is an anomaly in the AI ​​model in that section (1010).

[0125] Temporary data noise, etc. may occur, and the third anomaly detection unit may determine an anomaly when the difference value or RMSE value continuously exceeds the threshold value for a certain period of time. In other words, noise processing may be performed based on a preset time for temporary threshold exceedances, etc. This may be set in various ways depending on the data stability, sensitivity, etc. of the system to which this embodiment is applied.

[0126] As described above, the AI ​​model performance monitoring device according to the present disclosure can accurately detect anomalies by analyzing input data through various operations. Furthermore, by performing dual monitoring of input and output data, the cause of anomalies can be easily identified.

[0127] Below, the entire system in which the above-described operations are performed is described with reference to drawings as an example, and is briefly described to avoid redundant explanations.

[0128] FIG. 11 is a diagram illustrating the overall flow of an artificial intelligence model performance monitoring operation according to one embodiment.

[0129] Referring to FIG. 11, baseline data (1110) of a database (1100) can be input into an input data quality monitoring entity (1120). The baseline data (1110) can be used to produce statistical information, and statistical information is produced (1121) using the baseline data (1110). Statistical information on the produced baseline data (1110) is stored (1122).

[0130] Additionally, baseline data (1110) is also input to the anomaly detection model (1124). The baseline data (1110) can be used by the anomaly detection model (1124) for RCF learning (1125).

[0131] Additionally, baseline data (1110) can be input into an artificial intelligence model performance monitoring entity (1130). When baseline data (1110) is input into an artificial intelligence model (1131), output data is collected to generate a baseline result (1132). The baseline result (1132) may be a reference value or reference output data for accuracy comparison as described above.

[0132]

[0133] *When the pre-stop operation using baseline data (1110) is completed, production data (1115), which is input data, can be input. When the production data (1115) is input, entity 1120 uses the production data (1115) to calculate statistical information, retrieves the stored baseline statistical information (1122), and compares (1123) the statistical information using the production data (1115) with the statistical indicator. The comparison result can be transmitted to the monitoring entity (1140) and provided to the user.

[0134] Production data (1115) is also input to an anomaly detection model (1124). The anomaly detection model (1124) calculates an anomaly score (1126) using the input production data (1115). Based on the result of the anomaly score calculation, the occurrence of an anomaly can be detected. The result of the anomaly detection can be transmitted to a monitoring entity (1140) and provided to the user.

[0135] Production data (1115) is also input into the artificial intelligence model performance monitoring object (1130). The artificial intelligence model (1131) is identical to the artificial intelligence model (1131) using baseline data (1110).

[0136] The output data of the artificial intelligence model (1131) is compared (1133) with the results (1132) based on baseline data and used to detect whether an abnormality has occurred in the artificial intelligence model (1131). The results of detecting whether an abnormality has occurred in the artificial intelligence model (1131) can be transmitted to a monitoring entity (1140) and provided to a user.

[0137] The monitoring object (1140) can send a notification to a registered user when a preset alarm criterion is met.

[0138] Above, the input data quality monitoring entity (1120), the model performance monitoring entity (1130), and the monitoring entity (1140) are described separately according to their functions. As described above, the artificial intelligence performance monitoring device according to the present disclosure may include all of the above entities. Alternatively, it may include only some of them.

[0139] In this way, the present disclosure can efficiently detect various abnormal situations and monitor the performance of AI models already installed and operating in industrial settings. Furthermore, it can obtain the cause of anomalies, enabling accurate measures to be taken in response to the anomalies.

[0140] Below, the aforementioned embodiments and the operation of the AI ​​model performance monitoring device are described again in a time-series flowchart. Each step described below can be merged, split, or rearranged. Furthermore, to avoid unnecessary duplication, specific examples are omitted, and the content described above can be applied to each step in any combination.

[0141]

[0142] FIG. 12 is a diagram for explaining a method for monitoring artificial intelligence model performance according to one embodiment.

[0143] Referring to FIG. 12, the method for monitoring the performance of an artificial intelligence model may include a first abnormality detection step of storing input data received from a database for a preset period of time to generate statistical information, and comparing the statistical information with preset reference statistical information to detect whether an abnormality occurs in the input data (S1200).

[0144] For example, the first anomaly detection stage can store input data inputted into an artificial intelligence model for a preset period of time. In other words, the first anomaly detection stage uses a buffer to store input data for a set period of time to detect the occurrence of an anomaly.

[0145] The first anomaly detection step can detect the occurrence of anomalies in the input data using statistical information about the input data. For example, the statistical information may include at least one of the distribution, sum, standard deviation, minimum value, and maximum value of the input data.

[0146] The first anomaly detection step compares the input data's statistics with baseline statistics to detect anomalies in the input data. To achieve this, baseline statistics must be set in advance.

[0147] For example, the preset baseline statistical information may include at least one statistical indicator among a baseline distribution, a baseline mean, a baseline sum, a baseline standard deviation, a baseline minimum value, and a baseline maximum value, which are generated using baseline data received from a database before the input data is received. The statistical indicators of the input data and the baseline data must be identical.

[0148] The first anomaly detection step can be set to detect whether anomalies occur in input data using various statistical indicators.

[0149] For example, the first anomaly detection step can detect an abnormality in the input data by comparing the same statistical indicators of the statistical information and the preset reference statistical information and if the difference exceeds the preset threshold range, the occurrence of an abnormality in the input data can be detected. For example, the first anomaly detection step can generate statistical information using input data stored for a certain period of time. The first anomaly detection step compares the same statistical indicators of the statistical information on the preset reference statistical information and the input data to determine whether the difference exceeds the preset threshold range, thereby determining whether an abnormality has occurred in the input data buffered for a certain period of time. The storage period of the input data can be set to be the same as the storage period for generating statistical information on the baseline data. Alternatively, it can be set differently. This information can be applied to statistical indicators that are represented by a single value. For example, in cases where statistical information is represented by a single value, such as a sum, standard deviation, maximum value, and minimum value, the first anomaly detection step can compare the difference between the statistical indicators and determine whether the difference exceeds the threshold range.

[0150] As another example, the first anomaly detection step can detect the presence of anomalies in input data using a function that calculates the difference between the probability distribution of the reference distribution and the distribution included in the statistical information. For example, if the statistical indicator is not a single value, the first anomaly detection step can detect the presence of anomalies in the input data using the distribution among the statistical indicators. For example, the presence of anomalies can be detected by comparing the reference distribution for a certain period of baseline data generated using the baseline data with the distribution for a certain period of input data generated using the input data.

[0151] To this end, the first anomaly detection step can use a function to calculate the probability distribution difference between each distribution. For example, the first anomaly detection step can use a function such as the Kullback-Leibler divergence to determine whether an anomaly has occurred based on whether the difference between probability distributions exceeds a certain level. In addition, various functions can be applied to calculate probability distribution differences.

[0152] In addition, the artificial intelligence model performance monitoring method may include a second anomaly detection step of inputting input data into a preset anomaly detection model and detecting whether an anomaly occurs in the input data based on an output value of the anomaly detection model (S1210).

[0153] For example, the second anomaly detection stage can input input data into a pre-trained and configured anomaly detection model. The second anomaly detection stage can detect whether anomalies occur in the input data using the output values ​​produced by the anomaly detection model.

[0154] For example, the second anomaly detection step can detect that an anomaly has occurred in the input data when the output value of the preset anomaly detection model that calculates an anomaly score based on the number of divisions for isolating the input data from the baseline data exceeds a preset threshold value.

[0155] For example, an anomaly detection model can use a tree-structured model. The anomaly detection model can be trained using baseline data. Alternatively, the anomaly detection model can calculate the number of divisions for dividing a two-dimensional plane to separate newly input data from the baseline data based on the distribution of the baseline data. The anomaly detection model can calculate the inverse value of the number of divisions as an outlier score. Therefore, a higher outlier score can be determined to indicate an anomaly in the input data. To this end, the second anomaly detection step can compare a threshold value with the outlier score, and if the outlier score is equal to or exceeds the threshold value, it can be determined that an anomaly has occurred in the corresponding input data.

[0156] As another example, the second anomaly detection step may use the same number of segmentations to isolate the input data, with a lower number indicating an anomaly in the input data. In this case, the second anomaly detection step may determine that an anomaly exists in the input data if the number of segmentations falls below a threshold.

[0157] For example, anomaly detection models can use the Random Forest algorithm. Another example is the Random Cut Forest. In addition, various models can be applied as anomaly detection models, such as those that can detect when specific data is out of sync or deviates from the baseline data by converting this into numerical values.

[0158] The method for monitoring the performance of an artificial intelligence model may include a third abnormality detection step of inputting input data into a preset artificial intelligence model and detecting whether an abnormality occurs in the artificial intelligence model using the output data of the artificial intelligence model (S1220).

[0159] For example, the third anomaly detection stage can detect whether an anomaly has occurred by verifying the output data of the artificial intelligence model itself.

[0160] For example, the third anomaly detection step can set a reference value using accuracy information derived from baseline data received from a database before receiving input data and then inputting the data into the AI ​​model. In other words, a reference value for detecting anomalies in the AI ​​model can be set using baseline data.

[0161] The third anomaly detection stage can detect an anomaly in the AI ​​model when the accuracy of the output data relative to the input data is evaluated to be below a reference value. For example, the third anomaly detection stage can input a certain level of baseline data into the AI ​​model and measure accuracy using the output data relative to the baseline data. The reference value can be set in advance by applying a certain level of offset to the measured accuracy information.

[0162] The third anomaly detection stage measures the accuracy of output data for input data, and determines that an anomaly has occurred if the accuracy falls below a preset standard value.

[0163] Accuracy can be measured by comparing predicted output data in industrial settings to actual values ​​generated after a certain period of time. Therefore, the third anomaly detection stage can detect anomalies in the AI ​​model when accuracy falls below a certain level.

[0164] However, there may be cases where the accuracy of output data cannot be measured for various reasons. This can be difficult to measure, such as when comparing predicted output data with actual data, or when measuring accuracy in AI models, such as models that generate actual drive control signals rather than predicted behavior. Therefore, considering these cases, the third anomaly detection stage can utilize deviations between output data.

[0165] For example, the third anomaly detection step can input baseline data received from a database into an AI model before receiving input data, and determine whether the deviation between the output data and the reference output data exceeds a preset threshold. The third anomaly detection step can detect an anomaly in the AI ​​model if the deviation is determined to be above or exceed the threshold.

[0166] To this end, the third abnormality detection step can use baseline data to check the output data of the artificial intelligence model in advance for a certain period of time, and store the output data changes over time as reference output data.

[0167] For example, the third anomaly detection stage can detect the occurrence of an anomaly in the AI ​​model by storing reference output data in the time domain for a certain period of time and comparing the output data with the reference output data stored in the time domain at regular intervals. In other words, the third anomaly detection stage can compare the output data measured in real time by the input data based on the time flow of the stored reference output data, and detect an anomaly if the deviation between the two output data exceeds a certain level.

[0168] If necessary, the method for monitoring the performance of an artificial intelligence model may further include a monitoring step of monitoring at least one of whether an abnormality occurs in input data and whether an abnormality occurs in an artificial intelligence model, and providing a notification by distinguishing the type of abnormality (S1230).

[0169] For example, the monitoring step may receive information regarding the results of abnormality detection generated in the first abnormality detection step, the second abnormality detection step, and the third abnormality detection step. The monitoring step may generate a control signal so that each piece of received information is output through an output device. In addition, the monitoring step may process the abnormality occurrence information so that the occurrence of an abnormality for the same input data is displayed by considering the correspondence between input data for mapping the occurrence of each abnormality.

[0170] For example, the first anomaly detection step stores input data for a certain period of time, which may result in a time difference from the second anomaly detection step, which detects whether anomalies occur for each input data. The monitoring step may take this into account and process information about whether anomalies occur for each input data so that the information is displayed correspondingly through an output device. Information about whether anomalies occur for an artificial intelligence model may be displayed correspondingly to the input data.

[0171] Through this, the monitoring stage can classify whether anomalies detected in the output data of an AI model are due to input data anomalies or AI model degradation. Furthermore, the monitoring stage can provide users with the cause and timing of the anomaly through the classified results.

[0172] Through the aforementioned actions, the AI ​​model performance monitoring method can monitor and provide notifications for any abnormalities in the AI ​​model and input data during operation. Furthermore, if an abnormality occurs in the AI ​​model, the AI ​​model performance monitoring method can also provide notifications by distinguishing whether the cause of the abnormality is due to AI model degradation or input data abnormalities.

[0173] In this way, the present disclosure can efficiently detect various abnormal situations and monitor the performance of AI models already installed and operating in industrial settings. Furthermore, it can obtain the cause of anomalies, enabling accurate measures to be taken in response to the anomalies.

[0174] The above description is merely an illustrative example of the technical idea of ​​the present disclosure, and those skilled in the art to which the present disclosure pertains will appreciate that various modifications and variations can be made without departing from the essential characteristics of the technical idea of ​​the present disclosure. In addition, the present embodiments are not intended to limit the technical idea of ​​the present disclosure but rather to explain it, and therefore the scope of the technical idea of ​​the present disclosure is not limited by these embodiments. The scope of protection of the present disclosure should be interpreted by the claims below, and all technical ideas within a scope equivalent thereto should be interpreted as being included within the scope of the rights of the present disclosure.

[0175]

[0176] CROSS-REFERENCE TO RELATED APPLICATION

[0177] This patent application claims priority under 35 USC § 119(a) to Korean Patent Application No. 10-2023-0182612, filed December 15, 2023, the entire contents of which are incorporated herein by reference. Furthermore, this patent application claims priority in countries other than the United States for the same reasons, the entire contents of which are incorporated herein by reference.

Claims

1. In the artificial intelligence model performance monitoring device, A first abnormality detection unit that stores input data received from a database for a preset period of time to generate statistical information, and compares the statistical information with preset reference statistical information to detect whether an abnormality occurs in the input data; A second abnormality detection unit that inputs the above input data into a preset abnormality detection model and detects whether an abnormality occurs in the input data based on the output value of the above abnormality detection model; and An artificial intelligence model performance monitoring device including a third abnormality detection unit that inputs the above input data into a preset artificial intelligence model and detects whether an abnormality occurs in the artificial intelligence model using the output data of the artificial intelligence model.

2. In paragraph 1, The above preset standard statistical information is: An artificial intelligence model performance monitoring device including at least one statistical indicator among a baseline distribution, a baseline mean, a baseline sum, a baseline standard deviation, a baseline minimum value, and a baseline maximum value generated using baseline data received from the database before the input data is received.

3. In paragraph 2, The above first abnormality detection unit, An artificial intelligence model performance monitoring device that detects that an abnormality has occurred in the input data when the same statistical indicator of the above statistical information and the above preset reference statistical information is compared and exceeds a preset threshold range.

4. In paragraph 2, The above first abnormality detection unit, An artificial intelligence model performance monitoring device that detects whether an abnormality occurs in the input data by using a function that calculates the difference in probability distribution between the above-mentioned reference distribution and the distribution included in the above-mentioned statistical information.

5. In paragraph 1, The above preset anomaly detection model is, An artificial intelligence model performance monitoring device, which is an unsupervised algorithm model that is trained using baseline data received from the database before the input data is received.

6. In paragraph 1, The above second abnormality detection unit, An artificial intelligence model performance monitoring device that detects that an anomaly has occurred in the input data when the output value of the preset anomaly detection model that calculates an anomaly score based on the number of divisions for isolating the input data from the baseline data exceeds a preset threshold value.

7. In paragraph 1, The above third and above detection unit, Before the above input data is received, baseline data received from the above database is input into the artificial intelligence model, and the accuracy information produced is used to set a reference value, An artificial intelligence model performance monitoring device that detects that an abnormality has occurred in the artificial intelligence model when the accuracy information of the above output data is evaluated to be below the above reference value.

8. In paragraph 1, The above third and above detection unit, An artificial intelligence model performance monitoring device that detects that an abnormality has occurred in the artificial intelligence model when the deviation between the output data and the reference output data output by inputting baseline data received from the database into the artificial intelligence model before the input data is received exceeds a preset threshold value.

9. In paragraph 8, The above third and above detection unit, An artificial intelligence model performance monitoring device that stores the above-mentioned reference output data for a certain period of time in the time domain, and compares the above-mentioned output data with the above-mentioned reference output data stored in the time domain at certain intervals to detect whether an abnormality occurs in the above-mentioned artificial intelligence model.

10. In paragraph 1, An artificial intelligence model performance monitoring device further comprising a monitoring unit that monitors at least one of an abnormality occurrence in the input data and an abnormality occurrence in the artificial intelligence model, and provides a notification by distinguishing the type of abnormality occurrence.

11. In the method of monitoring the performance of an artificial intelligence model, A first abnormality detection step for storing input data received from a database for a preset period of time to generate statistical information, and comparing the statistical information with preset reference statistical information to detect whether an abnormality occurs in the input data; A second abnormality detection step for inputting the above input data into a preset abnormality detection model and detecting whether an abnormality occurs in the input data based on the output value of the above abnormality detection model; and An artificial intelligence model performance monitoring method including a third abnormality detection step of inputting the above input data into a preset artificial intelligence model and detecting whether an abnormality occurs in the artificial intelligence model using the output data of the artificial intelligence model.

12. In paragraph 11, The above preset standard statistical information is: A method for monitoring performance of an artificial intelligence model, the method comprising: generating at least one statistical indicator among a baseline distribution, a baseline mean, a baseline sum, a baseline standard deviation, a baseline minimum value, and a baseline maximum value using baseline data received from the database before the input data is received.

13. In paragraph 12, The above first abnormality detection step is, An artificial intelligence model performance monitoring method for detecting that an abnormality has occurred in the input data when the same statistical indicator of the above statistical information and the above preset reference statistical information is compared and the same statistical indicator exceeds a preset threshold range.

14. In paragraph 12, The above first abnormality detection step is, An artificial intelligence model performance monitoring method for detecting whether an abnormality occurs in the input data by using a function that calculates the difference in probability distribution between the above-mentioned reference distribution and the distribution included in the above-mentioned statistical information.

15. In paragraph 11, The above preset anomaly detection model is, A method for monitoring the performance of an artificial intelligence model, which is an unsupervised algorithm model that is trained using baseline data received from the database before the input data is received.

16. In paragraph 11, The above second abnormality detection step is, An artificial intelligence model performance monitoring method for detecting that an anomaly has occurred in the input data when the output value of the preset anomaly detection model, which calculates an anomaly score based on the number of divisions for isolating the input data from the baseline data, exceeds a preset threshold value.

17. In paragraph 11, The above third abnormality detection step is, Before the above input data is received, baseline data received from the above database is input into the artificial intelligence model, and the accuracy information produced is used to set a reference value, An artificial intelligence model performance monitoring method for detecting that an abnormality has occurred in the artificial intelligence model when the accuracy information of the above output data is evaluated to be below the above reference value.

18. In paragraph 11, The above third abnormality detection step is, An artificial intelligence model performance monitoring method for detecting that an abnormality has occurred in the artificial intelligence model when the deviation between the output data and the reference output data output by inputting baseline data received from the database into the artificial intelligence model before the input data is received exceeds a preset threshold value.

19. In paragraph 18, The above third abnormality detection step is, A method for monitoring the performance of an artificial intelligence model, which stores the reference output data for a certain period of time in the time domain and compares the output data with the reference output data stored in the time domain at regular intervals to detect whether an abnormality occurs in the artificial intelligence model.

20. In paragraph 11, A method for monitoring performance of an artificial intelligence model, further comprising a monitoring step of monitoring at least one of an abnormality occurrence in the input data and an abnormality occurrence in the artificial intelligence model, and providing a notification by distinguishing the type of abnormality occurrence.

Images